When Seeing Is No Longer Believing: Evidence Isn't Dying, It's Moving

THE THESIS. For a century and a half, the image was the gold standard of public truth. A photo could convict in court, open a newscast and fix the historical record. That privilege is over. A deepfake — synthetic video or audio impersonating a real person — is now indistinguishable from the real thing and costs pennies. Our thesis has two parts. First: the race to detect fakes is technically lost, and the sooner we accept it, the better. Second: this does not mean «nothing can be believed anymore». It means trust is changing address. It no longer lives in the pixels; it lives in provenance — who captured the file, who signs it, who verifies it. It is a hard transition. But it is not the end of truth; it is the end of a shortcut that lasted 150 years.

THE LOSING RACE OF DETECTION. A deepfake detector is a program trained to tell real images from synthetic ones. In the lab, the best ones approach 96% accuracy, according to industry compilations. In the real world they collapse: the Deepfake-Eval-2024 benchmark, built from actual fakes found in circulation, recorded performance drops of 45% to 50% in the most cited detection models. The problem is structural, not an engineering bug awaiting a fix. Every detector learns the flaws of the generators that already exist. Every new generator removes exactly those flaws. The forger always moves last, and with the advantage. A recent academic paper says it in its title: this is «an unwinnable arms race» (arXiv, 2025).

Humans do worse. In an experiment by the verification firm iProov (2025), only 0.1% of some 2,000 participants correctly identified all the real and fake content shown to them. And the cost is already tangible. In January 2024, an employee of the engineering firm Arup transferred $25 million in 15 payments after a video call in which every participant, including the «CFO», was a deepfake, according to Hong Kong police. Cloning a voice takes between 3 and 30 seconds of public audio, as CBS News and the FBI have documented: the agency attributed $352 million in losses among Americans over 60 to AI-assisted scams in 2025 alone, according to US press reports. Deloitte projects that generative-AI-enabled fraud will reach $40 billion in the US by 2027. The most effective response today is a humble one: families agreeing on a code word to verify each other over the phone. Shared secrets are cutting-edge technology again.

THE LIAR'S DIVIDEND. The most corrosive effect of deepfakes is not that we believe lies. It is that the guilty can deny truths. Legal scholars Bobby Chesney and Danielle Citron named it the «liar's dividend»: when anything could be fake, shouting «it's a deepfake!» costs nothing. The data is in. A study published in the American Political Science Review (Schiff, Schiff and Bueno) ran five experiments with more than 15,000 US adults. The result: when a politician responds to a real scandal by claiming «misinformation», their support goes up. And it works better than apologizing or staying silent. The one piece of good news: the excuse still performs poorly against video evidence. That shield, however, weakens as deepfakes become normalized. The dividend is already being collected in court: the defense of a January 6 Capitol riot defendant suggested video evidence could be deepfaked, and Tesla's lawyers went as far as arguing that recorded statements by Elon Musk should not be admitted because they might be synthetic. Our read: the main damage is not done by the viral fake video, but by the cheap doubt that contaminates every true one.

SIGN THE ORIGIN, DON'T CHASE THE COPY. If detecting fakes is unworkable, the alternative is certifying what is authentic. That is the bet behind C2PA, an open standard — now an ISO norm — that attaches «content credentials» to every photo or video: a cryptographically signed history recording which camera (or which AI model) created the file and what edits it underwent. Adobe, Google, Meta, OpenAI, Sony, Nikon and Leica are on board; the initiative claims over 6,000 members. Leica was first to bring it to hardware: its M11-P camera has signed every image at the moment of capture since 2023. But the limits are serious, and they must be told. One: almost no smartphone signs its photos natively; Samsung's Galaxy S25, for example, only labels AI-edited images. The vast majority of the world's content is born without credentials. Two: most social networks strip that metadata when recompressing images. The chain of trust breaks exactly where it is needed most. Three: cryptography fails too; Nikon suspended the feature on its Z6 III and revoked its certificates after a signing vulnerability was discovered in September 2025. And we would add a design risk: the absence of a signature must never become proof of falsehood. If it does, we will penalize the witness with an old phone against whoever controls certified cameras.

COURTS AND NEWSROOMS RE-ARM. The two institutions that live off visual evidence are already under renovation. In US federal justice, the advisory committee on evidence rules decided not to amend, for now, the rule governing the authentication of evidence (Rule 901), but proposed a new Rule 707: machine-generated evidence would have to pass the same reliability filter demanded of a human expert witness. The committee vote is scheduled for May 2026; if it clears the full process, it would take effect in December 2027. The underlying message is clear: video no longer authenticates itself; you will have to prove where it came from. In journalism, the Associated Press bans generative AI in its news photography and requires verifying and labeling any synthetic material it reports on. Outlets such as the BBC, AP, Reuters, AFP and The New York Times have begun publishing images with content credentials, according to the standard's initiative, and Reuters is piloting with Canon and the Starling Lab workflows that certify a photo from the lens to the page. Photojournalism is not dying: it is being revalued. The signed agency photo will be worth more than ever against the anonymous flood.

OUR READ: LESS APOCALYPSE, MORE PLUMBING. Time to separate panic from data. The Knight First Amendment Institute (Columbia University) analyzed 78 election deepfakes from 2024, and its conclusion dismantles the dominant narrative: cheap fakes made without AI — crops, false captions, out-of-context videos — appeared seven times more often than AI-generated content. Harvard's Ash Center titled its election review «the apocalypse that wasn't». Disinformation is above all a demand problem — we want to believe what proves us right — not a technology supply problem. That is why we hold to our measured optimism. In the short term, the problems are real and will grow: industrialized fraud, the liar's dividend, slower and costlier litigation. In the long term, history offers perspective: journalism verified facts for centuries without photography — with witnesses, documents and source triangulation — and it will do so again with better tools. The implication for readers: distrust loose content and trust chains of custody (who publishes it, what signature it carries, who else confirms it). The implication for companies and institutions: provenance stops being a technical detail and becomes critical infrastructure, just as the HTTPS padlock — the encryption protecting websites — was for e-commerce. Nobody says the internet killed shopping because anyone can build a fake storefront: we built trust plumbing, and online commerce flourished. The same will happen with images. Seeing will no longer be believing. Verifying, though, will be easier than ever.