It’s the morning of the first day of trial. Your opponent calls her first witness who testifies about a video he says was taken at the accident scene. The video clearly shows your client running the red light. The witness is pointing at the screen and saying the video is a fair and accurate depiction of what he observed. The judge nods his head knowingly and looks at you. Your client is tugging at your sleeve and whispering something. There is something vaguely off about the footage.
What do you do? How do you attack the presumed authenticity? Should you? What if the video is more or less authentic but has been enhanced? Should you mention that? How?
Deepfakes and evidence created or enhanced by AI are going to become increasingly prevalent. There are numerous examples but few solutions or answers for lawyers like the above, for judges who are evidentiary gatekeepers, and for jurors who are often the ultimate decision-makers in court.
That’s why what the Visual Evidence Lab at University of Colorado Boulder recently created and did is important. The Lab gathered 20 some experts from academia, law, media forensics, journalism, and human rights practice in April of this year to discuss the use of video and AI for a full day and to talk about the problems that AI can and is creating in our courtrooms. The group released a report entitled, Video’s Day in Court: Advancing Equitable Legal Usage of Visual Technologies and AI.
While the focus of the group was on video evidence, much of what was discussed is applicable to other forms of non-documentary evidence. The group talked about three key things: systematic public access to and storage of video evidence, how to place guidelines on the interpretation of video evidence by judges and juries to mitigate bias and properly interpret the evidence, and the issues posed by the impact of AI on video evidence to better establish and ensure reliability and integrity.
The Access Problem
The group was concerned about access since unlike documentary evidence, video evidence is haphazardly stored. Why is that important? It prevents researchers and others from being able to grasp the scope of the problem and the risks it poses. It also precludes a meaningful analysis of the characteristics that might indicate a deepfake: “These visual materials cannot become a proper part of common-law jurisprudence either because lawyers and judges are not able to refer in any reasoned fashion to decisions of other courts regarding comparable videos.”
Frankly, I had not thought of this issue. But as we shall see later in the report discussed below, the lack of the ability to understand the scope and magnitude of the problem hampers the ability to systematically deal with it.
You can’t solve a problem with anecdotes instead of facts. But anecdotes are all we have right now.
And the access problem is only the beginning.
The Interpretation Problem
The impact of video evidence is different than documentary evidence in ways that are often misunderstood. There’s lots of psychology research that shows perception of video evidence can be more selective, biased, and shaped by what the report calls motivated reasoning, that is, using the evidence to support a preexisting conclusion.
In addition, the video medium can be manipulated to shape interpretations. Things like playback speed can alter the perception of video evidence: it makes the depicted action seem more deliberate. Other factors including camera angle and field of view are important. The report concludes, “Despite the multiple factors shaping interpretation and decision-making, judges, lawyers, and jurors are largely unaware of the various influences on how they construe what they see in a video.”
Put bluntly, video evidence, by its very nature, impacts decision-making in ways that are different than other evidence. There is precious little study of how this impacts decision-making in the courtroom and how altering or enhancing the video can impact that reasoning. Without that, it’s hard to know what is fair and how to define what is impartial when it comes to decision-making.
For example, is it fair for a jury to be presented with an enhanced video to better demonstrate a bloody and brutal injury? Or does that place jurors too close to the victim and interfere with fairness?
The Impact of AI
All of these issues are compounded by AI, the report concluded. It’s hard to confidently distinguish whether a video accurately depicts what it is being offered to show, the standard test of authenticity. Three questions arise:
- The difficulty detecting and verifying AI-created media
- The uncertainty about what kind of enhancement is permissible in court
- The fear that deepfakes may become more prevalent
Here’s the problem: as noted by the report, the Advisory Committee on Federal Evidentiary Rules decided in May of this year that no changes to Evidentiary Rule 901 which governs authenticity were necessary. Why? Because the Committee concluded so few deepfakes had been offered as evidence. (Of course, that assumes that all “deepfakes” had been found, labelled, and that labelling recorded in a way that could be accessed, which gets back to the first problem.) The Lab report notes:
The central challenge is how to establish robust authentication standards that can withstand scrutiny, without simultaneously creating verification systems that compromise people’s right to confront evidence or endanger the human rights of media creators and witnesses.
The report also noted that courts have long allowed the use and admission of technologically enhanced media like enlarged photos and interactive 3D models. But AI tools bring new levels of enhancement not seen before.
Moreover, the ease of use and affordability of these tools make them ubiquitous. Things like changes to resolution, brightness, contrast, sharpness, and other features allow video evidence (and photographic evidence for that matter) — features we all use every day, by the way — to be presented in new and persuasive ways.
Here’s a real-world example of a problem with video. In a previous life, I was a swim official. One of the calls a swimming official makes is to make sure in relay events no swimmer leaves the blocks before his teammate touches the wall. The only way to do that is to stand right next to the block. I can’t tell you how many times a spectator would come to me with a video taken 30 yards away to dispute a call.
That video, of course, is not an accurate depiction of what actually happened. But the spectator would extrapolate what actually happened from that video.
The question is at what point do those kinds of enhancements cross the line between what is convenient and proper and become a deepfake? We have no firm, universal rules to determine this. Without these rules, inequalities exist which undermines a consistent application of the rule of law.
There is, by the way, a proposed amendment to Evidentiary Rule 707 that would apply the Daubert standard of reliability to determine the admissibility of AI-enhanced and -generated evidence. It is open for public comment until February 2026.
All of this, combined with the fear that deepfakes are going to become more and more prevalent, all raise issues of evidentiary integrity, says the report.
What Is There to Do?
The Colorado gang didn’t just stop at identifying a problem, they came up with several recommendations to get us to some solutions:
- The development of standards for labeling, storing, securing, and archiving video evidence. This would include a data strategy along with a decentralized architecture that would enable use and analysis of that data.
- The development of visual evidence training for judges (e.g., how to probe and ask relevant questions) to better perform their role as gatekeepers.
- The development of research-based guidance to help jurors better evaluate video evidence.
- Systematic research into the prevalence of deepfakes in court to develop safeguards for AI-generated evidence.
- The issuance of ethics opinions on the offering of known or suspected AI-generated or -enhanced evidence.
According to the report:
Judges must be prepared to handle cases involving AI-generated and AI-enhanced video evidence. Improving notice and disclosure for AI-enhanced evidence can help safeguard reliability without further exacerbating the inequality of access to justice.
The Report Conclusion
The Report concluded as follows:
The development of a long-term infrastructure for storing and accessing evidentiary videos, research-based training for judges, instructions for jurors, and safeguards for the admission of AI-based evidence will advance the consistent and fair use of video and AI technologies in the pursuit of justice.
Some Final Thoughts
Yes, the report is short on concrete, practical solutions. It’s one thing to say we need to do things like educate judges. It’s another thing to create training modules and roundtables to do just that. The former is easy, the latter harder.
But what the Lab has done is a start. It’s a studied, inclusive, and fair examination of a problem that’s only going to get worse without action. While the devil is often in the details, you don’t get to the details without understanding the problem you are trying to solve. That’s what the Colorado group is doing. That’s what we need more of if we as a profession are going to successfully confront the problem.
Until we get serious about understanding the scope of this problem, we’re just playing courtroom roulette with the truth.
Stephen Embry is a lawyer, speaker, blogger, and writer. He publishes TechLaw Crossroads, a blog devoted to the examination of the tension between technology, the law, and the practice of law.
The post The Deepfake Courtroom Problem: A Colorado Blue Ribbon Study Sheds Some Light And Offers A Start To Solutions appeared first on Above the Law.