SDSignal Desk

ChatGPT and Codex text gets an invisible EU watermark as the AI Act’s transparency rules bite

· Updated · Signal Desk newsroom

Analysis of reporting by TechCrunch. How we work

Image: TechCrunch

TechCrunch reports eligible EU ChatGPT and Codex users on all plans will get an invisible textGrain watermark in coming weeks to meet AI Act rules in force since August 2. Developers worldwide can opt in on select API models. OpenAI says a missing watermark does not prove human authorship.

Key takeaways

  • Eligible EU ChatGPT and Codex users on all plans will receive an invisible text watermark.
  • The EU AI Act’s transparency rules have required machine-detectable marking since August 2.
  • API developers anywhere can opt in on select models; it is off by default outside the EU.
  • Swapping 10% of words for synonyms cut detection from about 92% to 66% in OpenAI’s tests.
  • Detector access is initially limited to approved researchers and expert organizations.

What is changing

Over the coming weeks, eligible ChatGPT and Codex users on all plans in the European Union will get an invisible watermark on generated text, TechCrunch reports. The EU AI Act’s transparency rules, in effect since August 2, require AI companies to mark generated content in a machine-detectable way. Developers anywhere can turn the watermark on for select API models now; it is off by default, and OpenAI is not making it a global default. A companion look at the robustness data is in our earlier piece (EU-default watermark), alongside OpenAI’s own announcement (watermarking in ChatGPT and Codex).

The method, called textGrain and described in a technical report co-written with researchers from the University of Pennsylvania and Yale, uses a secret key to subtly bias next-word choices; across hundreds of such nudges, a detector with the key can identify the pattern from text alone. Anthropic, Google, Meta, Microsoft and OpenAI are among companies committed to the EU’s code of practice on AI-generated content.

Benefits and stated limits

Provenance tools help people judge what they read, and a mark embedded in word choices survives copy and paste better than metadata. OpenAI says the watermark does not identify the user and that it saw no meaningful change in model performance with it on. If both hold, the cost to ordinary users is low.

OpenAI is candid about limits. Its tests show that swapping 10% of words for synonyms drops detection from about 92% to 66%. Short passages, math answers and translated text are harder to catch. The company says a missing watermark does not prove a human wrote something. A detector with a meaningful miss rate and an evasion path works best against people who are not trying to hide, such as a student who used ChatGPT openly for a draft, and worst against those who are. Schools or employers treating a positive as proof of misconduct, or a negative as proof of innocence, will make both kinds of error, which is why limiting detector access to approved researchers and expert organizations at first is a cautious choice.

Authorship and competition

OpenAI notes a watermark can show that an OpenAI system generated or processed part of a passage, but not how much human judgment went into it. Anthropic’s plan to watermark Claude text worldwide drew pushback from users who said they supplied the instructions, context and decisions while the model was the tool, an argument likely to grow as more professional writing passes through AI.

The Wall Street Journal reported in 2024 that OpenAI held back an earlier watermark partly over fears users would switch to rivals that did not mark text. Regulation resolved that collective-action problem in Europe; outside it, the off-by-default setting suggests the commercial concern remains. More in our regulation coverage, OpenAI coverage and generative AI coverage.

Who feels it

EU users
Output carries an invisible mark OpenAI says does not identify them.
Educators and employers
Detection is a signal, not proof, given reported miss rates.
Developers
Can opt in on select API models worldwide.
Rival providers
The same EU obligation removes room to compete on not watermarking in Europe.

What to watch

  1. Whether detector access expands.
  2. Whether OpenAI makes watermarking default outside the EU, as Anthropic says it is doing.
  3. How EU regulators judge robustness given the synonym-swap results.

FAQ

Who gets the watermark?

Eligible ChatGPT and Codex users on all plans in the EU, over the coming weeks.

Why now?

The EU AI Act’s transparency rules, in effect since August 2, require machine-detectable marking.

Can it be evaded?

Partly; OpenAI’s tests show 10% synonym swaps cut detection from about 92% to 66%.

Does a missing watermark prove human authorship?

No, OpenAI says it does not.

Does it identify users?

OpenAI says it does not.

Read the original

Continue at the source.

TechCrunch

Companies: OpenAI

More on OpenAI

All OpenAI coverage →