Skip to content
by InfinyAI ES

IA en un minutoNewsSpain and Europe

Spain and EuropeUpdated:

OpenAI will add invisible watermarks to ChatGPT and Codex text in the EU to comply with the AI Act

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

To meet the AI Act, OpenAI will put marks people cannot see into text that ChatGPT and Codex produce for users in the European Union, starting in the coming weeks. On the API, any customer in the world can already switch them on for some models, though they come turned off. Only approved researchers and expert organizations will get the detector, which in OpenAI tests fell from 92% to 66% when 10% of words were swapped.

OpenAI is going to place an invisible mark on the text ChatGPT generates in the European Union.

The company explained today how it plans to abide by the AI Act, the EU's law on AI, under which AI-generated text has to be identifiable by machines. Over the coming weeks, invisible watermarks will be added to text produced by ChatGPT and by Codex, the company's coding assistant, for eligible users on every plan, though only in the European Union. API customers anywhere in the world can already turn the watermarks on for certain models, but the default setting is off.

For the time being, the detector that reads these watermarks will be restricted to researchers and expert organizations that have been approved. OpenAI concedes the technology is at an early stage and has major limitations. In its tests, swapping 10% of the words for synonyms was enough to cut detection to 66%, down from about 92%.

The technology, called textGrain, works by slipping an invisible statistical signal into the words the model chooses, which a detector later searches for. OpenAI says textGrain equaled or beat other methods in its tests, SynthID for text among them, but short texts are harder to catch. With the false positive rate set at 1%, the watermark was found in about 80% of 200-token passages about psychology and in about 95% of those 400 tokens long, and far less often in mathematics. When 25% of the words are replaced, detection falls to 17%.

According to OpenAI, the watermark does not identify the user, does not measure how much of a text a person contributed and does not say whether the text is true, and its absence is no proof that a human was the author. In the coming weeks the company will also work with cloud partners so that watermarking is available for OpenAI models reached through their services, and it plans to publish the technology as open source.

Why it matters · analysis and opinion

The European law places the obligation on the provider, not on whoever uses the text, and OpenAI has chosen to comply only where it is required to: in the European Union and in ChatGPT and Codex, while in the API the decision is left to each customer. What remains to be settled is almost the most important part, who will be able to check the watermarks. The detector stays for now with approved researchers, so neither the public nor businesses will be able to use it to tell whether a text came from ChatGPT. For people who write with AI the change is invisible; for those who want to check, there is no tool yet.

Official source: OpenAI · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.