The Yale Law Journal

May 2017

Machine Testimony

Andrea Roth

abstract. Machines play increasingly crucial roles in establishing facts in legal disputes. Some machines convey information—the images of cameras, the measurements of thermometers, the opinions of expert systems. When a litigant offers a human assertion for its truth, the law subjects it to testimonial safeguards—such as impeachment and the hearsay rule—to give juries the context necessary to assess the source’s credibility. But the law on machine conveyance is confused: courts shoehorn them into existing rules by treating them as “hearsay,” as “real evidence,” or as “methods” underlying human expert opinions. These attempts have not been wholly unsuccessful, but they are intellectually incoherent and fail to fully empower juries to assess machine credibility. This Article seeks to resolve this confusion and offer a coherent framework for conceptualizing and regulating machine evidence. First, it explains that some machine evidence, like human testimony, depends on the credibility of a source. Just as so-called “hearsay dangers” lurk in human assertions, “black box dangers”—human and machine errors causing a machine to be false by design, inarticulate, or analytically unsound—potentially lurk in machine conveyances. Second, it offers a taxonomy of machine evidence, explaining which types implicate credibility and how courts have attempted to regulate them through existing law. Third, it offers a new vision of testimonial safeguards for machines. It explores credibility testing in the form of front-end design, input, and operation protocols; pretrial disclosure and access rules; authentication and reliability rules; impeachment and courtroom testing mechanisms; jury instructions; and corroboration rules. And it explains why machine sources can be “witnesses” under the Sixth Amendment, refocusing the right of confrontation on meaningful impeachment. The Article concludes by suggesting how the decoupling of credibility testing from the prevailing courtroom-centered hearsay model could benefit the law of testimony more broadly.

author. Assistant Professor of Law, UC Berkeley School of Law. The author wishes to thank George Fisher, Erin Murphy, Daniel Richman, David Sklansky, and Eleanor Swift for extensive feedback. For technical information I am grateful to Nathaniel Adams, Jennifer Friedman, Angelyn Gates, Jessica Goldthwaite, Andrew Grosso, Allan Jamieson, Dan Krane, and Terri Rosenblatt. I am also indebted to Ty Alper, Laura Appleman, David Ball, Ryan Calo, Edward Cheng, Catherine Crump, David Engstrom, Dan Farber, Brandon Garrett, Andrew Gilden, Robert Glushko, Chris Hoofnagle, Edward Imwinkelried, Elizabeth Joh, Owen Jones, Robert MacCoun, Jennifer Mnookin, Deirdre Mulligan, Melissa Murray, John Paul Reichmuth, Pamela Samuelson, Jonathan Simon, Aaron Simowitz, Christopher Slobogin, Avani Mehta Sood, Rachel Stern, Karen Tani, Kate Weisburd, Charles Weisselberg, Rebecca Wexler, Tal Zarsky, the Berkeley Law junior faculty, and participants in the Berkeley, Stanford, Vanderbilt, and Willamette faculty workshops. Christian Chessman, Jeremy Isard, Purba Mukerjee, and Allee Rosenmayer offered excellent research assistance, and the Yale Law Journal editors offered invaluable editorial guidance.


In 2003, Paciano Lizarraga-Tirado was arrested and charged with illegally reentering the United States after having been deported.1 He admitted that he was arrested in a remote area near the United States-Mexico border, but claimed he was arrested in Mexico while awaiting instructions from a smuggler. To prove the arrest occurred in the United States, the prosecution offered the testimony of the arresting officers that they were familiar with the area and believed they were north of the border, in the United States, when they made the arrest. An officer also testified that she used a Global Positioning System (GPS) device to determine their location by satellite, and then inputted the coordinates into Google Earth. Google Earth then placed a digital “tack” on a map, labeled with the coordinates, indicating that the location lay north of the border.2 Mr. Lizarraga-Tirado insisted that these mechanical accusations were “hearsay,” out-of-court assertions offered for their truth, and thus inadmissible. The Ninth Circuit rejected his argument, even while acknowledging that the digital “tack” was a “clear assertion[],” such that if the tack had been manually placed on the map by a person, it would be “classic hearsay.”3 In the court’s view, machine assertions—although raising reliability concerns4—are simply the products of mechanical processes and, therefore, akin to physical evidence. As such, they are adequately “addressed by the rules of authentication,” requiring the proponent to prove “that the evidence ‘is what the proponent claims it is,’”5 or by “judicial notice,”6 allowing judges to declare the accuracy of certain evidence by fiat.

Mr. Lizarraga-Tirado’s case is emblematic of litigants’ increasing reliance on information conveyed by machines.7 While scientific instruments and cameras have been a mainstay in courtrooms for well over a century, the past century has witnessed a noteworthy rise in the “‘silent testimony’ of instruments.”8 By the 1940s, courts had grappled with “scientific gadgets” such as blood tests and the “Drunk-O-Meter,”9 and by the 1960s, the output of commercially used tabulating machines.10 Courts now routinely admit the conveyances11 of complex proprietary algorithms, some created specifically for litigation, from infrared breath-alcohol-testing software to expert systems diagnosing illness or interpreting DNA mixtures. Even discussions of the potential for robot witnesses have begun in earnest.12

This shift from human- to machine-generated proof has, on the whole, enhanced accuracy and objectivity in fact finding.13 But as machines extend their reach and expertise, to the point where competing expert systems have reached different “opinions” related to the same scientific evidence,14 a new sense of urgency surrounds basic questions about what machine conveyances are and what problems they pose for the law of evidence. While a handful of scholars have suggested in passing that “the reports of a mechanical observer” might be assertive claims implicating credibility,15 legal scholars have not yet explored machine conveyances in depth.16

This Article seeks to resolve this doctrinal and conceptual confusion about machine evidence by making three contributions. First, the Article contends that some types of machine evidence merit treatment as credibility-dependent conveyances of information. Accordingly, the Article offers a framework for understanding machine credibility by describing the potential infirmities of machine sources. Just as human sources potentially suffer the so-called “hearsay dangers” of insincerity, ambiguity, memory loss, and misperception,17 machine sources potentially suffer “black box” dangers18 that could lead a factfinder to draw the wrong inference from information conveyed by a machine source. A machine does not exhibit a character for dishonesty or suffer from memory loss. But a machine’s programming, whether the result of human coding or machine learning,19 could cause it to utter a falsehood by design. A machine’s output could be imprecise or ambiguous because of human error at the programming, input, or operation stage, or because of machine error due to degradation and environmental forces. And human and machine errors at any of these stages could also lead a machine to misanalyze an event. Just as the “hearsay dangers” are believed more likely to arise and remain undetected when the human source is not subject to the oath, physical confrontation, and cross-examination,20 black box dangers are more likely to arise and remain undetected when a machine utterance is the output of an “inscrutable black box.”21

Because human design, input, and operation are integral to a machine’s credibility, some courts and scholars have reasoned that a human is the true “declarant”22 of any machine conveyance.23 But while a designer or operator might be partially epistemically or morally responsible for a machine’s statements, the human is not the sole source of the claim. Just as the opinion of a human expert is the result of “distributed cognition”24 between the expert and her many lay and expert influences,25 the conveyance of a machine is the result of “distribut[ed] cognition between technology and humans.”26 The machine is influenced by others, but is still a source whose credibility is at issue. Thus, any rule requiring a designer, inputter, or operator to take the stand as a condition of admitting a machine conveyance should be justified based on the inability of jurors, without such testimony, to assess the black box dangers. In some cases, human testimony might be unnecessary or, depending on the machine, insufficient to provide the jury with enough context to draw the right inference. Human experts often act as “mere scrivener[s]”27 on the witness stand, regurgitating the conveyances of machines. Their testimony might create a veneer of scrutiny when in fact the actual source of the information, the machine, remains largely unscrutinized.

Second, the Article offers a taxonomy of machine evidence that explains which types implicate credibility and explores how courts have attempted to regulate them. Not all machine evidence implicates black box dangers. Some machines are simply conduits for the assertions of others, tools facilitating testing, or conveyances offered for a purpose other than truth. But “silent witnesses” that convey images and machines that convey symbolic output—from pendulum clocks to probabilistic genotyping software—do implicate black box dangers. These claim-conveying machines vary widely in their complexity, opacity, sensitivity to case-specific human manipulation, and litigative or nonlitigative purpose, and they might involve a low or high risk of inferential error absent a further opening of their black box. But they should be recognized, in the first instance, as credibility-dependent proof. As it turns out, courts have often shown promising intuitions about black box dangers in their attempts to regulate machine conveyances. But those attempts, particularly with respect to proprietary algorithms created for litigation, have too often been incoherent or incomplete. Meanwhile, commentators sometimes conflate credibility-dependent machine evidence with machine tools, conduits, or conveyances offered for a purpose other than truth when describing the influx of machine evidence into criminal trials.28

Finally, the Article offers a new vision of testimonial safeguards for machine sources of information. For several reasons, the Article does not advocate a broad rule of exclusion, akin to the hearsay rule, for “risky” machines.29 First, the hearsay rule itself could not easily be modified to accommodate machines, given its focus on the oath, physical confrontation, and cross-examination. Second, a broad category of exclusion might be less appropriate for machine sources than for human sources, whose frailties and foibles largely motivated the rise of machine-generated proof to begin with.30 Third, even with respect to human declarants, the hearsay rule is already highly unpopular for categorically excluding so much relevant evidence while being riddled with exceptions that are largely tradition-based and empirically unfounded.31 Instead, this Article focuses on safeguards that would offer the jury key foundational facts or context32 to better assess the accuracy33 of machine conveyances.

Lawmakers should first consider design, input, and operation protocols to improve accuracy, much like the protocols that govern breath-alcohol machines and, in some states, eyewitness testimony.34 Such front-end protocols could include software testing, machine-learning performance evaluations, and variations of “adversarial design,”35 in which competing perspectives are incorporated at the design stage into the variables and analytical assumptions of algorithms. Next, lawmakers should consider pretrial disclosure and accessrules for machines, especially machine “experts.” These rules might allow litigants to access machines before trial to test different parameters or inputs (much like posing hypotheticals to human experts). The rules might also require public access to programs for further testing or “tinkering”;36 disclosure of “source code,”37 if necessary to meaningfully scrutinize the machine’s claims;38 and the discovery of prior statements or “Jencks material”39 of machines, such as COBRA data for breath-testing machines.40 Lawmakers should also continue to require authentication of machine-related items to ensure that a machine conveyance, whether a DNA-typing printout or email, is what the proponent says it is.41

For machines offering “expert” evidence on matters beyond the ken of the jury,42 lawmakers should clarify and modify existing Daubert and Frye reliability requirements for expert methods43 to ensure that machine processes are based on reliable methods and are implemented in a reliable way. Daubert-Frye hearings are a promising means of excluding the most demonstrably unreliable machine sources, but beyond the obvious cases, these hearings do not offer sufficient scrutiny. Judges generally admit such proof so long as validation studies can demonstrate that the machine’s error rate is low and that the principles underlying its methodology are sound.44 But validation studies are often conducted under idealized conditions, and it is precisely in cases involving less-than-ideal conditions—degraded or highly complex mixtures difficult for human analysts to interpret—that expert systems are most often deployed and merit the most scrutiny. Moreover, machine conveyances are often in the form of predictive scores and match statistics, which are harder to falsify through validation against a known baseline. For example, even if a DNA expert system rarely falsely includes a suspect as a contributor to a DNA mixture, its match statistics might be off by orders of magnitude because of a host of human or machine errors, potentially causing jurors to draw the wrong inference. Courts applying Daubert-Frye to software-generated statements should treat software engineers as part of the relevant scientific community and determine reliability not only of the method, but also of the software implementing that method, based on industry standards. In some cases, courts would likely need to access proprietary source code to assess the code’s ability to operationalize an otherwise reliable method.

Beyond the admissibility stage, an opponent should be allowed to impeach machines at trial, just as the opponent can impeach human witnesses and declarants even when a judge deems their assertions reliable.45 Lawmakers should allow impeachment of machines by inconsistency, incapacity, and the like, as well as by evidence of bias or bad character in human progenitors. Lawmakers might even impose live testimonyrequirements for human designers, inputters, or operators in certain cases where testimony is necessary to scrutinize the accuracy of inputs, as the United Kingdom has done in criminal cases. Courts could also give jury instructions for certain machines typically under- or overvalued by jurors, akin to those used for human declarants like accomplices. And they could impose corroboration requirements, akin to those imposed on accomplice testimony and confessions, for certain risky machines or machines whose results lie within a certain margin of error. Such requirements might be grounded in concerns not only about accuracy, but also about public legitimacy in cases where the sole evidence of guilt is machine output.46

Finally, in criminal cases, machine sources of accusation—particularly proprietary software created for litigation—might be “witnesses against” a defendant under the Confrontation Clause.47 Accusatory machine output potentially implicates the central concerns underlying the Clause in three ways. First, if substituted for the testimony of witnesses otherwise subject to credibility testing, machine testimony allows the State to evade responsibility for accusations. Second, the State’s ability to shape and shield testimony from scrutiny through proprietary black box algorithms is analogous to the ex parteaffidavit practice that preoccupied the Framers. Third, machines are potentially unreliable when their processes are shrouded in a black box. While machines generally cannot be physically confronted, they can be impeached in other ways, and courts and scholars should revisit cases in which the Supreme Court appears to recognize implicitly that “confrontation” includes a right of meaningful impeachment.

Part I of this Article argues that some machine evidence implicates credibility and catalogs black box dangers—potential testimonial infirmities of machine sources. Part II offers a taxonomy of machine evidence, explaining which types do and do not implicate credibility and exploring how courts have attempted to regulate different machine conveyances under existing law. Part III suggests testimonial safeguards for machines, including both credibility-testing mechanisms that would target the black box dangers and methods of confronting accusatory machine conveyances under the Sixth Amendment. The Article concludes by explaining how the law of testimony more broadly could be improved by decoupling credibility testing and the hearsay rule and refocusing safeguards for all testimony on a right of meaningful impeachment.

I. a framework for identifying credibility-dependent machine evidence

This Part argues that some machine evidence implicates the credibility of its machine source—that is, the machine’s worthiness of being believed. It then offers a framework for describing the testimonial infirmities of machines, cataloging the black box dangers of falsehood by design, inarticulateness, and analytical error—caused by a variety of human and machine errors at the design, input, and operation stages—that might cause a factfinder to draw an improper inference from a machine source of information.

A. Machines as Sources Potentially in Need of Credibility Testing

How testimony48 differs from alternative ways we come to know facts has been the subject of debate. Epistemologists generally recognize a distinction between “testimony” and “non-informational expressions of thought.”49 Legal scholars have also suggested a distinction between “testimony” and other evidence. Nineteenth-century treatise writer Thomas Starkie described “testimony” as “information derived . . . from those who had actual knowledge of the fact,”50 and physical evidence as objects or conduct capable of being assessed through “actual and personal observation” by the jury.51

Both physical and testimonial evidence can lead to decisional inaccuracy. A jury asked to draw inferences from physical evidence, such as a large blood-stained serrated knife allegedly found in the defendant’s purse after the murder, must be given the tools to determine that the large blood-stained serrated knife is, in fact, the same knife that was found in the defendant’s purse. This process of “authenticating” the knife might require testimony of witnesses who found the knife, and that testimony might have its own set of credibility problems. But factfinders can assess, based on their own powers of observation and reasoning, the probative value of the knife’s physical properties.

Testimonial evidence presents different challenges for decisional accuracy. Even if the factfinder’s powers of observation and inference are working well, she might draw an improper inference if the source is not worthy of belief. In the hopes of offering juries sufficient context to assess the probative value of human testimony,52 American jurisdictions have adopted rules of exclusion,53 disclosure,54 impeachment,55 and corroboration,56 and, to a lesser extent, jury instructions57 and rules of production,58 to screen out the most egregiously noncredible human sources and—if testimony is admitted—to empower factfinders with information sufficient for them to assess accurately a source’s credibility.

Predictably, lawmakers and scholars disagree about precisely which human acts and utterances should be subject to these safeguards. But they all invoke the same potential infirmities—the so-called “hearsay dangers”—of human sources: insincerity, inarticulateness, erroneous memory, and faulty perception.59 For example, scholars seem to agree that so-called “implied assertions”—acts and utterances not intended by the source as an assertion, but that convey the source’s belief that a condition is true and are offered to prove the truth of that belief—trigger credibility concerns because their probative value turns on the source’s perceptive abilities.60 But courts generally exempt “implied assertions” from the hearsay rule because other infirmities, such as insincerity, are unlikely to arise.61

Lawmakers and scholars should likewise be open to viewing machine acts and utterances as dependent on credibility, if their probative value turns on whether a machine suffers testimonial infirmities. A handful of scholars have acknowledged that a machine, if it conveys information relied upon by others, offers testimonial knowledge, or a type of “instrumental knowledge” “closely related” to testimonial knowledge.62 A handful of courts have also used words like “credibility” and “impeachment” to describe machine sources.63

Two theoretical objections to the concept of “machine credibility” might be raised at the outset. The first is, as some courts and litigants have insisted, that machine conveyances are simply the hearsay assertions of the machine’s human programmer or inputter.64 This argument offers a strategic payoff for some litigants, particularly criminal defendants, as it would exclude the machine conveyance absent live testimony of the programmer. The argument also has intuitive appeal. Even the most sophisticated machines today are bundles of metal and circuitry whose journey from “on” switch to output begins with the instructions laid out for them; even robots “are not capable of deviating from the code that constitutes them.”65

Ultimately, though, this argument fails. That a programmer has designed a machine to behold and report events does not mean the programmer herself has borne witness to those events. As the Fourth Circuit noted in rejecting such an argument with respect to a gas chromatograph, “[t]he technicians could neither have affirmed [n]or denied independently that the blood contained” drugs “because all the technicians could do was to refer to the raw data printed out by the machine.”66 The argument also fails to recognize the phenomenon of machine learning67 and other unpredictable aspects of modern machine operation; “[p]rogrammers do not, and often cannot, predict what their complex programs will do.”68 Indeed, machine learning “can lead to solutions no human would have come to on her own,” including patentable inventions69 and award-winning literature.70

That is not to say that humans bear no responsibility for machine output. A programmer might be legally responsible for machine output that is socially harmful71 or have output imputed to him under a fairness-based “opposing party admission”-type doctrine.72 A programmer might also be partially epistemically responsible for machine output because she gives the machine its analytical parameters and instructions. In the human context, an expert witness’s Ph.D. advisor, or witnesses interviewed by an expert in the course of rendering an expert opinion, might be partially epistemically responsible for the expert’s opinions. Evidence scholars call this phenomenon “distributed cognition,” and it is a characteristic of all expert testimony,73 including that informed by technology.74 It is why experts are allowed to testify based on hearsay: otherwise, the proponent would be forced to call the Ph.D. advisor, and friends and family of a patient diagnosed with a mental illness in part based on such witnesses’ representations, to the stand.75 In the case of machines, juries might sometimes need the testimony of the machine’s “advisor”—the programmer—to adequately assess credibility, particularly since the machine cannot use its own judgment in deciding how much to rely on the instructions or assertions of its programmer.76 But any ruling allowing the programmer to testify should not be based on the premise that the programmer is the true declarant of the machine’s conveyance of information.

The second theoretical objection might be that machine sources are inherently different from human sources because machines do not engage in thought. But that premise, too, is questionable. While Western science has been dominated for centuries by a “passive-mechanistic” view that treats artificial beings as lacking agency, a competing line of thought has insisted that machines have agency, just like living beings, in that their actions are neither random nor predetermined.77 The father of modern computing, Alan Turing, famously suggested that a machine should be described as “thinking” so long as it could pass as human upon being subject to text-based questioning by a person in another room.78 “Machine cognition” is now an established field of study,79 and some have argued for a new “ontological category” for robots, between humans and inanimate objects.80

Although it seems clear that machines lack the ability to engage in moral judgment or to “intend” to lie, the need for credibility testing should not turn on whether a source can exercise moral judgment. The coherence of “machine credibility” as a legal construct depends on whether the construct promotes decisional accuracy, not on what cyberneticists or metaphysicists have to say about whether a machine can ever achieve “real boy” status. Legal scholars have similarly acknowledged that the question whether machine-generated communications can be “speech” for purposes of the First Amendment is “necessarily a normative project” that rests on one’s conception of why certain communications are labeled “speech” at all.81 If one believes “speech” as a legal category is intended primarily to protect explicitly political expressions, then much algorithm-generated speech might not be covered. If it is intended to promote truth by expanding the marketplace of ideas, then more machine speech might be covered.82 In the same respect, the question whether to subject machine evidence to credibility testing is a normative project. If one views the law of testimony as intended to promote decisional accuracy, and if black box dangers are not sufficiently guarded against under existing laws treating machine evidence as simply physical objects, then “machine testimony” is a category worthy of study.

B. Black Box Dangers: Causes of Inferential Error from Machine Sources

This Section explores the potential testimonial infirmities of machine sources. Some courts and scholars assume that “machines . . . fall outside the scope of hearsay ‘because the hearsay problems of perception, memory, sincerity and ambiguity have either been addressed or eliminated.’”83 It is true that machine conveyances are not “hearsay,” but not because they are immune from testimonial infirmities. While machines might be incapable of “memory loss,” a given machine conveyance—just like a human assertion—might be false or misleading because the machine is programmed to render false information (or programmed in a way that causes it to learn to do so), is inarticulate, or has engaged in analytical missteps.84

1. Human and Machine Causes of Falsehood by Design

Merriam-Webster defines “insincere” as “not expressing or showing true feelings.”85 A machine does not have “feelings,” nor does it suffer moral depravity in the form of a questionable character for truthfulness. But it could be deliberately programmed to render false information or programmed to achieve a goal in a way that leads the machine itself to learn to utter falsehoods as a means of achieving that goal.

Falsehood by human design. First, humans can design a machine in a way they know, or suspect, will lead a machine to report inaccurate or misleading information. A designer could choose to place inaccurate markings on a mercury thermometer’s side, or choose to place alcohol instead of mercury in the bulb during construction, both causing falsity by design. One recent example is the discovery that “[r]ogue [e]ngineers”86 at Volkswagen used “covert software” to program diesel vehicles to report misleading emissions numbers during pollution tests.87 Similarly, two Time Magazine journalists were able to determine, through Turing Test-like questioning, that a robot-telemarketer was programmed to falsely claim she was a real person.88 When one asks Siri, “are you a liar?”, her response is “no comment.”89 So long as programming technology exists, and motives to lie or cheat exist, programmers face the “temptation to teach products to lie strategically.”90

Falsehood by machine-learned design. A machine might also “teach” itself to lie as a strategy for achieving a goal. Once algorithms with “billions of lines of code” and “an enormous number of moving parts are set loose,” they go on to “interact with the world, and learn and react,” in ways that might be unpredictable to the original programmers.91 Even if a human does not program a machine to render false or misleading information, the machine can teach itself to lie if it learns that deception is a good strategy to reach a goal programmed into it.92 In one study, “hungry” robots “learned” to suppress information that clued in other robots to the location of a valuable food source.93 A legal system should establish safeguards to detect and avoid false or misleading machine testimony, whether the falsity is due to human design or machine-learning.94

2. Human and Machine Causes of Inarticulateness

Like a human source, a machine source might utter information that is inarticulate in a way that leads an observer to draw the wrong inference, even if the machine is otherwise nondeceptive and well designed to render an accurate claim. A machine’s reasons for being inarticulate are, like its reasons for being deceptive, different from those of a human witness. A machine does not slur its words due to intoxication or forget the meaning of a word. But a machine can be imprecise, ambiguous, or experience a breakdown in its reporting capacity due to human design, input, and operation errors, as well as machine errors caused by degradation and environment.

Human design. Human design choices—unless disclosed to the factfinder—can lead to inferential error if a machine’s conveyance reflects a programmed tolerance for uncertainty that does not match the one assumed by the factfinder. Imagine a human eyewitness tells a police officer at a lineup that he is “damn sure” the man who robbed him is suspect number five. Assume that if the defendant were able to cross-examine the eyewitness in court, the witness would clarify that, to him, “damn sure” means a subjective certitude of about eighty percent. But if the eyewitness never testifies and the prosecution calls the officer to relate the witness’s hearsay account, the factfinder might inaccurately infer that “damn sure” means a subjective certitude of ninety-nine percent. Machine conveyances might suffer the same ambiguity. If IBM’s Watson were to start conducting autopsies and reporting to factfinders—using a subjective scale—the likely cause of death in criminal cases based on a diagnostic algorithm, factfinders would not know—based solely on Watson’s output that the decedent “most likely” suffered from a particular condition—whether their own tolerance for uncertainty matched Watson’s. DNA match statistics generated by software, offered without information about the size of potential sampling error in the population frequency estimates used, would be another example,95 as would medical diagnosis algorithms, where software designers must make decisions about how far to tolerate false negatives and positives.96 This sort of failure to articulate a tolerance for uncertainty produces ambiguity. In contrast, a machine that is over- or underconfident in its assessment—that is, one that states a level of uncertainty about its assessment that does not correspond to the actual empirical probability of the event—suffers another sort of infirmity,97 whether an analytical error or falsehood by design.

Human operation. A human operator could also create ambiguity leading to inferential error by placing the machine in circumstances where its conveyance of information is misleading. Again analogizing to human testimony, imagine a person in a room overheard saying “Brrr—it’s cold.” A party now offers the statement as proof that the room was generally cold. In truth, the room was warm, but the declarant was standing directly in front of an air conditioning duct, a fact that would likely remain hidden absent the declarant’s live testimony.98 In the same respect, a thermometer placed in front of the air duct, if the reading is presented in court as an accurate report of room’s temperature, might cause the factfinder to draw the wrong inference.99

Machine degradation and malfunction. Due to entropy, machines stray from their original designs over time and potentially err in articulating their calculations. A digital thermometer’s battery might wear out to the point that “eights” appear to be “sixes.” A bathroom scale might be bumped such that—absent consistent calibration—it no longer starts its measurements at zero, thus overreporting weight. One could conceive of these errors as “machine errors,” because the machine has lost its ability to articulate, or as human maintenance errors, because an operator failed to take corrective action. The critical point is that, when left unattended, machines can malfunction in ways that manifest as inarticulate conveyances.

3. Human and Machine Causes of Analytical Error

In the early days of computing, some philosophers rejected the idea that a machine could “perceive” anything.100 Now, numerous universities have laboratories dedicated to the study of “machine perception,”101 from the development of hardware allowing machines to approximate human senses such as touch, vision, and hearing, to aesthetic judgment about art.102 Some machines are much cruder, “perceiving” only in the sense of interacting with and analyzing data. Given these ongoing debates about the differences between machine and human perception, I use the term “analytical error” rather than “misperception” to capture machine errors analogous to human cognitive and perceptive errors.

Human design. Analytical errors can stem from programming mistakes, beginning with inadvertent miscodes. Miscodes are inevitable; “bugs and misconfigurations are inherent in software.”103 In several cases, programmers have failed to program computer codes that could accurately translate legal code.104 Likewise, programmers have miscoded crime-detecting and forensic identification tools, which has led to inaccurate analysis of allelic frequencies, embedded in DNA-typing software to generate match statistics;105 to glitches in Apple’s “Find My iPhone” App that have led victims of iPhone theft and loss to the wrong locations;106 and to a “minor miscode” in a probabilistic DNA-genotyping software program that affected the reported match statistics in several cases, though generally not by more than an order of magnitude.107 Other notorious miscode examples include the Therac-25, a computer-controlled radiation therapy machine that “massively overdosed” six people in the late 1980s based on a software design error.108

Human design could also lead a machine to utter false or misleading information where the programmer makes inappropriate analytical assumptions or omissions. Programmers must incorporate a number of variables to ensure that machine estimates are accurate. For example, programmers must design breath-alcohol machines to distinguish between ethyl alcohol, the alcohol we drink, and other substances, such as acetone, that present similar profiles to a machine relying on infrared technology.109 They must also program breath-alcohol machines with an accurate “partition ratio” to calculate blood-alcohol level from the suspect’s breath-alcohol level, a ratio that some defense experts say differs nontrivially from person to person.110 An expert review of the “Alcotest 7110” source code found that, although the device was “generally scientifically reliable,” its software had several “mechanical and technical shortcomings.”111 This review prompted the New Jersey Supreme Court to require modifications to the machine’s programming to guard against misleadingly high readings.112 Moreover, in modeling highly complex processes, a programmer’s attempt to account for one variable might inadvertently cause another variable to lead to error. For example, Tesla now believes that the fatal crash of one of its self-driving cars into a truck trailer might have occurred because the car’s radar detected the trailer but discounted it as part of a design to “tune out” certain structures to avoid “false braking.”113

A programmer’s conscious or unconscious bias might also influence algorithms’ predictions or statistical estimates. For example, software designers have created compliance and risk-management software with “automation biases” to favor corporate self-interest,114 and Facebook recently rigged its “trending topics” algorithms to favor ideologically liberal content, a result the company insists was caused by “unconscious bias” on the part of human curators.115 And algorithm-generated credit scores and dangerousness “scores” may entrench bias by incorporating racially-correlated variables.116 In addition to designer bias, user patterns can inadvertently skew algorithms. For example, the WestlawNext algorithm may have the “potential to change the law” by biasing results away from “less popular legal precedents” and rendering those precedents “effectively . . . invisible.”117

Even if a programmer is not “biased” in the sense of making choices to further a preconceived goal, her analytically controversial choices can affect the accuracy of the machine’s scores and estimates. For example, in the DNA context, programmers have the power to set thresholds for what to count as a true genetic marker versus noise in determining which markers to report on the graphs used in determining a match.118 Programmers of DNA mixture interpretation software must also decide how conservative their estimates should be with respect to the probability of unusual events—such as small amounts of contamination during testing—that directly affect interpretation.119 Beyond the interpretation of the DNA sample itself, programmers must make judgment calls that affect the software’s report of a match statistic, such as determining the appropriate reference population for generating estimates of the rarity of genetic markers.120

Machine-learning in the design stage. Machines themselves might also augment their programming in ways that cause analytical errors. Machines learn how to categorize new data by training on an existing set of data that is either already categorized by a person (“supervised learning”) or is categorized by the computer itself using statistics (“unsupervised learning”).121 The fewer the samples in the training set,122 or the more that future data is likely to look different from the training set over time,123 the greater the chance the algorithm will draw an incorrect inference in future observations. Errors might occur because the machine infers a pattern or linkage in the limited data set that does not actually mirror real life (“overfitting”).124 Or the machine might try to account for too many variables, making the data set inadequate for learning (the “curse of dimensionality”),125 a reason that match-dating websites catering to narrower subgroups predict matches better.

In the crime-detecting context, imagine a machine like the Avista SmartSensor126 that teaches itself, after seeing how police categorized three hundred street level interactions through surveillance camera footage, that a person who shakes hands three times in a row is likely engaged in a drug transaction. Even if this new decision rule were reasonable based on the machine’s sample, an inference in a future case that two people are engaged in illegal activity based on that new programming might be incorrect. Alternatively, a machine might inaccurately infer that a crime is not occurring.

Human input and operation. Some machines do not require further human input, post-design, before conveying information. A mercury thermometer, for example, does not require a person to input information or physical objects before reporting ambient temperature. Even a highly complex “lay” machine, such as a robot security guard reporting what it has seen, is able to convey information based solely on its programming and the events it perceives. On the other hand, many machines do require human input to convey information. These human inputs can be either “physical” or “assertive,” but both types of input can lead to erroneous machine conveyances.

Assertive input encompasses information that humans enter into machines. Most “expert systems”—programs rendering complex analysis based on information fed to it by humans—require inputters to provide case-specific information, and those types of machines might misanalyze events or conditions if fed the wrong inputs. For example, DNA mixture interpretation software might require a human analyst to upload the DNA profile information of a typed sample before conducting its analysis. Similarly, a medical diagnosis expert system might require a human doctor to upload patient information.127

The potential for error stemming from expert systems’ reliance on the assertions of human inputters is analogous to the potential for error from human experts’ reliance on the assertions of others. The law of evidence generally shields juries from human testimony that merely repeats the assertions of others. Thus, as a general rule, lay witnesses are forbidden from testifying to statements made by others, on grounds that the hidden declarant’s testimonial capacities cannot be tested.128 But human experts may base their opinions in part on otherwise-inadmissible assertions made by other people, so long as those assertions are of the type “reasonably relied upon” by experts in the field.129 A human psychologist’s assertion that the defendant suffers from schizophrenia is likely a product of her schooling, the treatises and articles she has read, and the interviews she conducted with the defendant’s friends and family. In short, her assertion is a product of what evidence scholars have called “distributed cognition.”130 While distributed cognition is an inevitability of expert testimony, the possibility that these other assertions are false necessarily injects another potential source of error into an expert’s, or expert system’s, analysis.

Other problematic inputs leading to a false machine conveyance might be physical rather than assertive. For example, an operator of a breath-alcohol machine who fails to wait long enough after a suspect vomits before commencing the test runs the risk that the machine will mistake residual mouth alcohol for alcohol in deep lung air and inaccurately estimate the suspect’s blood-alcohol level.131 A computer-run DNA analysis on a crime-scene sample contaminated with residue from a suspect’s sample may, without correct control tests, falsely convey that the two samples match.132 “False” inputs might even include the failure to remove inputs that were correct when initially inputted, but have since become outdated. For example, the failure to scrutinize law enforcement databases for old, resolved warrants has led computer systems to falsely report to officers in the field that a suspect has an outstanding warrant.133

Machine error. Finally, analytical error can stem from machine malfunction due to degradation or environmental factors. A digital thermometer left rusting in the rain might experience a glitch in its computational process and render an incorrect result. A voltage change might cause a breath-testing machine’s process to malfunction during its analysis, leading to inaccurate results.134 An initially functioning computer program might experience “software rot,” a deteriorating and outdating of code over time that, if not subject to periodic software review that could detect such deterioration, could cause a machine to render false or misleading information. Or even an errant animal might be to blame.135 In 2009, according to an Air Force spokesman, a control room temporarily lost contact with Reaper and Predator drones at an American Air Force command base after a cat wandered in and “fried everything.”136


The fact that machine evidence might implicate black box dangers does not necessarily mean it should be excluded or even subject to special safeguards. It may be that for a particular type of conveyance, the likelihood that black box dangers would both exist andbe discounted by the jury is low, and that the cost of exclusion or production of further contextual information is too high. The goal of this Article is not to allow opponents of machine evidence to capitalize on the cachet of labels like “credibility” in arguing for broad exclusion of potentially risky machine conveyances.137 Rather, it is to force lawmakers, scholars, courts, and litigants to recognize that some machine sources will likely benefit from at least some of the credibility-testing mechanisms we use in the human context, for some of the same reasons that human sources benefit from such testing.

II. a taxonomy of machine evidence

Armed with the black box dangers framework, this Part explores which machine acts and utterances implicate credibility, and how courts have attempted to regulate them. As it turns out, courts, scholars, and litigants have often implicitly recognized that some machines do what witnesses do: they make claims relied upon by factfinders for their truth. But these intuitions have not translated into a systematic regime of machine credibility testing.

A. Machine Evidence Not Dependent on Credibility

Some human acts and utterances do not implicate the credibility of the actor or speaker. Evidence that a defendant was having an affair might be offered as circumstantial proof of a motive to kill his wife. A party may offer evidence that a person said “it’s cold out here” after an accident merely to prove the person was conscious and able to speak at the time of the statement, and not to prove that the temperature was actually low. These acts and utterances do not implicate the sincerity, articulateness, memory, or perception of the human actor. Instead, they are essentially akin to physical objects, whose mere existence the proponent invokes in persuading the factfinder to draw a particular inference.

Like human acts and utterances, machine testimony does not always raise concerns about the credibility of the machine source itself. Machine evidence does not implicate the black box dangers—the testimonial infirmities of machine sources—when the machine acts simply as a conduit for the assertions of others; when it simply performs an act that facilitates scientific testing; or when its conveyance is offered for a purpose other than its truth.

Critically, it is not the complexity or type of machine that determines whether machine evidence implicates credibility. The most opaque, complex, biased, manipulable machine imaginable might produce evidence that is not dependent on credibility, for example, if the evidence is a printout offered simply to show that the machine’s ink cartridge was functioning at the time. Likewise, proprietary email software that simply offers a platform for the emailed assertions of human communicators, themselves offered for their truth, does not implicate black box dangers simply because it is proprietary. These types of machine evidence might affect decisional accuracy by implicating authenticity concerns, requiring proof that the machine result is what the proponent says it is—an email actually written by Aunt Mary, or a printout from a particular machine. But they do not implicate black box concerns.

1. Machines as Conduits for the Assertions of Others

Some machines act as “conduits” for the assertions of people, and thus do not implicate the black box dangers.138 For example, if I write my friend an email stating that “John ran the red light,” and a party introduces my email in a civil suit as proof that John ran the red light, the assertion is mine, not the machine’s. The same logic would apply to tape recorders and dictographs, text messages, website or social media content, and any “electronically stored information” (ESI),139 such as databases listing entries made by employees.140

The line between a machine conduit and a machine source implicating black box dangers is not necessarily a bright one.141 For example, automatic transcription services such as Google Voice can be “extremely inaccurate” under certain conditions, such as when a speaker has a heavy accent.142 Google Voice might therefore raise the specter of analytical error, and thus might require credibility testing, in a way that a tape recorder does not. The ability of email, internet content, or a tape recording to be manipulated, however, does not render the resulting product the conveyance of a machine source rather than a conduit. Rather, the doctored information would be akin to a doctored transcript or fabricated physical object. The admission of such evidence may turn on authenticating whether the human declarant actually made the statement, but it raises no novel issue of machine credibility.143And, usually, a proponent of ESI is required to authenticate the information by showing the input and recording process was regular.144 Authentication ensures that the computer faithfully rendered another person’s assertion. The person’s assertion itself, of course, is subject to all the usual safeguards that apply to human testimony.

2. Machines as Tools

Machine evidence also does not implicate black box dangers when offered to show that human witnesses used the machines as tools to facilitate their observations. Examples might include a laser that facilitates latent fingerprint collection or bloodstain pattern recognition; a magnifying glass or reading light that facilitates handwriting analysis; a gas chromatograph that facilitates the separation of a substance that can then be analyzed by a human or mass spectrometer; and a machine that takes a small amount of DNA and, through repeated heating and cooling cycles, makes millions of copies of the DNA to facilitate later testing.145 Machine tools are analogous to human laboratory technicians who maintain and operate equipment, or who otherwise offer assistance during testing. Of course, human technicians might deliberately tamper with results in a way that machines would not, unless programmed to do so. But the technicians’ actions, while consequential, are not treated as credibility-dependent assertions under hearsay law.

Like the actions of human technicians who facilitate testing, the actions of machine tools are different from machine and human sources that convey information. Instead, the actions of machine tools are akin to physical objects or natural processes.146 A gun or tape recording cannot be “impeached” because they make no claims; they are authenticated, and then offered to the jury for what they are worth. The same is true for evidence of a machine action offered simply to prove the machine committed a certain act, such as mixing two substances together. The act is relevant for whatever inferences can be directly drawn from it. Similarly, where a machine tool merely illuminates facts for a human observer, the observation and report relied upon by the factfinder is ultimately that of the human witness, not of the machine.147 A human expert might make a mistake, of course: “[m]icroscopic studies” require “‘a sincere Hand, and a faithful eye’ to examine and record ‘the things themselves as they appear.’”148 Those who criticize microscopic hair analysis as a means of forensic identification do so on the grounds that examiners suffer cognitive bias and lack any probabilistic basis for determining the probative value of an alleged match,149 not on grounds that the microscope itself has made an underscrutinized claim.

In contrast, the opinion of a human expert—or an expert system—can be impeached, and the opponent should have the chance to do so.150 There are difficult cases at the margins, where the difference between a machine tool facilitating human observation and a machine source engaging in its own observation is subtle. A thermal imaging device, for example, while in one sense an object facilitating human observation, is also an observer and interpreter itself, within the confines of its design and inputs. The device’s own credibility—whether its conveyance might be false by design, inarticulate, or analytically unsound—is implicated.

3. Machine Conveyances Offered for a Purpose Other than Proving the Truth of the Matter Conveyed

A machine’s act or utterance, even if explicitly conveying a claim, does not implicate black box dangers if it is not offered to prove the truth of the claim. In the human context, an act or utterance not offered for its truth does not implicate the so-called “hearsay dangers” (and thus, even if made out of court, does not implicate the hearsay rule) because the inference to be drawn by the factfinder does not “involv[e] a ‘trip’ into the head of the person responsible . . . .”151 In the same respect, when a jury can draw the requested inference from a machine act or utterance with no trip into and out of the machine’s analytical process, the machine’s believability is not at stake.

For example, if a machine’s printout were offered merely to prove that the machine’s ink toner was functional at the time of printing, then the evidence would not pose a black box problem. The printout is nothing more than a physical object, which the factfinder observes and from whose mere existence the factfinder can draw the proponent’s requested inference. Similar logic would apply to statements sent by FBI malware to computers suspected of having visited certain illegal websites, offered not for their truth but to show that the computers then sent information back to the FBI.152 Likewise with evidence in a fraud case that a red light camera programmer has chosen an unreasonably short “grace period” to generate revenue for the city.153 The probative value of the statement stems not from its “communicable content,” but from its “perceptual content.”154

B. Machine Evidence Dependent on Credibility

This Section explores how courts, scholars, and litigants have historically treated machine evidence that does implicate credibility; that is, machines whose acts and utterances are offered for the truth of some claim they convey in a way that implicates the black box dangers. Even as these groups appear to recognize the “testimony”-like nature of certain machine evidence, these episodes of recognition have never converged to form a single coherent doctrine of machine testimony. Instead, lawmakers have dealt with machine sources through a patchwork of ill-fitting hearsay exceptions, confusing authenticity rules, and promising but inadequate reliability requirements for expert methodologies.

As this Section also explains, machine sources that implicate credibility vary in their characteristics: some are simple, some are complex; some are transparent, some are opaque; some are highly stable, while others are highly sensitive to degradation or human input and operation errors; and some are created in anticipation of litigation, while others have a nonlitigative public or commercial use. Some machine sources convey images, while others explicitly convey information through symbolic output. These characteristics may determine whether a machine source should be subject to particular safeguards, but even the simplest, most transparent, most stable, and most regularly made instrument is a “source” if its output depends on credibility for its probative value.

1. “Silent Witnesses” Conveying Images

When offered as proof of an event or condition they purport to have captured, photographs and films implicate the testimonial capacities of the camera itself, as influenced, of course, by human choices. Jennifer Mnookin, in her exploration of the socio-legal history of the photograph, notes the many courts and commentators who referred to the photograph in its early days in testimonial terms: a “sworn witness,”155 a “dumb witness,”156 a “mirror with a memory,” in the words of Oliver Wendell Holmes,157 and even—to the skeptics—a “most dangerous perjurer,”158 a witness that, because it cannot be cross-examined, “may testify falsely with impunity.”159 Again, these descriptions were not simply metaphor. They reflected a qualitative difference between photographs and mere physical evidence:

[P]hotographs, unlike murder weapons, . . . tell a story about the world, making a difficult-to-refute claim about how a particular location looked at one instant . . . . [T]o whatever extent this visual depiction is not tied to testimony, a competing, nonverbal account enters a space where the words of witnesses—and lawyers—are supposed to reign.160

John Henry Wigmore similarly described the x-ray machine as a conveyor of information, one that “may give correct knowledge, though the user may neither have seen the object with his own eyes nor have made the calculations and adjustments on which the machine’s trustworthiness depends.”161 Tal Golan describes the x-ray and other visual evidence as the emblem of a new class of “machine-made testimonies” of the late nineteenth century.162 Others have more explicitly argued that filmic evidence is inherently “testimonial”163 and “assertive in nature,”164 and have, in passing, analogized film to hearsay in arguing that its assertions potentially exhibit insincerity, misperception, and ambiguity.165

The camera is a relatively simple machine, in terms of its physical form and internal processes. But because photography is highly sensitive to human input and human bias, photographic evidence can easily mislead a factfinder. A cameraperson might intentionally or through unconscious bias166 choose a lens, filter, or angle to make a suspect look more sinister167 or guilty,168 make a wound seem deeper or shallower,169 or make a distance seem greater or smaller.170 Moreover, photographs and film do not provide factfinders with full context. Key aspects of an event or condition might be missed or obscured because of poor sound or visual quality of an image or film,171 potentially leading a factfinder to draw improper inferences. For day-in-the-life videos and other filmic evidence created expressly for litigation, the motivation for biased representation of facts—such as increasing a film speed to make a disabled subject look less injured172—might be particularly high. Photographs can also be modified or fabricated, just like any other physical object. After capturing an image, a photographer may choose to “reverse the negative” so that the right side of the photograph appears on the left side.173 But these possibilities of post-hoc human manipulation pose problems for authenticity, not credibility.

Courts’ treatment of photographic evidence reflects both a promising intuition that black box dangers exist and an unfortunate failure of imagination in fully regulating photographs as credibility-dependent evidence. In photography’s early days, courts admitted photographs only if the photographer testified about the process and certified the image’s accuracy.174 This rule addressed a fear that the public would view photographic images as infallible even as they proved highly manipulable.175 When requiring the photographer’s testimony became unsustainable, courts used a different tactic: they labeled the photograph as merely “demonstrative” of a witness’s testimony about an event, rather than as substantive evidence in its own right, thereby “demot[ing] the photograph from the nearly irrefutable to the merely illustrative.”176

But that fiction eventually collapsed as well. Photographs are now, along with films and x-ray images, “readily accept[ed]”177 in most American jurisdictions without an accompanying human witness, under a so-called “silent witness” theory.178 In any event, many photographic systems—such as surveillance cameras, red light cameras, and ATM video footage—are now automatic and collect images without a person behind the camera. Courts still require authentication to prove the photograph depicts what the proponent says it depicts, but such proof can typically be from the photograph alone179 under the theory that it “speaks for itself.”180 Because photographs are considered neither mere appendages to human testimony nor “testimony” under the law of evidence, they are caught in a netherworld along with other machine conveyances and underscrutinized for the presence of black box dangers.181

2. Basic Scientific Instruments

For well over a century, courts have implicitly acknowledged the credibility-dependent nature of the measurements of instruments, basing their admission on characteristics likely to minimize black box dangers. By the mid-nineteenth century, there existed a “public depot of scientific instruments” for “commercial” and “nautical purposes.”182 Many such instruments made their way into English and American courtrooms, including clocks, watches, thermometers, barometers, pedometers, wind speed measures, and “a variety of other ingenious contrivances for detecting different matters.”183 Although litigants would occasionally insist that an instrument’s measurement was inaccurate,184 courts afforded scientific instruments a presumption of correctness “akin to” the usual course of business hearsay exception for mercantile records offered for their truth.185 John Henry Wigmore, in his influential 1904 treatise, placed his discussion of “scientific instruments” under the rubric of hearsay rather than physical evidence, noting that the accuracy of instruments’ conveyances depends on the credibility of others:

The use of scientific instruments, apparatus, and calculating-tables, involves to some extent a dependence on the statements of other persons, even of anonymous observers. Yet, on the one hand, it is not feasible for the scientific man to test every instrument himself; while, on the other hand, he finds that practically the standard methods are sufficiently to be trusted . . . . The adequacy of knowledge thus gained is recognized for a variety of standard instruments.186

The 1899 edition of Simon Greenleaf’s evidence treatise similarly discussed scientific instruments in the hearsay context: in noting that “an element of hearsay may enter into a person’s sources of belief,” he used examples such as “reckoning by a counting-machine.”187 Relying on instruments’ regular production and use, modern courts often take “judicial notice” of their readings, without further foundation, on grounds that their accuracy is beyond reasonable dispute.188

As the presumption of correctness reflects, most basic scientific instruments are simple, transparent in terms of their design and process, not sensitive to human input error, and regularly made for a nonlitigative purpose. Yet as Part I made clear, even well-designed, simple, transparent instruments are still susceptible to errors of articulation when they have old batteries or worn markings, or to inferential errors based on an operator’s placement decision. And for some instruments, like the sextant, accurate output largely depends on inputter and operator skill.189 These potential flaws do not suggest the measurements of such instruments should be excluded, but that lawmakers should consider which operation protocols, impeachment mechanisms, and other safeguards sufficiently empower jurors to assess instrument credibility.

3. Computerized Business Records

Beginning in the mid-1960s, American courts faced litigation over the admissibility of computer records kept or created in the regular course of business.190While some computer records were “conduits” storing data inputted by humans, others were information generated by the computer as a source.191 Most federal courts did not observe this distinction and instead, by the mid-1990s, treated all computer records as requiring a foundation under a hearsay exception,192 perhaps bolstered by an oft-cited 1974 article opining that “[c]omputer-generated evidence will inevitably be hearsay.”193 These courts, like nineteenth-century courts facing the measurements of instruments, were rightly concerned about the sensitivity of computer-generated records to human design, input, and operator error. But their insistence upon regulating such records through a hearsay model had little grounding in law or logic.194

As more courts recognize both that computer-generated information is not hearsay and that it might still be inaccurate, some have looked to Federal Rule 901(b)(9) and its state analogs195 for guidance. A provision in Rule 901(b)(9) allows proponents to authenticate the results of a “process or system” by “describing [the] process or system” used to produce the result and showing it “produces an accurate result.”196 The rule, proposed in 1968, responded directly to computerized business records.197 Its original language provided, much like other traditional authentication rules, that a proponent prove that a system result “fairly represents or reproduces the facts which the process or system purports to represent, or reproduce.”198 But Judge Weinstein suggested adding the word “accurate” to the language,199 meaning that proponents of machine processes now have a choice between authenticating the result through proof that the process produces an accurate result and authenticating it through other means. For computerized business records, the authentication requirement of Rule 901(b)(9) may screen clearly unreliable processes,200 although as Part III makes clear, such records—like all machine conveyances—should also be open to impeachment and other scrutiny that provides the factfinder with additional context.

4. Litigation-Related Gadgetry and Software

Unlike the measurements of basic instruments and computerized business records, some machine-generated data are created specifically for civil or criminal litigation, motivating humans to design, input, and operate the machine to produce information favorable to the proponent.

A concern about this type of litigation-related bias—and an influential 1976 Second Circuit dissenting opinion expressing such concern—may have influenced courts from the 1970s to the early 2000s to treat computer-generated records as hearsay.201 In a contract dispute between an inventor and a patent assignee, Singer Company, the inventor offered the conclusion of a proprietary computer program that an anti-skid automotive technology was capable of being perfected by Singer for sale.202 Singer claimed the inventor’s refusal to disclose the “underlying data and theorems employed in these simulations in advance of trial” left the company without a fair and adequate opportunity to cross-examine the inventor’s expert witnesses.203 The majority concluded Singer had enough fodder to cross-examine the experts who relied on the program, without learning more about the program itself.204

In dissent, Judge Van Graafeiland declared he was “not prepared to accept the product of a computer as the equivalent of Holy Writ.”205 Instead, “[w]here . . . a computer is programmed to produce information specifically for purposes of litigation,” the product should be subject to greater scrutiny.206 He suggested that the party introducing such information should have to disclose the computer “program” before trial to the opposing party, so that the party has the “opportunity to examine and test the inputs, program and outputs prior to trial.”207 Ultimately, the judge insisted, where a party’s “entire case rests upon the accuracy of its computerized calculations,” judges should “subject such computations to the searching light of full adversary examination.”208 A court in the 1980s similarly admitted a program called “Applecrash,” which estimated the likely speed of a car during a collision,209 rejecting the opponent’s arguments for pretrial disclosure of the program’s processes on grounds that cross-examination of the human expert who relied on the program was sufficient.210

More recently, courts have ruled on the reliability of litigation-related, computer-generated conclusions that form the basis of human expert testimony. Courts tend to admit such evidence so long as validation studies prove the reliability or general acceptance (under Daubert or Frye, respectively) of the program’s methodology and the opponent can cross-examine the human expert.211 In Part III, I explore the limitations of existing reliability-based admissibility rules as a means of testing machine credibility. Meanwhile, courts have admitted other nonscientific algorithms with no Daubert scrutiny at all.212

In criminal cases, courts have also tended to subject the conveyances of gadgets and software created for law enforcement purposes to reliability tests,213 but have routinely admitted them. In the early twentieth century, courts faced a wave of fact-detecting, “ingeniously contrived”214 gadgets. The Harvard Law Review published a note in 1939 titled “Scientific Gadgets in the Law of Evidence,” chronicling ABO typing, blood-alcohol testing, deception tests, filmic evidence, and fingerprint and ballistic analysis.215 One scholar wrote in 1953 that the “whole psychological tone” of the new “scientific age” of the early twentieth century “embodie[d] an increasing reliance on gadgets.”216 Some of these gadgets explicitly conveyed information in symbolic output, and some were made expressly for law enforcement purposes. For example, radar guns and tachometers recorded car speed and were soon used in traffic prosecutions.217 The Drunk-O-Meter, unveiled in 1938, and the Breathalyzer, unveiled in 1954, recorded the concentration of alcohol in deep lung air.218

While these analog gadgets were often uncomplicated in their construction,219 they were sometimes maligned by judges and commentators in language suggesting concerns about black box dangers. The Breathalyzer, for example, was derided as “Dial-a-Drunk” because it forced a police officer to manually set a baseline before testing a suspect.220 And judges, apparently expressing concern over the black box opacity of certain guilt-detecting gadgets, warned that both the radar gun and the Breathalyzer might usher in an era of “push button justice.”221 Courts still occasionally reject a particular gadget or program as being unreliable enough to exclude under Frye or Daubert,222 but such cases are few and far between, particularly now that breath-alcohol testing is subject to so many front-end safeguards. Many states now limit the type of machines that can be used and enforce operation protocols to ensure accurate results.223

In subsequent decades, these gadgets have shifted from analog to digital forms, reducing certain aspects of their manipulability, but exhibiting a “creeping concealedness” in their opacity and complexity.224 While the Drunk-O-Meter required a human to do the arithmetic necessary to translate its color test and scale-measured breath weight into blood-alcohol content,225 modern breath-alcohol tests based on infrared and fuel cell technology offer a print-out report or digital screen reading.226 Radar gun software and output,227 as well as infrared spectrometers and gas chromatographs reporting drug levels in blood,228 have also graduated to digitized, software-driven forms.

A number of other modern computer-driven sources of information, built in anticipation of law enforcement use, now exist, including stingray devices that can record incoming and outgoing phone numbers to a cell phone;229 license plate readers;230 graphs of DNA test runs, purporting to show which genetic markers or “alleles” are present in a sample;231 red-light camera time-stamp data;232 address logs purporting to list IP addresses of users who have visited child pornography websites;233 database-driven computer reports of the closest handful of matching archived records to an inputted latent print or ballistic image from a crime scene;234 machine-learning crime-detecting programs;235 drug identification software that can identify particular cutting agents used, which might lead investigators to a particular dealer;236 and arson investigation software that offers an “answer” to whether debris suggests arson.237 And a number of software programs now exist that offer a “score,” based on several inputted variables, that represents the subject’s future dangerousness for purposes of criminal sentencing, parole, and civil commitment determinations.238 Most of these programs are proprietary.239

In particular, complex proprietary software has dramatically affected criminal cases involving DNA mixture interpretation. DNA has revolutionized criminal trials and is now ubiquitous as a means of forensic identification.240 But while some DNA samples comprise a large amount of a single person’s DNA and are relatively easy to analyze, other samples contain mixed, low-quantity, or degraded DNA. Drawing inferences about the number and identity of contributors in such complex mixtures is a difficult business. As one DNA expert noted, “[I]f you show ten colleagues a mixture, you will probably end up with ten different answers.”241 Recognizing the inherent limitations of manual methods,242 several companies now offer probabilistic genotyping software purporting to enhance the objectivity and accuracy of DNA mixture interpretation by automating the process both of calling “matches” and of generating a match statistic that explains the match’s significance—that is, how many people in the population would have a DNA profile consistent with the mixture purely by chance. As one program designer put it, we now have a “computer that interprets DNA evidence.”243 These systems differ in terms of the assumptions embedded in their source code and the form their reported match statistics take.244 Some developers have opened their source code to the public;245others, such as Cybergenetics’s “TrueAllele” program and New Zealand DNA expert John Buckleton’s “STRmix,” have not.246 Courts have nearly universally admitted the results of these programs over objection in Frye/Daubert litigation,247 and in at least one case, a defendant used results to convince prosecutors to support vacating his conviction.248

In one recent case, two expert DNA systems returned contradictory results based on the same factual input. In 2011, a twelve-year-old boy in Potsdam, New York was tragically strangled to death in an apartment he shared with his mother. Police suspicion fell upon Nick Hillary, a former college soccer coach who had dated the mother and who was upset about their breakup a few months earlier.249 Another former boyfriend, a deputy sheriff who had been physically violent with the mother, was cleared of suspicion based on a video showing him walking a dog several blocks away minutes before the incident. Rumors that another child may have killed the boy were also dismissed by police early on.250 Focusing on Hillary, police surreptitiously took his DNA from a coffee cup and the butt of a cigarette and compared it to dozens of samples from the scene and the boy’s body and clothing, with no resulting match.251 Nor did any DNA samples taken from Hillary’s car, home, or clothing match the boy’s DNA. But analysts could not determine whether Hillary might be a contributor to a DNA mixture found under the boy’s fingernail. Seeking a more definitive opinion, police in 2013 sent the DNA data to Mark Perlin, the creator of “TrueAllele.” In 2014, Perlin reported that “[t]he TrueAllele computer found no statistical support for a match” with Hillary.252 A year later, a new district attorney—elected on a promise to find the killer253—had the DNA data analyzed through STRmix, which reported that Hillary was 300,000 times more likely than a random person to have contributed to the mixture.254In September 2016, a trial judge excluded the STRmix results under Frye,255 and Hillary was subsequently acquitted.256

5. Other Complex Algorithms, Robots, and Advanced Artificial Intelligence

A host of other types of machine conveyances are routinely offered for their truth in court, sometimes to prove a criminal defendant’s guilt. Many of these conveyances come from machines created for general purposes, not for litigation, and many of those machines are driven by proprietary software. Common examples include Event Data Record information;257 automated telephone responses giving telephone number information;258 Google Earth satellite imagery and GPS coordinates;259 software-generated driving time estimates;260 “Find my iPhone” features used to track phone theft;261 and Fitbit data used to impeach an alleged rape victim’s claim about being asleep at the time of an attack.262 Other expert systems are now available and seem capable of being offered as evidence, such as those rendering medical diagnoses263 and automated language analysis,264 and mobile facial recognition technology and goggles offering real-time information about observed subjects.265

Perhaps the final frontier in law’s reliance on machine conveyances of information is the full automation of the act of witnessing. The jump from having an expert system render an opinion to having a robot266 or android deliver that opinion to a jury face-to-face does not seem particularly fanciful. As one blogger asked, “[I]s it far-fetched to imagine Watson’s now-familiar blue avatar someday sitting on the witness stand?”267 Even IBM’s senior vice president for legal and regulatory affairs has suggested that Watson might have a place in the courtroom as a real-time fact checker.268 And at least one legal scholar has suggested that artificial intelligence play the role of a court-appointed witness under Federal Rule of Evidence 706 in giving counsel to judges during Frye/Daubert hearings.269 Likewise, “robot police”270 and robot security guards271 are already in use and could presumably offer information, in a suppression hearing or at trial, about a suspect’s observed behavior.

Whether created for litigation or general purpose, these complex systems raise accuracy issues not adequately addressed by existing evidence law. The only clear legal rules that apply to them are basic rules of relevance and undue prejudice, authentication rules like Federal Rule 901(b)(9) requiring that a process produce an accurate result, and Daubert-Frye reliability requirements for human expert testimony. But as machine conveyances become ever more sophisticated and relied upon, factfinders need more information and context to assess machine credibility.

III. testimonial safeguards for machines

This Part offers a brief vision of new testimonial safeguards built for machine sources of information. It first considers credibility-testing mechanisms that the law of evidence could adopt, and then considers whether accusatory machine conveyances in criminal cases might implicate the dignitary and accuracy concerns underlying the Confrontation Clause.

A. Machine Credibility Testing

The purpose of credibility-testing mechanisms is not primarily to exclude unreliable evidence, but to give jurors the context they need to assess the reliability of evidence and come to the best decision.272 Indeed, in the machine context, a generalized rule of exclusion like the hearsay rule would harm the factfinding process, given the promise of mechanization as a means of combatting the biases of human testimony. With that in mind, this Section explores safeguards that would give jurors more context about a machine conveyance, without necessarily excluding the information as unreliable. In choosing whether to adopt such safeguards, lawmakers must consider issues of cost; efficiency; fairness; the likelihood that, without the safeguard, the jury will draw the wrong inference; and the likelihood that, with the safeguard, the jury will overestimate the value of the impeachment material and undervalue the evidence.

1. Front-End Design, Input, and Operation Protocols

The first means of both improving the accuracy of machine conveyances and producing contextual information helpful to juries is to develop better protocols for design, input, and operation. Front-end protocols are underused but not entirely absent in the context of human testimony: the New Jersey Supreme Court, for example, has recognized a number of front-end protocols that can prevent human bias in stationhouse eyewitness identifications.273 In the machine context, states have imposed protocols most conspicuously for breath-alcohol tests, requiring that testers use an approved machine and follow procedures targeting practices shown to produce ambiguity due to misplacement and input error.274 Such requirements need not be a condition of admission; in the breath-alcohol context, the failure to adhere to testing and operation protocols goes to weight, not admissibility.275 But breath-alcohol testing is an outlier in this respect, likely for reasons relating to the history of DUI jurisprudence and the political capital of DUI defendants;276 other types of forensic testing are not yet regulated by such a regime of detailed state-imposed protocols.

Generally, the more complex, opaque, and litigation-driven a machine’s processes, the more design protocols are helpful. First, it is difficult for the jury, through a facial examination of the assertion and through mere questioning of the source itself or herself, to determine the assertion’s accuracy: protocols help here for the same reason they are helpful in the stationhouse eyewitness identification process. Second, and putting litigative motive aside, the chance for inadvertent miscodes or analytical overreaching will be greater in machines that are highly complex or that attempt to model complexity, like self-driving car technology or Google Earth.

A jurisdiction might therefore require any software-driven system used in litigation to be certified as having followed software industry standards in design and testing. Though these standards are readily available,277 programmers typically do not adhere to them in designing litigation-related software and courts and legislatures do not use them as a condition of admission. One software expert affirmed that STRmix, a probabilistic genotyping program, had not been rigorously tested according to industry standards,278 and the program’s creators have had to disclose publicly multiple episodes of miscodes potentially affecting match statistics.279 Critical errors were also found during review of source code in litigation over the Toyota Camry’s unintentional acceleration problem.280 A software expert who reviewed the source code of the “Alcotest 7110,” a breath-alcohol machine used in New Jersey, found that the code would not pass industry standards for software development and testing. He documented 19,500 errors, nine of which he believed “could ultimately [a]ffect the breath alcohol reading.”281 A reviewing court found that such errors were not a reason to exclude results, in part because the expert could not say with “reasonable certainty” that the errors manifested in a false reading,282 but the New Jersey Supreme Court did cite the errors in requiring modifications of the program for future use.283 Exclusion aside, a more robust software testing requirement reduces the chance of misleading or false machine conveyances.

Even where software is well written to operationalize the intended method, the method itself might be biased in ways that could be avoided if the design process were less opaque. One scholar has advocated what he terms “adversarial design,”284 a means of building models that itself is political, reflecting normative controversies and compromises. If the process of developing risk assessment tools, credit score algorithms, or genotyping software were itself more adversarial, with input from all sides of contentious debates, we would presumably see less tolerance for analytical biases and fewer variables that correlate to race.285 Because of extant biases and racial variables, courts and legislatures should consider requiring that software used in criminal trials and sentencings be publicly designed and open-source. Experts have proposed similar public solutions to other black box scenarios, such as the credit scoring system.286 Public models would have the benefit of being “transparent” and “continuously updated, with both the assumptions and the conclusions clear for all to see.”287

When algorithms are privately developed, a public advisory committee could still promulgate requirements related to key variables or assumptions. For example, programmers of probabilistic genotyping software should not be the ones to choose the level of uncertainty that prompts a system to declare a DNA mixture “inconclusive” as opposed to declaring someone a potential contributor,288 or to choose their own estimate related to the frequency of certain phenomena, such as genetic markers or allelic drop-out. Developing such guidelines for the substance and scope of machine testimony would be analogous to the National Commission on Forensic Science’s recent call for human experts to cease using the phrase “reasonable degree of scientific certainty.”289

Programs that use machine-learning techniques might require their own set of protocols to promote accuracy. Data scientists have developed very different “evaluation metrics” to test the performance of machine-learning models depending on the potential problem being addressed. For example, testers might use a technique called “hold-out validation” to determine whether a “training set” of data used at the beginning of supervised learning is an appropriate set on which to train the machine.290

Beyond design, input and operation protocols may be important for machines particularly sensitive to case-specific human errors, from sextants to breath-testing devices. One means of encouraging and checking proper calibration is to require quality control and quality assurance logs, a practice currently part of most laboratory work. In the breath-testing context, the test results from each machine are automatically recorded and transmitted to an online data center, maintained and reviewed by the state.291 In the context of entering GPS coordinates into Google Earth, like the officer in Lizarraga-Tirado, one could imagine documentation requirements as well. Another check on inputs and operation would be to allow an opponent’s representative to be present for case-specific inputs and operation of a machine.

2. Pretrial Disclosure and Access

A number of pretrial disclosure and access rules already apply to human testimony. If the United States intends to use expert testimony in a criminal trial, it must disclose the qualifications of the expert and the bases and reasons for her testimony at the defendant’s request.292 The disclosure requirements in civil trials are even more onerous, requiring the expert to prepare a written report that includes the facts or data relied on.293 Proponents must not discourage witnesses from speaking with the opponent before trial,294 and in criminal trials, proponents must also disclose certain prior statements, or “Jencks material,” of their witnesses after they testify.295 These requirements offer notice of claims that might require preparation to rebut, the ability to speak with the witness before trial, and the ability to review prior statements for potential impeachment material.

Applying these principles to machine sources, a jurisdiction might require the proponent of a machine “expert”—a source that generates and conveys information helpful to the jury and beyond the jury’s knowledge—to disclose the substance and basis of the machine’s conclusion. As one DNA statistics expert told me, “I just want these expert systems to be subject to the same requirements as I am.” A jurisdiction might therefore require access to the machine’s source code, if a review of the code were deemed necessary to prepare a rebuttal of the machine’s claims.

Creators of proprietary algorithms typically argue that the source code is a trade secret or that it is unnecessary to prepare a defense to the machine’s conclusion so long as the opponent understands the “basic principles” underlying the machine’s methods.296 But it is not clear that trade secret doctrine would protect the source code of an algorithm used to convict or impose liability.297 Moreover, validity of method and validity of software-driven implementation of method are not equivalent; as one group of researchers has argued, “[c]ommon implementation errors in programs . . . can be difficult to detect without access to source code.”298

A jurisdiction might also require meaningful access to the machine before trial, so the opponent can both review the machine’s code, if it is disclosed, and also input different assumptions and parameters into the machine—for example, those consistent with the opponent’s theory of the case—to see what the machine then reports. TrueAllele offers access to its program to criminal defendants, with certain restrictions, but only for a limited time and without the source code.299 This sort of “black box tinkering” allows users to “confront” the code “with different scenarios,” thus “reveal[ing] the blueprints of its decision-making process,”300 but it also approximates the process of posing a hypothetical to an expert for purposes of preparing cross-examination related to the opponent’s theory. Indeed, the ability to tinker might be just as important as access to source code. Data science scholars have written about the limits of transparency301 and the promise of “reverse engineering” in understanding how inputs relate to outputs,302 as well as the benefits of “crowdsourcing”303 and “[r]uthless public scrutiny”304 as means of testing models and algorithms for hidden biases and errors.

A jurisdiction could also require disclosure of “Jencks material” for machine sources.305 If a party takes several photographs of an accident scene with different lenses and camera angles and cherry picks the best one to present in court, the remaining photographs should be disclosable as Jencks material of the camera. Similarly, the prosecution using probabilistic DNA software might be required to disclose the results of all prior runs of a machine of a particular sample under various assumptions and parameters.306Or consider a criminal case in which investigators find a latent fingerprint at a crime scene and run it through the federal fingerprint database system, which reports the top ten matching prints and allows a human analyst to declare if any is a likely match.307 State officials generally refuse defense requests for access to the other reported near matches, notwithstanding arguments that these matches might prove exculpatory.308

Likewise, a breath-alcohol machine’s COBRA data, which has been helpful in unearthing errors with such machines,309 might be more clearly disclosable and admissible for impeachment if the machine were treated as a witness. In a somewhat analogous case, the defendant in a 1975 tax fraud prosecution sought access to the IRS’s computer system’s previous reported lists of nonfilers, to determine whether any previous records were mistaken. The court did not dismiss the request out of hand, but ruled that the defendant had sufficient alternative means of testing the computer’s accuracy, including his own expert’s pretrial access to the IRS’s data processing systems.310

3. Authentication and Reliability Requirements for Admissibility

Just as certain categories of human sources are particularly “risky,”311 certain machine sources might be more risky than others because of their complexity, opacity, malleability, or partiality of purpose. Should a broad reliability-based rule of exclusion—akin to the hearsay rule—apply to machine conveyances that exhibit some combination of these traits? This Article does not advocate such a rule. The characteristics of machine conveyances do not lend themselves to categorical exclusion based on the lack of a particular characteristic or safeguard. While the hearsay rule focuses exclusively on human assertions rendered out of court, a categorical rule of exclusion for machines that focused on a particular level of complexity, opacity, manipulability, or litigative purpose would be difficult to draft and dramatically over- or underinclusive in terms of targeting truly “risky” machines. Even complex, opaque algorithms—like Google Earth—can offer highly probative, relatively accurate information that presumably should not be excluded from all trials simply because opponents lack access to, say, the source code. Indeed, a proponent of Google Earth results might reasonably be concerned that jurors will undervalue such results based on an opponent’s speculative or anecdote-based arguments about Google’s unreliability. Moreover, the hearsay rule itself is highly criticized and lacking in empirical foundation.312

Some countries do, in fact, have admissibility requirements for machine-generated reports of information, but these requirements are limited. In the United Kingdom, a “representation . . . made otherwise than by a person” that “depends for its accuracy on information supplied (directly or indirectly) by a person” is not admissible in criminal cases without proof that “the information was accurate.”313 But computer evidence in the United Kingdom is otherwise presumed, “in the absence of evidence to the contrary,” to be “in order,” and commentators have lamented the inability to meaningfully rebut software-generated conclusions.314Still other countries rely mostly on judicial discretion in determining the accuracy of machine conveyances,315 or allow such evidence so long as it is accompanied by a human expert.316

Of course, authentication rules should apply to machine sources: if output purports to be that of a particular machine, the jury should be able to rely on it as such. But authentication rules do not generally address the credibility or accuracy of a source.317 As discussed in Part II, federal authentication rules and state analogs include a provision targeted at the type of computerized business records existing in 1968, allowing authentication of a result of a process or system by showing that the system produces an accurate result. But even this rule is not by its terms an accuracy requirement; it is simply one allowable means of authentication among many for computerized evidence.318

To the extent some courts have interpreted Federal Rule 901(b)(9) as requiring proof that any result of a mechanical process be “accurate” as a condition of admission, they have done so largely within the realm of computer simulations offering “expert” opinions, importing a Daubert-like reliability analysis.319 I turn to this sort of reliability requirement for expert machines next. But it is worth noting that a general accuracy requirement, along the lines of 901(b)(9) or Daubert, might also be adopted to screen out unreliable machine processes that are not “expert,” such as the lay observations of a poorly programmed robot security guard.

Rules requiring the scientific or technical methods of expert witnesses to be reliable and reliably applied should also extend to machine sources, at least those whose conveyances relate to matters beyond the ken of the jury.320 Daubert and Frye technically do not apply to machine conclusions admitted without an accompanying human witness, although they could be modified to do so. Under current law, courts treat the machine as the method of a human expert, rather than as the expert itself, even when the expert is a “mere scrivener” for the machine’s output.321 As a result, any scrutiny of the machine’s conclusion through Daubert-Frye comes through pretrial disclosure of the basis of the human expert’s testimony, the pretrial admissibility hearing, and cross-examination of the human expert at trial. The machine itself is not subject to pretrial disclosure rules or impeachment, or any scrutiny equivalent to cross-examination.

A rule requiring that the machine itself follow a reliable, and reliably applied, method for reaching its conclusions would involve more scrutiny than a typical Daubert-Frye hearing currently offers. Most judges rely heavily on validation studies in concluding that a machine, whether it be the Intoxilyzer 8000 or TrueAllele, uses a generally reliable process to reach its result.322 But validation studies alone, showing a low false positive rate or an expected relationship between input and output,323 might be an inadequate basis upon which to declare a machine conveyance likely accurate. Predictive algorithms, for example, might suffer feedback loops that taint performance evaluation.324 In the forensic identification context, a machine might be assumed reliable because its conveyances have not been proven to have ever led to a wrongful conviction, a problematic metric given the difficulty in proof.325 Validation studies are also often conducted under idealized conditions unrepresentative of the challenges of real casework. In the DNA mixture context, precisely those mixtures deemed too challenging to resolve manually because of degradation or other issues are relegated to software to solve. Some software designers embrace this state of affairs; TrueAllele advertises that the company “always giv[es] an answer,” even in the “most challenging” mixtures.326 As one expert warned, “TrueAllele is being used on the most dangerous, least information-rich samples you encounter.”327

Because of its limitations, validation is a potentially incomplete method of ensuring the accuracy of machine reports in the form of statistical estimates and predictive scores:

Laboratory procedures to measure a physical quantity such as a concentration can be validated by showing that the measured concentration consistently lies with an acceptable range of error relative to the true concentration. Such validation is infeasible for software aimed at computing a[] [likelihood ratio] because it has no underlying true value (no equivalent to a true concentration exists). The [likelihood ratio] expresses our uncertainty about an unknown event and depends on modeling assumptions that cannot be precisely verified in the context of noisy [crime scene profile] data.328

Effective validation studies would help determine whether a DNA expert system tends to falsely “include” subjects as a contributor to a mixture. But validation studies are much less informative, at least in their current state, for demonstrating how accurately (or inaccurately) a system predicts the likelihood of a subject’s contribution.

Some experts have argued that access to the source code is the only meaningful way to determine whether a complex algorithm’s method is both reliable and reliably applied.329 This argument has intuitive appeal: even if an algorithm’s variables and analytical assumptions are transparent and seemingly valid, the software is the means by which those assumptions are actually implemented by the machine, and should itself be validated.330 Assuming there are no trade secret issues, access to source code seems obvious. On the other hand, transparency alone does not guarantee meaningful scrutiny of software.331 Source code is lengthy; TrueAllele has 170,000 lines of code.332 If opponents (or the public) had unfettered and indefinite access to the software to tinker with it, and if the software were subject to robust front-end development and testing standards, access to the code might not be critical.333 At the very least, software engineers should be deemed part of the “relevant scientific community” for determining whether a method is or is not generally accepted,334 rather than judging the reliability of software based on whether it is “relied on within a community of experts.”335

Notably, the two expert DNA systems that came to a different conclusion in the Hillary case have both been accepted in numerous jurisdictions under both Daubert and Frye. These basic reliability tests, unless modified to more robustly scrutinize the software, simply do not—on their own—offer the jury enough context to choose the more credible system. TrueAllele’s creator recently criticized several aspects of STRmix’s methodology in a strongly-worded letter to the FBI,336 and cited on its website a defense motion in another case calling STRmix “foreign copycat software.”337 But without more information about how each program arrives at its match statistic, the opposing party has few tools to impeach the credibility of that conclusion. The tools for impeachment lie buried in the machine’s black box.

4. Impeachment and Live Testimony

Whether a machine source survives an authenticity or reliability challenge, the opponent should still have an opportunity to impeach the sources credibility at trial. After all, even when an out-of-court human assertion is admitted under a reliability-based hearsay exception, the opponent can still impeach the declarant at trial using any of the declarants prior inconsistent statements, evidence of incapacity or bias, or character for dishonesty.338 Once an opponent has access to the prior statements of a machine, the opponent could likewise impeach the machines credibility, assuming a few modifications in existing impeachment statutes.339

Given the “distributed cognition” between man and technology that underlies machine conveyances, meaningful impeachment of the machine source might also involve scrutiny of the character or capacity of human programmers, inputters, and operators. Evidence that a human programmer has a character for dishonesty, for example, or might harbor bias because he has been paid money to develop a program for a particular litigant, is relevant to the likelihood of deception or bias in the machine’s design.

Trial safeguards would not necessarily involve the live testimony of the programmer, although such a requirement might make sense depending on the black box dangers implicated. The United Kingdom’s rule requiring accuracy of inputs, for example, requires the live testimony of the inputter when a machine representation relies on information provided by that inputter.340 Other countries subject computer-generated conclusions to the hearsay rule if at any point a human intervened in the machine’s processes for creating its record.341 In South Africa, merely signing a document printed by a computer is enough to convert the document to hearsay.342 But treating a machine conveyance as “hearsay” mistakenly ignores the machine’s role in distributed cognition. Under a hearsay model, the live testimony of the human is deemed not only necessary, but sufficient, as a means of testing the machine’s credibility. Cross-examination of the human expert might be insufficient to unearth the design, machine-learning, input, operator, or machine degradation errors that pervert the machine report upon which the expert relies. Accordingly, cross-examination does not seem to have helped in any of the wrongful conviction cases involving “junk science.”343

The United Kingdom’s solution of requiring the testimony of any inputter of information would, in the context of expert testimony, be a significant departure from American law, but one that might make sense. Under Federal Rule of Evidence 703 and its analogs, an expert can testify to an opinion, even if based on the hearsay of others.344 A human expert, at least, can be cross-examined on her decision to rely on the assertions of others, and in a few jurisdictions, the declarants of such assertions, if they are deemed sufficiently “testimonial,” must testify as a constitutional matter.345 Most machines, on the other hand, cannot be cross-examined, and do not exercise judgment—independent of the programmer—in deciding what sorts of assertions to rely upon or not.

Looking further ahead, a jurisdiction might wish to require in-court cross-examination or out-of-court depositions of machine sources capable of answering questions posed to them, such as Watson-like expert systems. Requiring an oath and physical confrontation would presumably offer no further relevant context for the jury, unless a robot were programmed to sweat or exhibit other physical manifestations of deception on the witness stand. But allowing questioning of a machine before the jury might offer some of the same benefits as questioning human witnesses on the stand, in terms of resolving ambiguities in testimony, posing hypotheticals to an expert source, or pressing a source related to an inconsistency.

5. Jury Instructions and Corroboration Requirements

As mentioned in Part I, certain forms of risky or routinely misanalyzed human assertions, such as accomplice testimony and confessions, are subject to special jury instructions. Nonetheless, jury instructions are an underused means of encouraging jurors not to under- or overvalue evidence they are prone to misunderstand or view with prejudice. With respect to machines, both dangers are present: juries might irrationally defer to the apparent objectivity of machines,346 or reject machine sources because of an irrational mistrust of machines’ apparent complexities, even when the sources are highly credible.347

Depending on the machine source, courts might directly inform juries about black box dangers. For example, where photographs are admitted as “silent witnesses,” the court could instruct the jury about lens, angle, speed, placement, cameraperson bias, or other variables that might make the image insincere or ambiguous as a conveyor of information. Sometimes, these black box clues will not be available, or will be obvious to the jury from its own experience.348 If not, the court should use jury instruction to educate the jury about the effect of these variables on the image they are assessing.349 In short, courts should warn jurors not to “conflat[e] the realistic and the real” by treating a photograph as offering “direct access to reality”350 rather than as offering the potentially biased or ambiguous result of a black box process.

One could also imagine corroboration requirements for certain machine sources, akin to requirements for confessions and accomplice testimony.351 One way of dealing with the difficulty of validating the statistical estimates of law enforcement-elicited complex proprietary algorithms might be to require a second opinion from another machine.352 In the Hillary case, a corroboration rule would have ended in a pretrial dismissal without having to endure a trial, because the machine experts did not agree on the defendant’s inclusion as a likely contributor to the DNA mixture. Another rule might require additional corroborative evidence of guilt if machine conveyances are within a certain margin of error.353 Such rules might be grounded either in concerns about accuracy, or in concerns about dignity or public legitimacy where a machine result is the only evidence of guilt.354 In Europe, for example, the General Data Protection Regulation prohibits citizens from being “subject to a decision” that is “based solely on automated processing,” if it has a legal or “similarly significant[]” effect on the citizen.355

My goal in cataloging these potential safeguards is not to insist upon particular rules. Instead, it is to catalog the reasonable possibilities, to make clear that any future regime of machine credibility testing should draw lessons from how human testimony has been regulated, and to offer fodder for future scholarly discourse about machine credibility.

B. Machine Confrontation

The foregoing Section discussed the extent to which certain types of machine evidence implicate credibility and thus might require credibility testing—analogous to human assertions—to promote decisional accuracy. This Section briefly discusses the related but different question of whether a machine source might ever be a “witness[] against” a criminal defendant under the Sixth Amendment’s Confrontation Clause. A handful of scholars have addressed this question, and most conclude that machines themselves cannot be “witnesses”; only their human progenitors can be.356 While the subject deserves Article-length treatment, this Section briefly takes it on and suggests that machine sources sometimes may, indeed, trigger a right of confrontation.

1. Machines as “Witnesses Against” a Criminal Defendant

The Confrontation Clause of the Sixth Amendment guarantees to a criminally accused the right to be “confronted with the witnesses against him.”357 The precise meaning of the term “witnesses” has been the subject of vigorous debate in the Supreme Court for decades. The doctrine that currently exists has been in place since 2004, but has been losing some ground and is unpopular among some scholars. This Section first takes what seems to be undisputed about the Clause’s origins and purpose, and situates machines within that broad discussion. It then offers some thoughts on where machines fit within existing Supreme Court doctrine defining “witness.”

One goal of the Confrontation Clause, if not its “ultimate goal,” is to “ensure reliability of evidence.”358 A would-be accuser who is forced to take the oath, physically confront the person he is accusing, and endure cross-examination is less likely to make a false accusation. If he does make a false accusation, he is more likely to recant upon having to look the falsely accused in the eye. And the jury will have a better chance to assess the likelihood of falsehood if it can examine the declarant’s physical demeanor in court.

But accusations made behind closed doors can also subvert the dignity of criminal procedure: there is “something deep in human nature that regards face-to-face confrontation between accused and accuser” not only as promoting accuracy but as “essential to a fair trial in a criminal prosecution.”359 The Supreme Court once quoted then-President Eisenhower with approval as declaring that “[i]n this country, if someone . . . accuses you, he must come up in front. He cannot hide behind the shadow.”360 To “look me in the eye and say that”361 is to recognize me as a full person, worthy of respect. Thus, accusers should not be able to “hide behind [a] shadow”;362 rather, they should “stand behind” the accusation.363 This theme of responsibility for the truth of one’s statement squares with epistemologists’ “commitment” theory of assertion, which argues that “to assert a proposition is to make oneself responsible for its truth.”364 Such rhetoric has led scholars to acknowledge that, in addition to protecting decisional accuracy, “confrontation doctrine should protect the system’s sense and appearance of fairness.”365

One immediate target of the framers who ratified the Sixth Amendment was the centuries-old practice of using sworn affidavits of witnesses, which justices of the peace took during ex partemeetings in a “modestly formal setting, likely the [justice’s] parlor,”366 in lieu of live testimony against a defendant at trial.367 While the justices did not necessarily intend for these affidavits to replace witness testimony at trial, the Crown began to use them for that purpose. Even if the justice questioned a witness in good faith, and even if the witness did not recognize the full accusatory import of her statements, the resulting affidavit often contained mistakes, ambiguities, omissions, questionable inferences, and a slant toward a particular version of events that could not be probed or corrected at trial.368 Moreover, the defendant had no opportunity to look the witness in the eye as the witness rendered her accusation. Finally, the affidavits were sworn and had all the trappings of formality, which might have unduly swayed jurors.369 Faced with such unconfrontable but impressive-looking affidavits, defendants stood little chance of disputing them, even though the documents suffered “hearsay dangers.” The human affiants, while not bearing witness in court, clearly served as “witnesses against” the accused for purposes of implicating a right of confrontation.370

The state’s use of accusatory machine conveyances to prove a defendant’s guilt seems to implicate many of the same dignitary and accuracy concerns underlying the framers’ preoccupation with in-the-shadows accusations and ex parteaffidavits. To be sure, a machine is not, as far as we now know, capable of taking moral responsibility for a statement, or of understanding the moral gravity of accusing someone of a crime. But people are capable of doing those things, and when they build a machine to do the job, something may be lost in terms of moral commitment, if the person who is morally or epistemically responsible for the accusation is not called to vouch for the accusation in court. The court that first labeled the radar gun “push button justice” akin to “push button war” spoke only eight years after Hiroshima.371 Some view a “push button war” as threatening in part because it is easier to wage when one does not have to see the people one is killing.372 Perhaps it is easier to accuse someone when one builds an algorithm to do so.

In turn, the more inscrutable a machine process, the more its accusatory conveyances threaten the dignity of the accused and the perceived legitimacy of the process. In Kafka’s In the Penal Colony, a machine is programmed to inscribe on a condemned man’s back the law corresponding to his offense, which ultimately tortures and kills him in the process.373 Only one official is left who is willing to run the device, and Kafka emphasizes the sinister indecipherability of the machine’s blueprints.374 The polygraph, too, was mistrusted in part because of its inscrutability.375One commentator in 1955 wrote that “[t]he fear or distrust of lie detectors is in part due to the conception that the machine itself will become a ‘witness.’”376 A justice of the Oregon Supreme Court even articulated a “personhood” argument against the polygraph, reasoning that parties should be “treated as persons to be believed or disbelieved by their peers rather than as electrochemical systems to be certified as truthful or mendacious by a machine.”377 As one scholar of data science noted, “even when such models behave themselves, opacity can lead to a feeling of unfairness.”378

Allowing the state to build or harness machines to render accusations, without also providing the defendant a constitutional right to test the credibility of those machine sources, resembles trial by ex parteaffidavit. The conclusions of proprietary software created in anticipation of litigation replaces live human testimony at trial and obviates the state’s need to put a human expert on the stand to explain her methods and inputs that prompted the accusatory conclusion. And like an affidavit taken by a justice of the peace, the accusatory output—particularly output from machines created by or under contract with the state—might be incomplete or implicitly biased, even if sincere or technically accurate. As one scholar put it, “raw data is an oxymoron”379: all machine output reflects human choices about input, just as a direct examination of a witness in a justice’s parlor reflects choices about what questions to ask. Some “raw data” will be more helpful to the government’s case than others. In the Hillary case, for example, the district attorney shopped around until she found an expert system that would include the suspect as a potential contributor to the DNA mixture.380 Moreover, just as the Framers were concerned that factfinders would be unduly impressed by affidavits’ trappings of formality, “computer[s] can package data in a very enticing manner.”381 The socially constructed authority of instruments, bordering on fetishism at various points in history, should raise the same concerns raised about affidavits.

To say that machines built for criminal accusation implicate the concerns underlying the Confrontation Clause is not to say that the programmer is the one true “declarant” of the machine’s accusatory conveyance. After all, the justice of the peace was not the true declarant of an affiant’s sworn testimony: the affiant’s own testimonial infirmities were at stake. Nonetheless, the justice’s role in creating and shaping the affidavit was relevant in viewing the affiant as a “witness” in need of confrontation. The “involvement of government officers in the production of testimonial evidence” presents particular “risk[s]” of abuse.382 Perhaps these possibilities loomed large for Justice Goodwin Liu as he dissented from an opinion of the California Supreme Court stating that machines cannot be witnesses under the Clause:

[A]s a result of ever more powerful technologies, our justice system has increasingly relied on ex parte computerized determinations of critical facts in criminal proceedings—determinations once made by human beings. A crime lab’s reliance on gas chromatography may be a marked improvement over less accurate or more subjective methods of determining blood-alcohol levels. The allure of such technology is its infallibility, its precision, its incorruptibility. But I wonder if that allure should prompt us to remain alert to constitutional concerns, lest we gradually recreate through machines instead of magistrates the civil law mode of ex parte production of evidence that constituted the “principal evil at which the Confrontation Clause was directed.”383

Machine conveyances have become so probative and powerful that an algorithm like STRmix in the Hillary case can become the primary “accuser” in a criminal trial. While such software will surely help combat certain types of bias in forensic interpretation, it will create new types of bias a criminal defendant should have the right to explore.

If the Clause is concerned with unreliable, unconfronted testimony, then credibility-dependent claims that are likely unreliable and offered against the accused at trial should pose constitutional problems, particularly if the defendant does not have the opportunity to impeach the source. Several scholars have taken this view of the Clause, at least with respect to hearsay of human declarants,384 and it was the view of the Supreme Court before 2004.385 If unreliable, unconfronted testimony is the primary target of the Clause, then the accusatory output of proprietary software that has not been robustly tested would seem to be a problem potentially of constitutional magnitude.

Some scholars have suggested, along these lines, that the Clause be broadly construed, not only to guarantee courtroom testing of “witnesses,” but also to “safeguard[] the ability of a defendant to probe and to fight back against the evidence offered against him.”386 I think that view is right, with a slight modification. The Clause does use the word “witnesses,” and thus appears to address a particular kind of evidence—testimonial evidence. The Clause presumably has nothing to say about, for example, the state’s use of physical evidence, or of facts that are only relevant to the extent that another fact might be inferred from them. The Due Process Clause might govern the state’s failure to preserve or prove the integrity of physical evidence, but the Confrontation Clause presumably does not. In any event, there seems little reason to exempt unreliable machine sources from the definition of “witnesses” if reliability is the Clause’s primary target.

Even under current doctrine, many machine conveyances would seem to implicate the Confrontation Clause. In 2004, in Crawford v. Washington,the Court dramatically shifted its approach and declared that the Clause applies only to so-called “testimonial hearsay.”387 If hearsay is testimonial, the right to courtroom testing is nearly categorical; generally, only if the defendant had a prior opportunity to cross-examine a now-unavailable declarant would testimonial hearsay from that declarant be admissible.388 In turn, the question of what hearsay is “testimonial” has plagued lower courts since 2004. The Crawford Court adopted one of the definitions of “testimony” from Webster’s dictionary: “[a] solemn declaration . . . made for the purpose of establishing or proving some fact.”389 A “casual remark to an acquaintance,” however unreliable as evidence, would not be testimonial.390 On the other hand, statements in response to police interrogation are testimonial,391 unless the questioning appears primarily intended to resolve an ongoing emergency,392 because they resemble the old ex parteaffidavit practice. Presumably, volunteered accusations, where the declarant is aware of the potential prosecutorial consequences, are also squarely testimonial.393 Affidavits of forensic analysts, where the analyst certifies the reliability of the results of a laboratory process, are also generally testimonial,394 although the Court appears close to revisiting that rule.395

Under Crawford and its progeny, machines seem capable of producing testimonial evidence, given the fitting analogy to ex parteaffidavits. The primary sticking points are the Court’s perpetual focus on hearsay, which by definition refers only to the out-of-court statements of people, and its assumption that only a “solemn declaration or affirmation made for the purpose of establishing or proving some fact”396 can be testimonial.The focus on hearsay is, of course, understandable: the Framers were concerned primarily with human accusers, although bloodhound evidence presents an interesting point of comparison.397 But even some of the current Justices appear to recognize that the application of the Clause to so-called “raw data generated by a machine” is an open question with a nonobvious answer,398 much less the Clause’s application to machine experts or advanced AI witnesses. It is also true that a machine source does not make a “solemn declaration” for the “purpose” of establishing facts, if such language assumes thought, intent, and an understanding of the moral gravity of one’s accusation. Crawford took this phrase from a dictionary definition of testimony. While I sympathize with the view that Crawford’sfocus on solemnity might have been misguided and ignored broader definitions of “testimony” in the same dictionary entry,399 litigants have understandable difficulty convincing courts that machine conveyances are testimonial under this definition. Lower courts routinely hear, and reject, arguments that machine conveyances are covered by Crawford, in the context of digital infrared spectrometers and gas chromatographs reporting drug levels in blood;400 DNA typing results;401 breath test results;402 Google Earth location data and satellite images;403 red light camera timestamp data;404 and computer-generated “header” data.405 Some of these courts simply conclude that the Clause applies only to hearsay of persons, and no further analysis is required. Others correctly reason that machines are not aware of the prosecutorial consequences of their actions.

Even assuming the importance of solemnity in defining what evidence is “testimonial,” machine sources should not be given an absolute pass under the Clause. If the point of targeting solemnity is to capture what is particularly abusive about the state purposely relying on impressive but unconfronted allegations of crime as a substitute for testimony, then machine sources would seem to be squarely implicated. When a complex proprietary algorithm is wielded by the state to create testimonial substitutes for human testimony that implicate the black box dangers, in a way that allows humans to evade moral responsibility for the act of accusation, the fact that the algorithm does not itself understand how it is being used seems beside the point.

2. Rediscovering the Right of Meaningful Impeachment

While the word “witnesses” presumably limits the type of evidence covered by the Clause to evidence that is in some broad sense testimonial, there is little reason to narrowly construe “confront[ation]” as guaranteeing only the courtroom safeguards of the oath, physical confrontation, and cross-examination. Courtroom mechanisms are only one path to testing credibility, one that is entrenched in Anglo-American evidence law for a variety of historical reasons. As David Sklansky has put it, the Court’s focus on cross-examination is likely a product of its “fixation on the divide between common-law systems and civil-law systems” rather than the Clause’s true animating principles.406

The Supreme Court has stated that “confrontation” has a broader meaning, beyond its most literal sense of physical confrontation. In upholding a state practice of allowing child victims to testify outside the defendant’s presence by one-way closed circuit television, the Court in Maryland v. Craig noted that the “central concern” of the Clause is not to ensure an absolute right to physical confrontation, but “to ensure the reliability of the evidence . . . by subjecting it to rigorous testing.”407 “The word ‘confront,’ after all, also means a clashing of forces or ideas, thus carrying with it the notion of adversariness.”408 While the drafters of the Sixth Amendment clearly contemplated courtroom safeguards as the “elements of confrontation,” the Court made clear that face-to-face confrontation “is not the sine qua non of the confrontation right.”409 Instead, it is the right of the defense “to probe and expose [testimonial] infirmities.”410

Moreover, the Supreme Court seems to have implicitly recognized that the common-law right of confrontation contemplated a general right of meaningful impeachment, rightly focused on general credibility testing rather than on particular courtroom mechanisms. In Jencks v. United States411 and Gordon v. United States,412the Court required the prosecution to disclose witnesses’ prior statements—with no showing of materiality or favorability to the defense—so the defense itself could determine their “impeaching weight and significance,”413 and to avoid burying “important facts bearing on the trustworthiness of crucial testimony.”414 While Jencks and Gordon do not invoke the Sixth Amendment or a constitutional right of confrontation, at least one Justice later commented on the cases’ “constitutional overtones,”415 grounded in the “common-law rights of confrontation.”416 The cases stood for the “basic Jencks principle of assuring the defendant a fair opportunity to make his defense.”417 Such a right of impeachment would seem to contemplate credibility testing in general, not simply courtroom safeguards.

But with the passage of the Jencks Act quickly on the heels of these decisions in 1957, the underlying reasoning of cases like Jencks was lost. The Jencks Act by its terms applies only to witnesses who testify in court. But the purpose of that restriction, like the Act’s pronouncement that only “substantially verbatim” statements of the witness418 need be disclosed, was to ensure witness safety before trial, to avoid fishing expeditions, and to protect work product of government investigators.419 Even giving full force to these concerns, there would seem little reason not to extend the principles of Jencks to machine sources.

A right to meaningful impeachment of a nonhuman source might require much more, or less, than courtroom testing. Case-specific cross-examination of the programmer responsible for designing a software package may be unnecessary to probe the machine’s potential for falsehood by design, inarticulateness, or analytical error due to design malfeasance or mistake. Instead, the programmer could give live testimony before some type of scientific commission, and return to the commission every time the software is changed or updated. Such a commission might seem anathema to existing adversarial structures, but a similar proposal for “advisory tribunals” to assess conflicting expert testimony was made by Learned Hand over a century ago,420 and several bipartisan commissions have weighed in on how human forensic expert testimony should be presented.421

On the other hand, meaningful impeachment of a machine in a given case might require access to source code422 or, alternatively, written answers to interrogatories that are completed by humans but that question the machine as if it were on cross-examination, such as “what population frequency statistics are you using in calculation of your likelihood ratio?” or “what threshold do you use in deciding what to call a genetic marker versus “noise”?” Meaningful impeachment might also include, where feasible, the presence of a defense expert at the time of testing to discourage and unearth case-specific input errors.423 And it might require, as in Jencks itself, disclosure of prior statements of machines even when the prosecutor might not consider them “exculpatory” and “material,” thus removing them from the scope of disclosure as a matter of due process under Brady v. Maryland.424

Some might argue that the admission of machine evidence, a fast-changing field to be sure, should not turn on slow-moving constitutional litigation based on shaky doctrine. Hard-and-fast rules requiring, for example, the live testimony of a programmer for certain types of software might prove both overly burdensome on the state and unnecessary to meaningful impeachment. Perhaps, as a matter of strategy, reformers should focus their efforts on a workable, nonconstitutional impeachment standard for machine sources. But to immunize accusatory machine output from the Clause’s reach entirely seems to be the wrong answer, at least as a theoretical, if not strategic, matter. Daubert and Frye are not constitutional requirements, and a state tomorrow could choose to admit relevant and authenticated machine conveyances with no credibility testing whatsoever.

In other contexts, the Sixth Amendment has a standard-based application that seems to work well without hard and fast rules that unduly curtail judicial discretion or burden parties. For example, the denial of certain lines of cross-examination is generally a matter within the sound discretion of the trial judge, but can rise to the level of a Sixth Amendment violation. Thus, a defendant who is prohibited “from engaging in otherwise appropriate cross-examination designed to show a prototypical form of bias on the part of the witness,” critical to the jury’s credibility determination, is denied his constitutional right of confrontation.425 A similar standard might find a constitutional violation where the defendant is curtailed from testing a key aspect of the credibility of a critical machine source.


This Article has argued that certain machine evidence implicates the credibility of a machine source, that the black box dangers potentially plaguing machine sources trigger the need for credibility testing beyond what is contemplated by existing law, and that accusatory machine conveyances can be “witnesses against” a defendant under the Confrontation Clause. It has also offered a glimpse of the sorts of evidentiary and constitutional rules that might eventually govern machine sources of information. While we may never fully resolve the agency paradox underlying modern science, one does not have to believe that machines are entities capable of independent “thought” to understand the need to test their credibility or cabin the state’s ability to hide behind their algorithmic accusations without robust credibility testing.

Exploring “machine testimony” reminds us that the law of human testimony has relied too heavily on a courtroom model of credibility testing and confrontation. Sometimes, the right to meaningfully impeach humansrequires more than simply cross-examination. The Jencks Act, for example, does not apply to human hearsay accusers, even though access to the prior statements of hearsay declarants to impeach them through inconsistency, even if not on cross-examination, might be critical to the defense.426 Federal Rule of Evidence 703 should perhaps require more scrutiny of assertions relied upon by human experts.427 Front-end protocols, like the ones governing eyewitness identifications in some states, should be considered for other types of human testimony as well, such as on-scene witness statements to police officers. And jury instructions and corroboration rules should perhaps be considered for other types of human testimony.428 Perhaps the sacred dichotomy between testimonial and physical evidence should itself be revisited; indeed, the Innocence Project has suggested treating eyewitness testimony as akin to trace evidence, the “result” of a process, just like courts have attempted to do with machine reports.429 Meaningful impeachment of an eyewitness might move beyond cross-examination and toward access to experts. While human brains are not equivalent to a computer’s black box,430 cognitive psychologists have much to share that could avoid leaving juries with misimpressions about the probative value of human testimony.

The message of this Article is hopeful. While the Anglo-American system of proof is imperfect, to say the least, its strength is in its flexibility, which “creates space for experimentation with new approaches and also reduces the pressure for radical surgery on the existing system.”431 Creating new rules for machine sources, and adapting existing rules to accommodate machine sources, will not radically change our system of proof. Instead, recognizing machine conveyances as credibility-dependent will bring this critical area of conceptual and doctrinal confusion into line with the values underlying existing testimonial safeguards for human witnesses. If we do that, there is every reason to believe evidence law can “weather the coming tempests in proof technology.”432