nuclai
Explainers

Video, voice, text: can we still spot AI?

By eye, less and less: in a study by Tavus, the company that makes it, 48% of participants took its video avatar for a human. Through marking, a little more: Anthropic and OpenAI add an invisible watermark to text produced by their AI. OpenAI’s applies only in the EU, and rewriting weakens it considerably.

Offprint: Video, voice, text: can we still spot AI?

29.10.2026 — Summary. By eye, less and less: in a study by Tavus, the company that makes it, 48% of participants took its video avatar for a human. Through marking, a little more: Anthropic and OpenAI add an invisible watermark to text produced by their AI. OpenAI’s applies only in the EU, and rewriting weakens it considerably.

Updated on 29.10.2026

By eye, less and less

On 1 October 2026, the American company Tavus unveiled Griffin, a model that generates, in real time, a person who speaks, listens and reacts on video. To test it, Tavus ran its own study with 54 participants. Each believed they were having a one-minute video call with another participant. On the other side was Griffin-Lite, a preliminary version of the model. 26 people out of 54, or 48%, thought they were talking to a human. We watched the demos published by Tavus, and it took us a good minute to spot the virtual agents.

The trend is clear enough: visual hyperrealism is no longer the only mark of a successful digital avatar. With Griffin, the difference now lies in the smallest details: natural gestures, a calculated hesitation, a hint of an accent that makes the exchange feel human.

Our reading: this technical finesse opens up a fascinating prospect for work, that of the made-to-measure colleague. By adapting the agent’s features, temperament and way of relating to how we ourselves work, we could reduce from the outset the human friction found in any team. Less energy lost negotiating personalities, more operational alignment: the virtual colleague would then become a direct amplifier of our skills.

Through marking, a little more

Another approach is to mark content at the moment it is produced. Anthropic led the way in August 2026: text generated by its Claude models carries a watermark, applied worldwide. In October, OpenAI followed by announcing its own, called textGrain, for text generated with ChatGPT and Codex, in the EU only.

The principle is similar for both. When the model hesitates between roughly equivalent words, a secret key subtly steers its choice. The reader sees nothing, but a detector can find the pattern. No invisible characters are added and the meaning of the text does not change. In both cases, the detector is not public: it is reserved for certain organisations, such as researchers or authorities.

The watermark also has its limits, quantified by OpenAI. On English passages of around 300 words (400 tokens), the detection rate falls from about 92% to 66% when 10% of the words are replaced with synonyms, and to about 17% with 25% synonyms. According to OpenAI, translation or heavy rewriting also weakens it. Finally, a watermark does not measure how much of a text is human work.

What does the law say?

In the EU, since 2 August 2026, the EU AI Act has required deepfakes to be disclosed as such, with a lighter regime for artistic or satirical works. Switzerland has no equivalent rule, but the Federal Data Protection and Information Commissioner (FDPIC) considers that the use of tools capable of altering the face, image or voice of an identifiable person should always be clearly indicated.

What still helps to spot it

None of these clues is proof, but together they often say more than the image itself.

  • Context. An unusual or urgent request, or one that comes outside the usual channels, deserves attention, however good the video or voice.
  • Source. Who is publishing, on which account, with what track record: these questions apply to generated content too.
  • Cross-checking. The same information, found on a known channel (the official website, a contact’s usual number), confirms or contradicts what was seen or heard.
  • The makers’ tools. OpenAI offers a public tool to check its images and audio; for text, detectors remain reserved for specialists. The absence of a mark proves nothing, though: it may have been removed, or the content comes from a tool that does not mark.

Our takeaway

Marking is making progress, but it depends on the maker, the region and what is done with the content afterwards. For a communications team, the question therefore works both ways: how to read the content it receives, and how to present the content it produces with AI. It is one of the subjects we cover in our support for communications.

This article is not legal advice.

By Sofia Mestre

Sources

Back to top