The increasing use of AI-generated evidence in jury trials presents complex challenges for the US court system and requires a delicate balance between leveraging helpful technology and guarding against potential deception
Key highlights:
-
-
-
AI evidence creates a credibility dilemma for juries — Jurors are prone to treating AI outputs as factual and authentic, which makes it difficult to distinguish between legitimate evidence and sophisticated deepfakes.
-
Current evidentiary rules are inadequate for AI — Traditional rules of evidence may not be compatible with AI’s ability to create hyper-realistic fabricated content and authenticate evidence.
-
Courts need proactive measures and adaptability — To navigate this new frontier, courts must implement comprehensive jury instructions, build boards of AI evidence experts, provide red flag training for all participants, and develop flexible legal guidelines that can be adapted to the rapid evolution of AI technology
-
-
AI evidentiary issues are presenting multi-layered challenges in the United States court system. To deal with them, courts must perform a careful balancing act: Embracing helpful technology while guarding against potential deception from AI-generated evidence, especially in jury trials.
Several issues make this balancing act a challenge. First, individuals — including those in juries — tend to treat AI outputs as authentic and factual. “We know that people often treat artificial intelligence, [and] the outputs that they receive from there, as factual, and that inflates credibility across the board,” says Jawwaad Johnson, Director of the Center for Jury Studies and Principal Court Management Consultant at the National Center for State Courts (NCSC). “We know that the artificial intelligence itself has become more believable.”
Second, audiovisual testimony can be more memorable in a juror’s mind than written testimony, according to Dr. Maura R. Grossman, a Research Professor at the University of Waterloo (Ontario) and an eDiscovery lawyer and specialist. And that can irreversibly influence juries, particularly if the audiovisual testimony is fabricated, because this reality makes deepfakes hard to unsee, Grossman explains.
The Honorable Erica R. Yew, of the Santa Clara County Superior Court in California, agrees that the novelty of this development can be a challenge. “In the past, video has been used to refresh someone’s recollection when they forgot something,” Judge Yew says. “And so now we are worried about videos being used to modify or change or corrupt someone’s memory.”
Finally, the liar’s dividend risks juries dismissing legitimate, properly authenticated evidence simply because AI manipulation is possible. “If we overdo it, we are going to make our jurors so skeptical of everything, and they will become cynical and question all evidence, even legitimate evidence,” Grossman says. “But if we give them no guidance, we certainly do not want them pulling out magnifying glasses” to ascertain authenticity on their own using ad hoc methods.
A recent webinar hosted by the National Center for State Courts and the Thomson Reuters Institute AI Policy Consortium, looked at how courts can navigate the complexity of these psychological and technical challenges. The webinar panel included Judge Yew, Grossman, Johnson, and Megan Carpenter, Dean and Professor of Law at the University of New Hampshire Franklin Pierce School of Law, as the moderator.
The panel discussed how AI can impact evidence that is both acknowledged and unacknowledged. For example, acknowledged AI-generated evidence can enhance expert testimony and improve juror comprehension, such as in accident reconstruction. This is only possible when AI methods are transparent and there is clear chain-of-custody of the evidence.
Unacknowledged AI-generated evidence, on the other hand, can include deepfakes and other falsified evidence that are intended to deceive and may be hard to detect. Courts and lawyers must balance skepticism, disclosure, and expert input in these cases to protect juries without paralyzing them.
Limitations of current rules of evidence
Panelists also laid out options for practical legal frameworks that govern how acknowledged and unacknowledged AI-generated evidence can be admitted and evaluated in courtrooms. Different rules apply, of course, depending on the type of evidence. More specifically, in matters of acknowledged AI-generated evidence, validity, reliability and bias are primary considerations; and in matters concerning potential unacknowledged AI-generated evidence, the primary issue for judges and juries is to determine authenticity.
The current rules regarding evidence may not always be compatible with AI-generated content in several critical ways. Traditional authentication rules under Federal Rule of Evidence 901 assumes that evidence originates from reliable sources; however, given AI’s ability to create hyper-realistic deepfakes that are indistinguishable from authentic content, this assumption may not always be correct. In addition, current self-authenticating document provisions in Rule 902 may inadvertently admit fabricated evidence that has been processed through official channels, such as AI-generated documents filed with government agencies that then become official records.
Further, the technical sophistication of generative AI compounds these challenges. AI systems use a training method in which two algorithms compete to get better, which makes it harder to detect them, Grossman explains, adding that current automated-detection tools often fail in these cases, and even human experts can only provide probability assessments rather than definitive determinations of authenticity.
In her work with Judge Paul W. Grimm, Grossman has proposed two key reforms. The first, related to acknowledged AI-generated evidence, seeks to incorporate aspects of the Daubert standard into current evidentiary rules. Instead, the Federal Rules Advisory Committee decided to draft a new rule — proposed Federal Rule of Evidence 707 — which applies the expert reliability standards found in current rules to machine-generated evidence offered without expert testimony. This solution ensures AI-generated evidence meets the same reliability requirements as expert testimony while also addressing concerns about validity, bias, and methodological soundness.
In scenarios involving unacknowledged AI-generated evidence, such as in cases in which the authenticity of the evidence is disputed, Grossman and Judge Grimm proposed that if there is evidence that a jury could reasonably believe that suggests that the evidence is and is not authentic, judges should use a balancing test, which weighs how much the evidence actually helps prove something important in the case (known as probative value), against the risk that it will unfairly inflame, mislead, confuse, or waste time (referred to as the prejudicial value).
What can courts do now
Courts must take critical steps to deal with AI evidence in jury trials as it becomes increasingly sophisticated. The NCSC’s Johnson cites the importance of ensuring court participants are prepared. “There’s a very technical aspect to this discussion,” Johnson says. “And there is certainly a space for education… not just for individuals serving on a jury, but for the people who help juries.”
Most importantly, implementing comprehensive jury instructions that help jurors understand their evolving responsibilities in evaluating digital evidence allows them to gather valuable insight. And having judges allow jurors to ask questions about how AI is used in evidence during the process of screening questionable evidence can improve jury comprehension.
In addition, two additional actions for courts include:
Building a board of AI evidence experts — Courts may need to appoint AI-detection specialists to ensure fair evaluation when parties lack resources for private experts. Judge Yew points out that there is precedent to having a group of people whom the courts can call upon as experts on retainer, such as those used in competency hearings to determine an individual’s competence capacity to withstand a criminal case.
Offering “red flag” training — Courts should also enable attorneys, judges, and jurors to spot suspicious evidence through training. Indeed, court participants should learn how to scrutinize too good to be true evidence, particularly when original devices or documents are unavailable for examination, Grossman advise. Elaborate explanations for unavailability should trigger heightened scrutiny and potential expert analysis to verify authenticity before the evidence reaches a jury.
Finally, the panelists say they believe that, over the long term, any legal framework requires flexible rules that can adapt to rapidly evolving technology. Given the swift pace of AI development, rigid regulations would quickly become obsolete. Instead, adaptable guidelines that focus on principles like reliability, transparency, and fairness will better serve the future legal proceedings in which AI evidence is involved.
To learn more about AI evidentiary issues, visit the AI Policy Consortium hosted by NCSC and the TR Institute and the specific resources for judges
