When the Prime Minister's Voice Isn't the Prime Minister's

Researchers tracked 82 deepfakes targeting public figures across 38 countries in a single year. Watermarking catches the ones that play by the rules. Detection has to catch the rest.

A fabricated video of Keir Starmer criticizing his own party circulates online. A fake clip has Taiwan's Ko Wen-Je making accusations he never made. In Turkey, a deepfake links an opposition leader to a terrorist group. Ahead of India's Bihar assembly elections in late 2025, an AI-generated video of Narendra Modi's late mother appears, criticizing his politics. Researchers tracked 82 such deepfakes targeting public figures across 38 countries in a single 12-month span, and the raw volume of deepfake files in circulation has climbed from roughly 500,000 in 2023 to a projected 8 million in 2025.

Some of these are political. Many are financial: fake endorsement videos of Justin Trudeau and Claudia Sheinbaum have been used to push investment scams, borrowing a trusted face to sell something the real person never touched. Either way, the platform hosting the content is the one left holding the moderation problem, at a scale no team of human reviewers can keep pace with.

The industry's first answer was watermarking: content-credential schemes that tag media at the point of generation, so a platform can check the tag and know what it's looking at. That works for generators that participate. It does nothing for the ones that don't, and the generators that matter to a fraud or disinformation campaign are precisely the ones with no interest in tagging their own output. Watermarking verifies origin. It doesn't detect fakery.

What closes that gap is detection that works on the content itself, independent of whether it was tagged, tracing the visual and audio artifacts a synthetic generation process leaves behind rather than trusting a label the generator chose to include or leave out. That's a harder problem, and it's also the fastest-growing segment in the field: deepfake and synthetic-media detection is estimated to reach $5 to 8 billion by 2030, growing at 43 to 48 percent a year, because platforms have run out of easier options.

It's also a moving target. A detector trained on last year's deepfake generation techniques will miss this year's, the same way frontier vision-language models, tested against realistic synthetic identities, currently accept them as genuine 94 to 100 percent of the time. Detection only holds if it's continuously retrained against generation techniques as good as, or better than, whatever's circulating.

That's the pairing that makes detection durable: a generator that keeps producing the adversarial material a detector needs to stay current, and a detector trained on it before the next wave of fakes arrives rather than after. TessLabs supplies exactly that engine, built and proven inside a live identity-verification business handling real fraud at population scale for eight years.

Read the white paper to see how the detection stack is benchmarked.

Case study: political-deepfake tracking via Recorded Future and CETaS.