Part Three: AI on Trial – Admissibility of AI-Generated Evidence

Artificial intelligence (AI) is no longer a tool reserved for research laboratories and technology companies. It has entered courtrooms, not only as an aid to docket management or legal research but as a source of evidence. From documents generated by large language models (LLMs)^[1] to deepfake videos^[2] and AI-assisted data analysis^[3], courts now confront a pressing question: can such evidence be admitted? The existing rules of evidence were not designed with these technologies in mind, yet they remain the framework judges must apply to ensure that trials are fair, reliable, and consistent with due process. So, the stakes for the judiciary could not be higher.

The Rise of AI-Generated Evidence

AI can produce a wide array of materials that parties may seek to introduce as evidence: synthesized images or videos^[4], machine-translated transcripts^[5], algorithmic predictions^[6], and even summaries of complex data sets^[7]. What sets AI evidence apart from traditional sources is the increased difficulty in verifying authenticity and reliability. Traditional evidence usually has a clear chain of custody and/or human witness to explain its origins. AI-generated material often does not. This raises critical challenges for judges, who serve as gatekeepers under the Rules of Evidence.

When conducting evidentiary analysis under the Frye standard^[8] the focus is on whether the underlying scientific method or principle is generally accepted within the relevant expert community.  Under the Daubert/Rule 702 framework^[9], the analysis is broader: judges must determine whether the evidence is not only relevant but also reliable, with the proponent bearing the burden of proof by a preponderance of the evidence.  Under the Daubert framework and the amended Rule 702, judges must ensure that the proponent demonstrates the reliability of AI evidence by a preponderance of the evidence. At the heart of both approaches, however, the questions remains those of relevance, reliability, and authenticity. AI-generated evidence may satisfy the test of relevance, but the processes that create it can be opaque, biased, or vulnerable to manipulation. Accordingly, when applying evidentiary standards, judges must require that proponents establish both the reliability and authenticity of such evidence.

Deepfakes and the Threat to Authenticity

Perhaps the most troubling form of AI evidence is the deepfake — an image, video, or audio file that appears real but is entirely fabricated. Deepfakes can be created with alarming ease, raising the possibility that fabricated evidence could be introduced to sway juries or intimidate witnesses.

Authentication under Rule 901^[10] requires the proponent to demonstrate that the evidence “is what it purports to be.”^[11]With deepfakes, this burden becomes far more complex. Judges may require expert testimony to explain the techniques used to create or debunk such material. Yet this adds significant costs to litigation and risks overwhelming courts with technical disputes. More troubling still, even genuine evidence may be challenged as fake, fueling what some scholars call the “liar’s dividend”: the erosion of trust in all forms of evidence.

Reliability and the “Black Box” Problem

Reliability is a core concern in admitting AI-generated evidence. Courts often rely on expert witnesses under Rule 702^[12] to explain scientific or technical evidence, but AI complicates this analysis. Many systems, particularly those built on deep learning, operate as “black boxes,”^[13] with internal reasoning opaque even to developers. How can a judge evaluate the reliability of a process that cannot be explained?

This issue came to the fore in Wisconsin v. Loomis,^[14] where the defendant challenged the use of a proprietary risk-assessment algorithm at sentencing. The court upheld its use, reasoning that it was merely one factor among many. The United States Supreme Court denied certiorari despite the fact the defendant was unable to examine the algorithm’s methodology because the vendor claimed it was a trade secret.^[15] Cases like Loomis highlight the tension between proprietary protections and a defendant’s constitutional right to confront and challenge the evidence against them. If judges admit AI-generated evidence without adequate scrutiny, they risk undermining both fairness and constitutional due process.

The Role of Expert Testimony

Given the technical complexity of AI, courts will increasingly rely on expert witnesses to establish authenticity and reliability. These experts must be able to explain how the AI system was trained, what data it used, and how outputs were validated. But not every litigant can afford such experts, creating an equity problem. Wealthier parties may be able to marshal expert testimony, while under-resourced litigants may be left vulnerable, creating the potential for a two-tiered evidentiary system. Judges must therefore be mindful of the access-to-justice implications when AI evidence is at issue.

The ethical responsibility here is clear. Judges must exercise heightened vigilance, ensuring that experts are qualified and that their testimony rests on sound scientific principles. Lawyers, in turn, must not blindly accept AI outputs but must verify their accuracy before presenting them in court. The duty of competence now extends to understanding the limits of AI evidence.

Privacy, Data, and Chain of Custody

AI evidence also raises questions about privacy and chain of custody. AI systems often rely on massive datasets, some of which may include sensitive or unlawfully obtained information. If AI evidence is generated from improperly sourced data, courts must consider whether admitting it would violate privacy laws or ethical obligations.

Chain of custody presents another hurdle. Traditional evidence must be carefully tracked to ensure it has not been altered. With AI, ensuring the integrity of digital inputs and outputs requires new protocols. Courts may demand documentation showing how data was collected, processed, and analyzed. Without such safeguards, the risk of tampering or error looms large.

Public Trust and the Legitimacy of AI Evidence

Ultimately, the admissibility of AI evidence touches on more than technical rules; it implicates public trust in the justice system. If litigants and the public believe that outcomes rest on unreliable or incomprehensible machine processes, confidence in the courts will erode. Conversely, if courts are seen as proactive in scrutinizing AI, setting clear standards, and safeguarding fairness, legitimacy can be preserved.

Transparency and communication are essential. Courts should be open about when and how AI evidence is admitted, what standards were applied, and why certain safeguards are necessary. This not only educates the public but also reinforces the judiciary’s role as a guardian of fairness.

Keeping Judges in Control

The guiding principle must be that AI serves justice, not the other way around. Judges cannot cede their constitutional role as arbiters of fact and law to machines. AI may assist by providing information, but it cannot weigh credibility, apply moral reasoning, or consider the human consequences of a decision.

The ethical duty is clear: judges must remain firmly in control, using AI evidence cautiously and only when its reliability can be demonstrated. Lawyers must likewise ensure that their use of AI complies with professional responsibilities of competence, candor, and fairness.

Conclusion: The Future of AI Evidence

While the dangers of AI evidence are significant, it is important to recognize the potential benefits as well. Properly validated AI tools can assist courts in managing complex data, identifying patterns more quickly than humans, and even reducing costs in cases with massive discovery burdens. When admitted responsibly, AI-generated evidence has the potential to enhance efficiency and accuracy, provided judges apply rigorous safeguards to ensure that fairness and due process remain paramount.

The courtroom of the future will almost certainly see more AI-generated evidence, ranging from predictive analytics to multimedia exhibits. The question is not whether AI will appear in trials—it already has—but how courts will respond. Admissibility must remain anchored in the traditional principles that have long guided evidence: relevance, reliability, authenticity, and fairness.

Other Articles in this series

Introduction: Artificial Intelligence and the Courts: A Blog Series from Justice Speakers Institute
Part 1: AI in the Courtroom: Opportunities and Risks
Part 2: AI in the Courts: Ethical Challenges
Part 3: AI on Trial – Admissibility of AI-Generated Evidence
Part 4: Judicial Decision-Making: Transparency, Accountability, and the Judicial Role
Part 5: Courts of the Future-Innovation, Access, and Global Trends
Part 6: Judging the Machine-Lessons, Guardrails, and the Path Forward

---

Citations

^[1] A large language model is a type of artificial intelligence system that has been trained to understand, generate, and work with human language.

^[2] A deepfake video is a type of synthetic media created using artificial intelligence (AI) that makes it appear as though someone is saying or doing something they never actually did.

^[3] AI-assisted data analysis is the use of artificial intelligence tools and techniques to help people examine, interpret, and make sense of data more effectively and efficiently

^[4] Synthesized images or videos are artificially created or altered media, often produced by artificial intelligence, that depict events, people, or scenes which may never have actually occurred.

^[5] Machine-translated transcripts are written records of spoken or written language that have been automatically converted from one language to another by computer software, without human translation or editing.

^[6] Algorithmic predictions are forecasts or decisions generated by computer algorithms that analyze data patterns to estimate future events, behaviors, or outcomes such as those used in risk and needs assessments.

^[7] Summaries of complex data sets are simplified explanations or condensed reports that highlight key patterns, trends, or findings from large or complicated collections of information, making them easier to understand and use.

^[8] Frye v. United States, 293 F. 1013 (D.C. Cir.1923). Though now considered the minority rule in the United States, it remains the law in several major states, including California, Texas, and New York.

^[9] Daubert/Rule 702, which is based upon the Supreme Court decision in Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993), is now the majority approach within the United States. Under this approach expert evidence is admissible only if it is grounded in reliable principles and methods, applied properly to the facts, and shown to assist the judge or jury in understanding the evidence or determining a fact in issue.

^[10]  Rule 901. Authenticating or Identifying Evidence

^[11] Id.

^[12]  Rule 702. Testimony by Expert Witnesses

^[13] A black box is a system, device, or process whose internal workings are hidden or unknown, but whose inputs and outputs can be observed and analyzed.

^[14] Loomis v. Wisconsin, 881 N.W.2d 749 (Wis. 2016)

^[15] https://www.scotusblog.com/cases/case-files/loomis-v-wisconsin/ Loomis v. Wisconsin

INTERESTED IN AI AND THE COURTS?

Artificial Intelligence is transforming justice. But it also raises complex questions of ethics, fairness and accountability. At the Justice Speakers Institute (JSI), we provide training, consulting, and expert presentations to help courts, policymakers, and legal professionals navigate these challenges responsibly.

Contact us today to learn how JSI can support your organization in understanding and implementing AI in the courts.