OpenAI is forcing invisible 'tattoos' on your ChatGPT text in the EU
It seems OpenAI finally decided to give in to EU bureaucrats. By injecting invisible statistical 'watermarks' into ChatGPT and Codex, the company is attempting to prove that machines are writing our emails and essays.
The implementation uses a technology called textGrain, which subtly shifts the statistical probability of word choices during generation. Instead of adding obvious characters or weird symbols, the system nudges the model to pick specific synonyms based on a secret pattern. This means the output is statistically distinct without looking different to the human eye.
For now, this feature is restricted to users within the EU, though OpenAI API customers worldwide can toggle it for specific models. The company admits this is hardly foolproof; their internal tests show that identifying AI-written content is a game of chance. While they can catch about 80% of 200-token chunks, changing just 10% of the words causes the detection probability to tank significantly.
This initiative follows similar moves by Anthropic with their Claude models and Google with Gemini. The goal is to comply with strict EU AI laws that demand machine-readable identification for synthetic content. Even with this tech, the company warns that a positive detection isn't proof of total AI automation—it just confirms the model had a hand in the process.
Ultimately, this is a cat-and-mouse game where the mouse is already halfway out the door. Marking AI text is about as effective as putting a padlock on a screen door; it satisfies the regulators' need for 'provenance' while barely slowing down anyone determined to rewrite a few sentences. It’s a performative gesture for the lawmakers, not a real solution to the truth-gap.
Source: OpenAI
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.