The use of artificial intelligence to write motions and briefs, not to mention the occasional opinion, containing hallucinated caselaw and quotations is already a well-known problem. Not only does it reflect the likelihood of a lawyer cheating the client out of competent representation and fees charged for work done by computer rather than the retained counsel, but with a modicum of efforts, it’s revealed to be an AI sham and ends with the client getting slammed and the lawyer being humiliated and sanctioned. All as it should be.

Eugene Volokh points to a series of websites created by Damien Charlotin to track AI cases.

Damien Charlotin, who runs the invaluable AI Hallucination Cases site (2200+ cases worldwide and counting) has a new one, AI Deepfakes in Court, listing 26 cases worldwide so far. His summary:

This database tracks decisions dealing with deepfakes submitted as evidence, averred or alleged, including texts and documents such as forms, records and correspondence. A failure to authenticate an exhibit is recorded separately from any finding that it was AI-generated.

It does not track cases about deepfakes as a subject: the focus is on cases started on any ground, where parties (allegedly or verily) submitted deepfakes as evidence to bolster their claims. This is distinct from using AI to fabricate citations, which is covered by the AI Hallucination Cases database, or using AI as evidence (e.g., to offer a translation or a common sense opinion), which is covered by the AI Evidence Cases database.

Hallucinations are one thing, but AI deepfakes? Years ago, I began the “but for video” series that addressed the fact that few outside of the criminal defense bar believed that cops behaved poorly, sometimes violently, because there was no proof of wrongdoing. “Why would the cop do that,” a judge would ask, “when he doesn’t know the defendant and has no reason to lie about him or harm him?” An excellent, but irrelevant, question. Defendants would complain and judges would shrug because there was no evidence beyond the defendant’s word. Judges tend not to take the defendant’s word very seriously, in case you were wondering.

Then came video proving conclusively that these things happened. Why remains a mystery in many cases, although it was easily enough attributed to police culture, bigotry and power trips. But with video of the interaction showing a cop needlessly beating or killing a defendant, the “why” question no longer mattered as much. There is was, in black and white. The cop did it and the video proved it.

Enter AI deepfakes. In the beginning, they tended to be pretty obvious, giving people six fingers or two left feet. They could be funny and embarrassing, even though obvious, as there’s something about seeing that makes it more real than it deserves to be. Indeed, docudramas have seriously twisted our grasp of reality as real life stories and images are intertwined with fake political narratives and dialogue in order to make people believe that the lie presented really happened. And it works. It works really well.

But AI deepfakes have gotten better and, unsurprisingly, found their way into the courtroom. The problem is that judges are no better equipped to deal with them than they were with pre-video claims of police misconduct. If it looks real and comes into court with a story that sounds remotely possible, who is to say that it’s a deepfake?

This raises two related issues. The first is that a deepfake video is submitted into evidence, apparently “proving” a fact that is totally false. The second is that a real video is submitted into evidence and challenged as a fake, thus raising doubts that the video proves anything.

Sure, there are putative experts who are available for hire to opine about whether a video is real or AI, but that costs money that the defense often lacks. In addition, just because someone holds themself out to be an expert doesn’t make them one, as lawyers are already too well aware.

In order to admit a video into evidence, the party offering it must lay a foundation to establish its authenticity. Given the ubiquity of security cameras, this isn’t particularly hard to do. A cop testifies that it was taken from video of a bodega across the street and, boom, that’s good enough. After that, the challenge is to weight and sufficiency, which can be a nearly insurmountable burden. Remember Brandolini’s Law:

The amount of energy needed to refute bullshit is an order of magnitude larger than the energy needed to produce it.

To recognize this is to belabor the obvious in the age of AI. Video is exceptionally powerful evidence, and merely testifying that it never happened doesn’t make the jury unsee what is proffered as “what really happened.” It’s facile to believe that this can be addressed by regulating AI deepfakes, as if ignoring the regulations somehow reveals the fact that it’s a deepfake. Absent affirmative proof that the video is fake, it can prove impossible to distinguish real from deepfake, or vice versa.

Whether the problem is “seeing is believing” or “believing is seeing,” the existence of AI deepfakes is upon us and the quality will improve until distinguishing real from fake will be essentially impossible. The days of “but for video” were heady for a while, but those days are numbered unless a way is found to address a future of fake videos proving lies beyond a reasonable doubt.