Deepfake Evidence in Court: How Judges Are Deciding What's Real in 2026

Every litigator has spent the last two years worrying about the wrong end of the AI evidence problem. The conversation has been dominated by hallucinated case citations, a lawyer’s own AI tool inventing an authority that gets caught and sanctioned. That problem is serious, but it is self-correcting: a fake case is checkable against a real database in minutes. The harder problem, and the one with almost no settled answer yet, is what happens when the evidence itself, the video, the audio recording, the WhatsApp screenshot, the voice note, might be synthetic, and there is no equivalent of a case citation database to check it against.

This is now described in the evidence literature as an authentication paradox: a deepfake detection tool can report an 86% probability that a piece of content is synthetic, and the content can still be admitted, because a probability score is not proof, and courts have no settled standard for what to do with a number like that. Judges, legislators and rule-makers in every jurisdiction The AI Bar covers are responding to this right now, at different speeds and in different ways.

How Each Jurisdiction Is Actually Responding

United States: Rewriting the Authentication Rule Itself

A formal proposal is before the Judicial Conference’s Advisory Committee on Evidence Rules to add a new subsection, Rule 901(c), to the Federal Rules of Evidence.

The mechanism is burden-shifting: once a challenger presents enough to support a finding that content may be AI-manipulated, the burden shifts to the proponent to satisfy ordinary Rule 901(b) authentication plus additional proof of reliability.
The judge, not the jury, decides admissibility under Rule 104(a).
Status: a rules suggestion working through committee, not yet adopted. But it signals where common law evidence doctrine is heading, away from a low authentication bar and toward a two-stage test triggered the moment a deepfake challenge is raised in good faith.

India: Two Frameworks That Only Work Together

Section 65B of the Bharatiya Sakshya Adhiniyam already imposes a demanding certification regime on electronic evidence.

It requires: the record comes from a computer used regularly for that kind of information, the data was fed in during ordinary business activity, the system operated properly throughout, and a certificate from someone in a responsible position over the device accompanies the evidence.
The gap: this framework was built for tampered or altered records, not content generated from nothing. A fabricated video can satisfy every element of section 65B while being entirely fake at its source.
The IT Rules 2026 close part of that gap without replacing section 65B, imposing labelling, watermarking and persistent metadata requirements on intermediaries for synthetically generated information, plus platform logging duties that create an independent provenance trail.
Practical consequence: authenticating contested evidence now requires both frameworks together. A section 65B certificate alone, without platform-side metadata and logs, is no longer sufficient practice for content you suspect is synthetic.

Singapore: When a Presumption of Reliability Becomes the Problem

Section 116A of the Evidence Act creates a rebuttable presumption that electronic records produced by a properly functioning device were accurately produced.

That presumption exists to stop ordinary printouts and database records requiring expensive forensic proof before admission.
Applied to a well-made deepfake, the same presumption becomes a liability: a synthetic video played from a properly functioning device can satisfy the presumption’s technical requirements without saying anything about whether the content itself is genuine.
Practical consequence: the burden is functionally on the challenging party to affirmatively rebut the presumption the moment manipulation is suspected, not to wait for the other side to prove authenticity.

United Kingdom: A Growing Problem With No Dedicated Rule Yet

No amendment to the evidence rules specifically for AI-generated content exists so far.

Family courts are already seeing the practical version of this: contested audio or video in custody and financial remedy proceedings, alleged to be manipulated with consumer-grade AI tools, decided under ordinary relevance and weight principles rather than any deepfake-specific test.
The risk: family proceedings often move faster than the forensic resources needed to test contested media.
Expect judicial guidance to follow the UK’s usual pattern on AI, cautious, principles-based, and reactive to specific cases rather than a wholesale rule change.

What This Means for Your Litigation Practice Now

The direction of travel is the same across jurisdictions: the presumption that a recording, image or document is what it purports to be is weakening, and the burden of proving otherwise is shifting earlier, onto whichever side raises the challenge first and most credibly. Three changes are worth making now, ahead of formal rule changes catching up.

Build authentication in at collection, not at challenge. Contemporaneous hash values, preserved metadata and documented chain of custody cost little to capture early and are often impossible to reconstruct once a deepfake challenge is raised later.
Treat a forensic authentication expert as standard for any contested electronic evidence with real stakes, the same way you’d already instruct a handwriting expert for a disputed signature. Detection tools remain probabilistic and largely proprietary, so the expert’s methodology, not just their conclusion, needs to survive cross-examination.
Raise authentication challenges early and specifically. A generic objection that evidence “could be AI-generated” carries little weight. A challenge identifying specific metadata gaps or provenance failures is what actually shifts the practical burden, whether or not your jurisdiction has formally codified that shift yet.

For India specifically, that first point translates into a concrete checklist:

Secure original devices and first-generation files, not forwarded copies.
Generate cryptographic hash values contemporaneously with collection.
Obtain a properly particularised section 65B certificate naming the device and certifying individual.
Request platform-side logs and synthetic content metadata early, before they age out of retention.
Instruct independent forensic analysis examining both the file and platform-side provenance markers, not either alone.

This is precisely the kind of evidentiary judgment we build into the Litigation Support module of The AI Bar Foundation for Lawyers, including how to instruct forensic experts on AI-generated content and how to frame an authentication challenge a court will actually engage with rather than dismiss as speculative. The tool that can fabricate evidence is not the same tool that can tell you whether it’s fake. That distinction is the whole of this problem, and it’s only getting more litigated from here.