Labelling AI content is not enough: the case for proof of authenticity

From 2 August 2026, anyone who publishes a deepfake or a text produced by an artificial intelligence system in Europe will have to declare it. This is the transparency obligation of Article 50 of Regulation (EU) 2024/1689, the AI Act, and it holds firm on that date: the Omnibus package discussed in Brussels does not postpone it. Whoever generates synthetic content will have to make it recognisable as such, with machine-readable marking and, in the case of deepfakes, with clear information to the people who see it.

It is a step forward. But it solves only half the problem. A label tells the viewer that a particular video, photo or audio clip was produced or altered by an AI. It says nothing about what is genuine. Flagging the fake is not the same as proving the true: the label is a warning about the synthetic, not a guarantee about the authentic. And that asymmetry matters, because trust is not rebuilt by marking what is artificial, but by demonstrating what is not.

The question the AI Act leaves open is therefore simple and uncomfortable: how do you prove that a piece of content is real? The answer is a paradigm that complements labelling rather than replacing it. Alongside the marking of the synthetic, we need certification at the source of the authentic: documenting the provenance, date, time and integrity of a piece of content at the moment it is created, before it circulates. That is the positive proof of the true. And it is what this article is about.

The structural limit of labelling: it flags the synthetic, it does not prove the authentic

Labelling has a limit built into it: it certifies what is artificial, but leaves everything else undefined. A piece of content without a label is not authentic for that reason alone. It might be genuine, or synthetic and never marked, or marked and then stripped. The absence of the signal is not proof of anything.

On top of this logical limit sits a practical problem of durability. The markings the AI Act calls for, from metadata to watermarks to provenance standards such as C2PA and Content Credentials, are fragile once content leaves its original environment. A share on a platform that recompresses files, a crop, a screenshot or a format conversion is enough for the embedded information to degrade or disappear. A watermark survives many manipulations but not all of them, and metadata is often removed at the very first hop.

The gap between the rule and reality is already measurable. An audit run by the Indicator observatory on a sample of roughly 516 suspect items identified about 169 of them as synthetic, and across many of the major platforms only a minority of AI-generated content was correctly labelled: the estimates gathered by Agenda Digitale put the share of correct labelling at around 31 to 33 per cent. In other words, two artificial items in three circulate without the mark they should carry. An obligation that in practice covers a third of cases is not enough to ground the trust of the person looking at the screen.

There is also a subtler effect. Once the label becomes the norm, its absence starts to be read as a guarantee of authenticity: "it does not say AI, so it must be real". That is exactly the reasoning an unmarked piece of synthetic content relies on to pass unnoticed. Labelling the synthetic reduces a risk. It does not build a proof.

The liar's dividend: when everything becomes suspect

Labelling, on its own, risks amplifying a side effect. The more obvious it becomes that any piece of content can be fabricated, the stronger the temptation to doubt everything, including what is authentic. This is the mechanism researchers call the liar's dividend: in an environment saturated with plausible fakes, someone filmed doing something real can dismiss the proof as "just another deepfake" and walk away.

The paradox is that labelling, designed to protect the public from the fake, can feed this dynamic. If the only available signal concerns what is synthetic, attention concentrates on suspicion rather than on verification, and the true and the false end up on the same footing of doubt. This is not a technological problem, it is a problem of trust: we have described elsewhere how the liar's dividend works and why it erodes the standing of evidence, testimony and documents.

Escaping this paradox means moving the point of leverage. As long as the only question asked is "is this content fake?", every answer stays contestable. The question that puts trust back on its feet is a different one: "is this content what it claims to be, and can we demonstrate it?".

Positive proof of the true: certification at the source as a complementary paradigm

Certification at the source is the documented record of a piece of content's characteristics at the exact moment it is captured: date, time, file integrity and provenance, fixed before the content enters circulation and can be copied, altered or taken out of context. It does not describe what a piece of content is not. It attests to what it is, when it was created, and that it has not changed since.

It is the exact inverse of labelling, and that is why it completes it. The label starts from content already in circulation and tries to signal its artificial origin after the fact, when the signals may already have degraded. Certification starts from the origin and binds the content to its history from the first instant, so that any later alteration becomes detectable by comparison. This is what makes the proof positive: not a judgement on the likelihood that something is fake, but a verifiable attestation of what is authentic. Anyone who wants to explore the foundation of this approach will find a dedicated analysis on the right to authenticity and positive proof of the true.

This paradigm does not replace the AI Act: it makes it operational where the rule stops. Article 50 establishes that the synthetic must be declared, and that remains an obligation to meet; anyone looking for the full picture can consult the analysis on the transparency obligations of Article 50. But no labelling rule answers the question facing the person who receives a piece of content and has to trust it: a court, a newsroom, a compliance office, a citizen. For them it is not enough to know that something is not marked as AI. They need to be able to demonstrate that it is genuine. The table below sets the two approaches side by side.

The two tools work together. Labelling reduces deception about the synthetic; certification grounds trust in the authentic. One signals, the other proves.

What it changes for businesses, media and institutions

For anyone who produces, publishes or assesses content, the shift from labelling to certification is not theoretical: it changes how you defend what is true. With roughly two AI items in three unlabelled on the platforms (Agenda Digitale), entrusting your credibility to the marking of the synthetic alone means leaving it in the hands of a system that covers a third of cases.

For businesses the stake is defensibility. A document, a photo of a construction site, the recording of a meeting or a screenshot of a conversation has value only if its provenance and integrity can be demonstrated when contested. In a setting where any piece of evidence can be dismissed as manipulated, certifying at the source turns a file into something you can rely on in front of a client, an authority or a judge. It is worth tying this need to the compliance checklist for businesses linked to the AI Act deadlines.

For the media the question is reputation. A newsroom that publishes an official video or a news photograph needs to be able to attest to its origin before it is reshared, cropped and recontextualised by others. A label on someone else's fake does not protect it; certification of its own material does. For institutions, finally, what is at stake is the resilience of public communication: a press release, an official act or an institutional video certified at the origin withstands pre-emptive challenge far better than one that can only be declared "not AI-generated".

How to certify authentic content at the source

Certifying at the source means capturing a piece of content with a forensic methodology and fixing its provenance and integrity at the very moment of capture, before it circulates. TrueScreen, the Data Authenticity Platform, does exactly this: it captures photos, videos, audio, screenshots and documents through a forensic process, verifies their integrity and certifies them with legal value, integrating the electronic seal and qualified timestamp of a third-party QTSP via API. TrueScreen is not a QTSP and does not issue qualified certificates: it is the platform that orchestrates the whole chain of custody and integrates into it the seal of a qualified provider, so that the proof is defensible.

The defining point is that certification happens at the birth of the content, not after the fact. Date, time, file integrity and provenance stay bound to the content in a way that remains verifiable over time, and any later alteration becomes detectable by comparison with the certified original. It is not a judgement on the likelihood that something is fake: it is a positive attestation of what is authentic, one that holds up against third parties.

A concrete example. A newsroom receives the official video of a statement from its communications office. Before publishing it, the newsroom certifies it at the source: at the moment of capture its origin, date, time and integrity are fixed. When the video is later reshared, cropped or placed next to suspect versions, the newsroom does not have to chase down every manipulated copy to deny it: it holds the positive proof that its own is the genuine original, with a verifiable chain of custody. The label would have flagged a possible fake; certification demonstrates the true.