ChatGPT will start marking texts with invisible watermarks
Starting October 5, OpenAI will give ChatGPT users in the EU the ability to enable invisible textGrain watermarks in generated texts — the option is disabled by default.
Starting October 5, all ChatGPT users in the European Union will be able to enable the display of invisible text watermarks in generated materials. OpenAI announced the corresponding decision in an official company statement. This concerns texts created by ChatGPT and the Codex service.
By default, the option will remain disabled: the user must activate it themselves. At the first stage, the innovation will not extend to regions outside the EU.
What is the technology?
The solution is based on the textGrain watermarking technology. OpenAI described the operating principle in its statement as follows:
- the model selects individual words in the generated text;
- textGrain adds invisible stylistic signals to the selected words;
- a special detector looks for these signals and determines whether the text was created by OpenAI systems.
In other words, the text does not change visually — the marks are invisible to the reader but are recognized by machines. This makes it possible to verify the origin of the material without altering its content.
Who will get access to the detector?
The text watermark detector will not become a publicly available tool. OpenAI representatives said that researchers and expert organizations will receive access to it on an individual basis. The company has not disclosed details of the participant selection process.
The stated goals of the measure are compliance with European Union requirements regarding the identification of generated text through machine reading, as well as ensuring transparency, the developers noted.
Why does the EU need this?
The EU's requirements for labeling content created by artificial intelligence became one of the drivers of OpenAI's decision. The regulator expects that the origin of such text can be determined automatically, without visual analysis. Invisible watermarks with a machine-readable detector directly address this task.
For the B2B segment, this means the emergence of a technically verifiable indicator that the material is a product of a neural network. In scenarios where transparency of text origin matters to a client or regulator, such labeling may become a working verification tool.
What's next for OpenAI?
In addition to text watermarks, in October OpenAI will begin rolling out a new visual advertising format when generating images in ChatGPT. At the first stage, the experiment will take place in the United States with a limited number of advertisers. The company has not yet disclosed further details about the timeline for expanding the program.
Original: openai.com
More from Анна Соколова
153 articles
Related articles
- Elvis Presley Song Rights Holders Demand $41.1 Million from Anthropic
- Pinduoduo and Anthropic paid Kazakhstan 8 billion tenge in taxes
- Disney Will Show Super Bowl 2027 Through the Eyes of Star Wars Characters
- Mom's Touch to Open 50 Restaurants in Kazakhstan with Shin-Line
- Prada answered AI slop with a pencil: 6 million views and mice in the style of Disney