OpenAI reveals new way to detect ChatGPT text without visible labels
&imwidth=600&imheight=450&format=webp&quality=medium "OpenAI reveals new way to detect ChatGPT text without visible labels")
OpenAI reveals new way to detect ChatGPT text without visible labels Photograph: (X)
OpenAI will introduce an invisible watermark for ChatGPT and Codex text in the EU to meet AI Act transparency rules. The hidden pattern can be detected by specialised systems, but OpenAI says editing, short text and translation can make detection less reliable.
OpenAI is preparing to add an invisible watermark to text generated by ChatGPT and Codex in the European Union, bringing a major change to how AI-written content can be identified.
The watermark is being introduced to comply with the transparency requirements of the EU AI Act. Unlike a visible label, it will not appear as a symbol or warning to readers. Instead, OpenAI will subtly influence word choices in generated text, creating a hidden pattern that specialised detection systems can identify. The company said the feature will begin rolling out to eligible ChatGPT and Codex users on all plans in the EU over the coming weeks.
How does OpenAI's invisible watermark work?
OpenAI's system, called textGrain, works inside the generated text itself. In simple terms, the model makes small, controlled changes to its word-selection process. When enough of these changes are combined, a detector with access to the relevant key can look at the text and identify the hidden pattern. Because the watermark is embedded in the words, it can remain attached when users copy and paste the text elsewhere. OpenAI said the watermark does not identify the person who generated the content. It also reported no meaningful impact on model performance when the system was enabled.
Can the ChatGPT watermark be removed?
This is where the technology has an important limitation. OpenAI's own tests found that editing can make the watermark harder to detect. Replacing just 10 per cent of words with synonyms reduced detection from about 92 per cent to 66 per cent in one test. Short passages, mathematical answers and translated text can also be more difficult to identify. That means a failure to detect the watermark does not prove that a person wrote the text. OpenAI will initially give detector access only to approved researchers and expert organisations to assess its reliability and responsible uses.
Trending Stories
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
Why is OpenAI doing this now?
The change comes as AI-generated content becomes harder to distinguish from human writing and regulators push technology companies towards greater transparency. The EU's AI Act transparency rules took effect on August 2 and require AI-generated content to be marked in a way that other systems can identify.
OpenAI said the watermark will initially be limited to the EU rather than becoming a global default. Developers using OpenAI's API can also enable the technology for selected models worldwide, although it is switched off by default. The move follows Anthropic's decision to watermark Claude-generated text worldwide. OpenAI had reportedly developed a similar system previously but delayed releasing it amid concerns that users might move to competing AI services without watermarks.
