In January 2024, a high school principal outside Baltimore was accused of an ugly, racist and antisemitic rant — an audio clip circulating through parent group chats that sounded exactly like him, word for word, inflection for inflection. He was suspended, threatened, and publicly ruined within a day. Months later, police determined the recording had been produced by the school's athletic director using AI voice cloning, as an act of revenge. The voice was a flawless forgery. The man it belonged to had never said any of it.
That is half of the new problem facing courts. The other half is worse: once a court accepts that audio and video can be forged convincingly, any defendant has an incentive to call the real evidence fake too. The deepfake walks into the courtroom from two directions at once — as a weapon of fabrication and as a shield of denial. The first cases have already worked their way through the system, and how they were resolved is a rough draft of the rules you may have to live under.
The Pikesville unmasking
The case of Eric Eiswert, principal of Pikesville High School in Baltimore County, became the first widely reported example of what digital forensics specialists now call "synthetic media as a weapon." A recording purporting to be Eiswert making racist and antisemitic remarks about students and colleagues spread across social media in January 2024. The reaction was immediate and real: suspension, public condemnation, threats against his family.
The truth surfaced not instantly and not from the recording itself, but from digital forensics experts who dissected the artifacts of synthetic speech: unnatural pauses, spectral seams where an AI model had stitched phonemes together, the absence of the room-tone background noise a genuine recording would carry. The Baltimore County investigation eventually charged the school's athletic director, who had a documented grudge against Eiswert. The case closed with a relatively fast unmasking — but that speed was the exception, not the rule. Debunking took weeks of specialist work, and it happened only after the real damage was already done.
The lesson of the case for courts is this: the technology for making a forgery has pulled well ahead of the technology for detecting one, and the speed of social media has outrun both. By the time a forensic examiner is ready to say "this is synthetic," the public has already reached its own verdict.
The liar's dividend: when the truth gets called a fake
The second precedent is stranger and, arguably, more dangerous — because it isn't about a forged piece of evidence at all. It's about the attempt to declare genuine evidence forged. In 2023, during litigation over a fatal crash involving Tesla's Autopilot in California, the company's lawyers argued in court filings that Elon Musk's own past public statements about Autopilot safety — including years-old interviews — could theoretically be deepfakes, and therefore should not automatically be treated as authentic evidence of what he'd said. The filing drew wide attention precisely because it showed the logic in its purest form: if any recording is potentially fake, a defendant can reject any inconvenient piece of evidence simply by asserting that it is.
Legal scholars call this "the liar's dividend" — a term researchers studying digital evidence law have used since the late 2010s: the more the public knows deepfakes exist, the easier it becomes for a genuinely guilty person to convince a court, or the public, to doubt real footage that actually incriminates them. This isn't hypothetical — it's already a defense strategy showing up in domestic violence cases, custody disputes, and criminal proceedings where a genuine video recording is the evidence at stake. According to attorneys practicing in family courts in the UK and the US, attempts to dismiss authentic audio recordings or messages as "AI-generated" have become noticeably more common as a tactic over the past two to three years — though no single body yet tracks the scale of the phenomenon reliably.
Who gets to weigh the evidence now
Courts move slowly because they have to move carefully: an error in either direction — accepting a forgery as truth, or dismissing truth as a forgery — wrecks a life. The Judicial Conference's Advisory Committee on Evidence Rules in the US has been publicly discussing, since 2023, an amendment to Rule 901 of the Federal Rules of Evidence — the rule governing how a party must "authenticate" digital evidence. The proposals boil down to making explicit that a party claiming a recording is a deepfake must offer a concrete basis for that claim rather than simply asserting the word — otherwise any piece of evidence can be neutralized by a bare accusation.
Alongside that, a whole forensic industry for synthetic media has grown up: blink-rate analysis on video, spectral analysis of voice recordings, file metadata review, and content-provenance watermarking — some camera manufacturers and AI platforms have begun embedding it under the C2PA standard. None of these methods offers a guarantee. The race between forging and detecting is structured so that every new detection method eventually becomes training data for the next generation of generators, which learns to defeat it.
Our record: In the system of Maat, Thoth governs the written word — the transmission of meaning through a sign. But like every archetype, Thoth has a shadow. Shadow Thoth is not lying itself; it is the instrumentalization of the very mechanism that makes a sign trustworthy — propaganda, forged documents, a voice impersonating a different voice. A deepfake in court is exactly this shadow taken to its limit: not a person lying under oath, but the recording itself — the thing a court has trusted as an honest witness for centuries — beginning to lie in place of the person. The scale of Maat weighs the heart against the feather of truth. What happens when the feather itself can be counterfeited?
What this means for you
You are unlikely to end up a party in a Tesla Autopilot lawsuit. But a divorce, a custody dispute, a workplace conflict where someone releases a "recording" of your voice or face — that scenario is far more plausible than it sounds, and ordinary cases like these — family court, employment disputes, defamation suits — are already colliding with deepfake evidence more often than the headline criminal trials, simply without the press coverage. Check whether your phone or primary camera has Content Credentials (the C2PA standard) enabled — some recent devices and camera apps already carry this in their settings: it is an invisible signature that, in some future dispute, could become your fastest way to prove a recording is genuine.
There is no technoskepticism to be had here for its own sake — courts are not going to abandon audio and video evidence, because the whole of modern life is recorded. But the rule being set right now, in these first, still-clumsy cases, is a rule about where the burden of doubt lands: on the person claiming a forgery, or on the person presenting the recording. As long as that burden sits in the wrong place, anyone with bad intent and access to a cheap AI tool gets a ready-made alibi, and any victim of a genuine forgery gets a ready-made way to have her word discounted a second time. Sorting that out isn't work for some future regulator. It's work every juror, every reader of the news, every person deciding whether to believe a recording does, before the recording gets to decide it for them.