Generative AI did not create one evidence problem. It created two, and the second is sneakier than the first.
The first is obvious: synthetic content good enough to pass. Fabricated videos, cloned voices, invented conversations — offered as if real.
The second is the mirror image, and it is already in courtrooms: the deepfake defense. When any recording could be synthetic, denial becomes free. "That video is a deepfake" costs nothing to say — and was said, notably, by Tesla's lawyers in 2023 about a widely-seen recording of Elon Musk's Autopilot statements. In 2025, a California judge dismissed a case upon finding that the plaintiffs had submitted AI-generated video testimony. Both directions of the problem are live.
The rules are moving
US rule-makers have spent several years on amendments aimed squarely at this. Proposed Rule 707 would subject machine-generated evidence offered without a supporting expert to the same reliability standard as expert testimony. A proposed Rule 901(c) addresses deepfake challenges directly, with a burden-shifting structure: a challenger must first produce enough evidence of fabrication; the proponent must then demonstrate genuineness. The proposals completed public comment in early 2026; if adopted, the earliest effective date is late 2027. States are moving too — Louisiana enacted the first framework for AI-generated evidence in 2025.
The details will evolve. The direction will not: the burden of showing authenticity is drifting toward whoever offers the content.
Why after-the-fact detection loses
The instinctive response — analyze the file, spot the fake — is a losing race. Detection tools chase generators that improve monthly, and a "probably authentic" score is exactly the kind of contestable opinion that keeps disputes alive.
The durable answer inverts the timing: instead of inspecting content afterwards for signs of fabrication, build the evidence of origin at the moment of capture. A record made in a controlled environment, with documented context, integrity fingerprints from the first second, a hash-linked custody chain, and independent time anchors, gives the "it's a deepfake" objection something concrete to break against: here is where it came from, when, and proof it has not changed since.
The honest limit — and the practical conclusion
Provenance-at-capture protects what passes through it from the moment it passes through. No tool can retroactively certify a stray file of unknown origin, and claims to the contrary deserve suspicion.
Which is exactly the practical conclusion: in a world where authenticity is challenged by default, the cheap move that works is to capture through infrastructure that documents origin as a matter of course. When the challenge comes — and the trend says it will — the answer is already in the record, not assembled in a scramble afterwards.