When Any Video Can Be Fake, Who Tells the Court?

A video used to be the end of an argument. Now it can be the start of one. Deepfakes make two kinds of injustice possible. A fabricated clip can convict someone who did nothing. And a genuine clip can be waved away by a defendant who says it was made by a machine.

Researchers call the second effect the liar’s dividend, a term coined by the American law professors Robert Chesney and Danielle Citron. The more people know that fakes exist, the cheaper it becomes to deny real evidence. Europe has spent most of its energy on the first problem.

Laws for the victim, not for the courtroom

A law that gives a remedy to someone whose face is used says nothing about what a judge does when a video appears in a trial. Where courts evaluate evidence freely, as is common in continental systems, no formal gate decides whether a recording may be considered. That worked well when fabricating a convincing video was a specialist job. It works less well when a teenager with a phone can do it in minutes. That leaves the justice system exposed.

In Europe, the criminal law moves faster than the evidence law

Italy. The Italian parliament approved a national AI law on 17 September 2025 that punishes people who create or spread harmful AI-generated content, including deepfakes, with one to five years in prison. Critics say it lacks an independent oversight body and limits on biometric recognition in public spaces.

Read also: Brussels Pushed Its Hardest AI Rules Back 16 Months

The EU. The AI Act requires deepfakes to be labelled as AI-generated. Labelling is a transparency duty that helps the public. It does not tell a criminal court how to test a video offered by the prosecution or the defence, and a forger will not label his work. The Act is a rule about what providers and deployers must do. It leaves the rules of evidence to the member states.

Read also: Europe’s Surveillance Rules Are Being Written Around the Exceptions

The United States as a comparison. American rule-makers have tried to answer the question directly and have stalled. The federal judiciary’s advisory committee on evidence rules examined a proposed Rule 707 on machine-generated evidence and a proposed Rule 901(c) on alleged deepfakes. On 7 May 2026 it did not advance Rule 707 and did not release the deepfake rule for public comment.

Under the draft, once an opponent made an initial showing that AI had fabricated an exhibit, the side offering it would have to prove it authentic on a higher standard. A Federal Judicial Center survey found 15 judges who had met deepfake issues, and a majority favoured a rule. Americans have not settled this either.

The new labelling code, and what it cannot do

On 20 July 2026 the Commission published a code of practice on marking and labelling AI-generated content, together with guidelines on the AI Act’s transparency duties. About 190 companies had signed by 31 July. The code asks providers to apply at least two machine-readable marking techniques to synthetic output, to offer a way of detecting marked content, and to make their systems interoperable by 2 February 2027.

Anyone who publishes a deepfake has to disclose it, using a common EU icon, with a lighter touch for evidently artistic or satirical work. The transparency duties apply from 2 August 2026, and some of the provider duties on marking are deferred to 2 December 2026. Breaches can cost up to 15 million euros or 3 percent of worldwide turnover.

This is a real step for the public square. A reader who meets a clip made with a mainstream tool will increasingly be able to see that it was synthetic. For a courtroom, the limits show quickly. A mark exists only if the tool applied it, and a model running on a private machine owes the code nothing. A mark can also be damaged by editing, which is presumably why the code asks for more than one technique. A forger who wants to convict a rival has no reason to sign up.

The deeper problem is the reverse case. A missing label proves nothing about whether a recording is genuine. If labelling becomes the norm, the unlabelled clip starts to look like the authentic one, and genuine footage that carries no provenance data, such as a phone video from a street corner, becomes easier to challenge. Labels help the audience. They do not replace testing in court.

How the liar’s dividend works in a trial

Consider the cost structure. A defendant who says a recording is fake does not have to prove it. Raising doubt is cheap, while a forensic examination of a video by a qualified expert is slow and expensive, and legal aid budgets are thin. If the prosecution cannot afford the examination, or the court is unsure who must pay for it, the doubt does its work.

The reverse is just as dangerous. A fabricated clip that looks real, offered by a party that can pay for a polished presentation, meets a court that has no routine for asking where the file came from. This is analysis, not a measured trend.

What a workable answer looks like

Three steps would help. A working group in the justice ministry to look at technical verification requirements and a national panel of experts. Training for judges and prosecutors in recognising manipulated material. And a criminal offence for submitting AI-generated evidence with the intent to mislead a court.

Two more belong on the list. Provenance: cameras and phones that sign recordings at the moment of capture make it possible to show that a file has not changed since. And disclosure: a party that challenges a recording as fake should be expected to state why, instead of simply asserting it.

The right to test the evidence

The right to a fair trial includes knowing the evidence against you and being able to test it. When any recording can be dismissed or fabricated, both halves of that right weaken.