The Courtroom's New Witness Can't Be Sworn In
Picture this: a fraud trial in a federal courtroom. The defense attorney steps forward and presents not a witness, not a document, but an AI-generated forensic accounting report—one trained on millions of financial records—that claims to dismantle the prosecution's entire theory. No human expert created it. No one can explain, line by line, how it reached its conclusions. And yet, there it sits, waiting to be admitted as evidence.
This isn't science fiction. This is the reality of American courtrooms in 2026. And it's forcing judges, lawyers, and legislators into the most consequential rethinking of evidence law since DNA testing walked through the courthouse doors decades ago.
I've spent months tracking how AI-generated evidence is reshaping litigation strategy—not in some hypothetical future, but right now. What I've found is both fascinating and deeply unsettling. The algorithms are already inside the courtroom. The question isn't whether they belong there. It's whether our legal system can evolve fast enough to prevent them from warping the very concept of truth.
The Evidence That Thinks for Itself
To understand what's happening, you need to appreciate the sheer breadth of what "AI evidence" now means. It's not a single technology. It's a sprawling category that touches virtually every type of case you can imagine.
In securities litigation, parties are offering predictive algorithms that analyze a decade of stock trading patterns to demonstrate the magnitude of market impact. In copyright disputes, AI systems compare visual data between two works to calculate "substantial similarity" with a precision no human eye could match. In criminal cases, prosecutors are using AI to enhance grainy surveillance footage—sharpening resolution, removing blur, and effectively creating visual information that didn't exist in the original recording.
And then there are the risk assessment tools. Systems like COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) evaluate a defendant's criminal history, demographic factors, and behavioral data to generate a score predicting the likelihood of reoffending. These scores now routinely influence bail decisions, sentencing recommendations, and parole outcomes across the United States.
Each of these applications raises a different set of questions. But they all share a common thread: the evidence is being generated, enhanced, or interpreted by a machine whose reasoning process is, at best, partially transparent. That's a problem for a legal system built on the principle that evidence must be tested, challenged, and understood before it can determine someone's fate.
The Deepfake That Killed a Case
If you want to see how fast things are moving—and how unprepared courts are—look at what happened in Alameda County, California, in September 2025.
In Mendones v. Cushman & Wakefield, Inc., self-represented plaintiffs submitted video exhibits that were supposed to capture witness testimony in support of their motion for summary judgment. Something was off. The judge, Victoria Kolakowski, noticed that the people in the videos appeared "robotic" and lacked natural expressiveness. The accents, cadence, gestures, and facial expressions of the individuals in the disputed videos were dramatically different from those in other verified recordings of the same person.
After closer examination, Judge Kolakowski made a finding that would have been unimaginable even five years ago: the exhibits were deepfakes—AI-generated fabrications of human testimony. The court dismissed the case with prejudice and imposed terminating sanctions, writing that such a sanction "serves the appropriately chilling message to litigants appearing before this Court: Use GenAI in court with great caution."
The Mendones case is widely considered one of the first reported instances of deliberately fabricated AI evidence being submitted in an American courtroom. But here's what should truly worry you: Judge Kolakowski also admitted that the court lacked "the time, funding, or technical expertise" to determine the authenticity of all the exhibits offered by the plaintiffs. She caught the problem because the deepfakes were, frankly, crude. What happens when they're not?
The "Deepfake Defense" and the Erosion of Trust
Here's the twist that most people miss when they think about AI and courtrooms. The more dangerous threat isn't just that someone will submit fake evidence. It's that the mere possibility of fake evidence will poison the well for everything that's real.
Legal scholars call this the "deepfake defense," and it's already being deployed. In Huang v. Tesla, discovery produced a video of Elon Musk making statements about the safety of Tesla's Autopilot feature. When the plaintiff submitted a Request for Admission asking Tesla to confirm the video's authenticity, Tesla refused—arguing that because Musk is a public figure frequently targeted by deepfakes, the company could "neither admit nor deny" whether the video was genuine.
The court rejected this argument, recognizing the dangerous precedent it would set. If any public figure could dismiss their own recorded statements by invoking the specter of AI manipulation, accountability would evaporate overnight. But the tactic revealed something important: the deepfake defense doesn't need to succeed to do damage. It just needs to plant doubt.
And doubt is spreading. In the Kyle Rittenhouse trial, the defense objected to the prosecution zooming in on an iPad video already admitted into evidence, arguing that Apple's pinch-to-zoom function uses AI to manipulate video. The prosecution couldn't produce an expert on the spot to refute this claim, and the court prohibited the zoom. In United States v. Reffitt, a January 6th case, the defense cross-examined an FBI agent about whether video evidence of the defendant at the Capitol might have been AI-manipulated—without providing any supporting basis for the claim. The court allowed the questioning.
Do you see the pattern? We're entering a world where genuine evidence can be undermined simply by uttering the word "deepfake." Jurors who are already 650% more likely to retain information from video testimony are now being told, with increasing frequency, that they can't trust what they're seeing. The irony is suffocating: AI is simultaneously making it easier to fabricate evidence and easier to discredit evidence that's perfectly real.
Proposed Rule 707: The Federal Government's Answer
Washington, to its credit, isn't ignoring the problem. In November 2024, amendments to the Federal Rules of Evidence were proposed to address AI-generated evidence. The most significant of these—Proposed Rule 707—has been making its way through the rulemaking process ever since. As of early 2026, it's in the public comment period, with a committee vote scheduled for May 2026.
Here's what Rule 707 would do: when machine-generated evidence is offered without a human expert witness, and that evidence would be subject to Rule 702 (the Daubert standard for expert testimony) if a human had generated it, the court may only admit the evidence if it satisfies the same reliability requirements. Specifically, the proponent must demonstrate that the AI output is based on sufficient facts or data, produced through reliable principles and methods, and reflects a reliable application of those methods to the facts of the case.
In plain English? You can't use an algorithm to sneak expert-level conclusions past a judge just because no human is technically "testifying." If an AI report would need to meet the Daubert standard coming from a person, it needs to meet the Daubert standard coming from a machine.
The proposal has its supporters and its critics, and both sides make compelling points. Proponents argue that Rule 707 closes a genuine loophole. Under the current patchwork of Rules 401, 402, 403, and 901, there's no clear mechanism for an opposing party to challenge the reliability of an AI system's underlying methodology. A party could submit an AI-generated damages calculation without any of the scrutiny that would attend a human expert's report. Rule 707 would change that.
Critics, however, raise practical concerns that deserve serious attention. Complying with Rule 707 would require expensive technical experts to analyze complex AI software—potentially creating a two-tiered system where well-funded litigants can use AI evidence and under-resourced ones can't. There's also a risk that the rule becomes a litigation weapon, with parties filing pretrial challenges to AI evidence not because they genuinely doubt its reliability, but to delay proceedings and drain opponents' resources. Some observers have already coined the term "Rule 707 hearing" by analogy to Markman hearings in patent law—suggesting these proceedings could become a routine, expensive fixture of complex litigation.
Perhaps the most pointed criticism is this: Rule 707 only applies to evidence that the proponent acknowledges was created by AI. It does nothing to help courts identify or exclude deepfakes and other fabricated evidence whose AI origins are concealed. That's a significant gap.
The States Aren't Waiting
While the federal rulemaking process grinds forward, state legislatures have been moving faster—sometimes much faster.
On August 1, 2025, Louisiana became the first state to establish a comprehensive framework for addressing AI-generated evidence. Under Act 250, attorneys in Louisiana are now required to investigate and verify the authenticity of digital evidence before offering it in court. It's a shift in professional responsibility that makes lawyers, not judges, the first line of defense against AI fabrication.
California has multiple legislative efforts underway. New York introduced A.B. 1338, addressing AI-generated evidence in litigation. And the movement isn't confined to traditionally tech-forward states. Illinois, through its Supreme Court and Judicial Conference Task Force, has released policies on AI use in the courts. These state-level efforts represent a patchwork, yes, but they also represent something more important: urgency. The states understand that waiting for federal rules to be finalized isn't an option when deepfake evidence is being submitted today.
The variation between states, though, creates its own problems. A piece of AI-enhanced video that's admissible in one jurisdiction might be excluded in another. For national litigation—think class actions, multidistrict litigation, cases involving companies operating across state lines—this inconsistency adds a new layer of strategic complexity. Lawyers will increasingly need to consider not just what evidence to present, but where the legal framework is most favorable for the kind of AI evidence they intend to use.
Algorithms in the Jury Box: Risk Assessment and Sentencing
I've been talking about AI evidence that parties submit. But there's another category that arguably matters more, because it operates with far less scrutiny: algorithmic tools used by the justice system itself.
Risk assessment algorithms—COMPAS being the most notorious—now shape outcomes for millions of defendants. These tools analyze variables including criminal history, employment status, and social connections to generate a recidivism score. Judges use these scores to inform decisions about bail, sentencing, and parole. And they do so in a system where judges at pretrial hearings sometimes spend an average of thirty seconds considering each case.
The constitutional problems are staggering. The Fourteenth Amendment guarantees equal protection and due process. But risk assessment tools trained on historical criminal justice data inevitably inherit the biases embedded in that data. Black men constitute approximately 13% of the male population in the United States but account for roughly 35% of the incarcerated population. When an algorithm learns from arrest records shaped by decades of racially disparate policing, it doesn't correct for that history—it amplifies it.
In State v. Loomis, a defendant challenged the use of a COMPAS risk assessment in his sentencing, arguing the tool was opaque and potentially biased. The court acknowledged the concerns but ultimately upheld the assessment's use, urging caution. That caution, however, hasn't translated into systematic change. Predictive tools remain in use across most U.S. states, and the fundamental tension between algorithmic efficiency and constitutional rights remains unresolved.
The feedback loop problem makes this worse. If an algorithm identifies a neighborhood as "high risk" based on historical arrest data, police deploy more resources there, leading to more arrests, which feeds back into the algorithm as confirmation of high risk. The data doesn't reflect where crime occurs; it reflects where police look for crime. And yet these outputs are treated with an authority that human judgment rarely receives.
The AI Enhancement Arms Race
Beyond deepfakes and risk assessments, there's a quieter revolution happening in how authentic evidence is processed before it reaches a courtroom. AI enhancement—using machine learning to sharpen blurry surveillance footage, amplify faint audio, or reconstruct partial images—is becoming standard practice. And it's creating a deeply unequal playing field.
In a Washington state criminal trial, the prosecution attempted to introduce surveillance video that had been enhanced through AI to compensate for low resolution and motion blur. The court denied admission, ruling that the enhancement posed dangers of unfair prejudice. The reasoning was sound: AI enhancement doesn't just clarify existing information. It generates new visual data—pixels that weren't in the original recording—based on the algorithm's predictions about what "should" be there. That's a fundamentally different thing from adjusting brightness or contrast.
But not every court has reached the same conclusion. Some judges have admitted AI-enhanced evidence; others have rejected it. There are no uniform standards for what constitutes acceptable enhancement versus impermissible alteration. And here's where the inequality bites hardest: AI enhancement tools are expensive. Sophisticated forensic AI services can cost tens of thousands of dollars per case. A well-funded prosecution or a corporate defendant can afford to enhance every piece of video and audio evidence. A public defender's office or a small-firm plaintiff's lawyer often cannot.
This creates an asymmetry in the quality and persuasiveness of evidence that has nothing to do with the underlying facts of a case—and everything to do with resources. If AI enhancement becomes routine, we may be building a system where the clarity of your evidence depends on the size of your legal budget.
Facial Recognition: The Silent Witness That Gets It Wrong
No discussion of AI in the courtroom is complete without confronting facial recognition technology—perhaps the most operationally deployed and constitutionally fraught AI tool in the criminal justice system.
In January 2025, reporting revealed that fifteen police departments across twelve states were using facial recognition systems to make arrests without direct evidence linking suspects to crimes beyond the algorithm's identification. The technology, which compares surveillance camera images against databases of photos, has been used to build cases and, in some instances, serve as the primary basis for charging decisions.
The accuracy problem is well-documented and severe. In Detroit, a facial recognition program misidentified suspects approximately 96% of the time and led to the wrongful arrests of several Black residents. Studies have consistently shown that facial recognition systems exhibit higher error rates for people of color, women, and older individuals. The technology's performance degrades further with poor lighting, unusual angles, or low-resolution cameras—precisely the conditions under which most law enforcement surveillance footage is captured.
When facial recognition output enters the courtroom as evidence, it carries the aura of scientific precision. Jurors hear that a computer matched a face to a database and intuitively assign that identification a weight that may far exceed its actual reliability. The human tendency to trust machines—especially ones that seem to operate beyond human cognitive limitations—creates an evidentiary distortion that defense attorneys are only beginning to learn how to counter.
Rewriting Trial Strategy: What Smart Lawyers Are Doing Now
The lawyers who are adapting fastest to this landscape aren't the ones with the biggest AI budgets. They're the ones who understand that AI evidence changes the strategic calculus of litigation at every stage—from discovery through closing arguments.
During discovery, expect new battles over AI provenance. If your opponent used AI to generate a damages model or enhance a piece of evidence, you're going to want to know everything: what system was used, what data it was trained on, what parameters were set, and what the error rate is. Under proposed Rule 707, this information would be necessary to challenge admissibility. But even without Rule 707, savvy attorneys are already building these requests into their discovery frameworks.
In pretrial motions, anticipate a surge in challenges to AI evidence under Daubert and Frye standards. The Weber court in 2024 established that AI-generated expert evidence should be subject to a Frye hearing before admission, with the proponent bearing an affirmative duty to disclose AI use. Other jurisdictions are following suit. If you're planning to introduce AI evidence, you need to prepare for a reliability hearing—and that means retaining experts who can explain your AI system's methodology in terms a judge will understand.
At trial, the deepfake defense is becoming a standard play. If you're offering video or audio evidence, prepare for the opposing side to question its authenticity. That means building a chain of custody for digital evidence, obtaining metadata, and potentially retaining forensic analysts who can testify to authenticity. Conversely, if you're defending against AI-enhanced evidence, you now have a powerful line of attack: challenge not just what the evidence shows, but what the AI added to what it shows.
For jury presentations, the new Rule 107 (governing illustrative aids) opens the door for AI-generated trial graphics, simulations, and visual recreations. These aren't offered as evidence—they're explanatory tools. But they're subject to a balancing test akin to Rule 403, meaning courts must assess whether an aid's usefulness is substantially outweighed by the danger of unfair prejudice. The attorneys who master this space will have a significant advantage in persuasion.
The Global Dimension
This isn't just an American problem. Courts worldwide are grappling with AI evidence, and the approaches are diverging in ways that matter for cross-border litigation, international arbitration, and human rights accountability.
The European Union, with its AI Act, has taken a risk-based approach that categorizes AI systems by their potential for harm. AI used in law enforcement and the justice system is classified as "high risk" and subject to stringent requirements around transparency, human oversight, and bias testing. The UK, post-Brexit, is charting a somewhat different course, emphasizing sector-specific guidance over comprehensive regulation. And in many jurisdictions across Africa, Asia, and Latin America, the regulatory frameworks simply don't exist yet—meaning AI evidence may be admitted or excluded based on individual judges' intuitions rather than established standards.
For practitioners in international disputes, this patchwork means that the admissibility and weight of AI evidence can vary dramatically depending on the forum. A predictive algorithm that's routinely relied upon in American sentencing might face serious challenges in a European court operating under GDPR and the AI Act's transparency requirements. Understanding these differences is no longer optional for lawyers who operate across borders.
The Road Ahead: What Needs to Happen
I'll be direct about what I think the legal system needs to do, and I'll be equally direct about where I think it's falling short.
First, judicial education must be treated as an emergency, not an initiative. The National Center for State Courts has published bench cards and guidance for judges confronting AI-generated evidence, and those are valuable. But a bench card is a band-aid on a structural problem. Judges need sustained, substantive training in how AI systems work—not at a computer science level, but at a level sufficient to evaluate the competing claims they'll hear from expert witnesses. The judge who spotted the deepfakes in Mendones did so through sharp observation. The next deepfake might not be so easy to catch.
Second, the legal profession needs mandatory disclosure rules for AI use. Louisiana's approach—requiring attorneys to verify the authenticity of digital evidence before submission—should be the floor, not the ceiling. Every jurisdiction should require parties to disclose when AI has been used to generate, enhance, or analyze evidence. The affirmative duty framework from the Weber case is a good model: put the burden on the proponent to identify AI involvement before the evidence is offered.
Third, the access-to-justice implications of AI evidence need to be confronted head-on. If challenging AI evidence requires expensive forensic experts, and introducing AI evidence requires costly technical preparation, we are building a system that advantages wealth over truth. Legal aid organizations, public defenders' offices, and pro bono networks need resources and training to engage with AI evidence on equal footing. Without this, Rule 707 and its state-level equivalents will simply become another mechanism through which the justice system's existing inequalities are amplified.
Fourth—and this is perhaps the hardest problem—we need to reckon with the philosophical implications of machine-generated truth. When an AI system enhances a surveillance video, it is making probabilistic judgments about what the missing visual data "should" look like. When a risk assessment tool generates a recidivism score, it is making predictions about human behavior based on correlations in historical data. These are not facts in the traditional evidentiary sense. They are computationally derived interpretations that carry the authority of mathematics without the accountability of human judgment. The law has always struggled with the boundary between fact and opinion, between observation and inference. AI evidence doesn't just blur that boundary—it threatens to erase it entirely.
Your Evidence Is Only as Good as the System That Created It
I started this piece with a scenario—an AI-generated forensic accounting report in a fraud trial. That scenario will become routine, probably within the next two to three years. The question isn't whether AI evidence will transform litigation. It already is. The question is whether we'll build the guardrails before or after the damage is done.
Right now, we're in the "before" window—barely. Proposed Rule 707 is under public comment. States like Louisiana and California are passing legislation. Judges are publishing bench cards and attending webinars. But the technology is accelerating faster than the institutions tasked with governing it. The deepfakes of 2025 were crude enough to spot. The deepfakes of 2027 likely won't be.
For lawyers, the imperative is clear: become literate in AI now, or risk being outmaneuvered by opponents who are. Understand how the tools work, where they fail, and how to challenge them. Build AI expertise into your trial preparation the same way you build in document review and expert witness preparation.
For judges, the message is equally urgent: you are the gatekeepers. The tools you have—Daubert, Frye, Rule 403, and soon perhaps Rule 707—are only as effective as your understanding of what you're gatekeeping against. Demand transparency. Require disclosure. Don't admit what you can't evaluate.
And for all of us, as citizens who might one day sit in a jury box or stand before a judge, the stakes are personal. The evidence that determines guilt or innocence, liability or exoneration, freedom or imprisonment is being increasingly shaped by systems that no single person fully understands. That's not inherently wrong—but it demands a level of vigilance, skepticism, and institutional reform that we haven't yet achieved.
The algorithm is on trial. The verdict will define what justice looks like for a generation.