OpenAI is preparing to mark ChatGPT and Codex text with an invisible signal across the European Union. The company announced its plans on 5 October 2026, following new European transparency rules for generative AI. British users will not be included in the initial regional rollout.
In its official announcement, OpenAI said it would โadd an invisible watermark to eligible ChatGPT and Codex text outputโ.
Developers using OpenAI’s API can already activate the feature for selected models worldwide. However, watermarking remains switched off by default for those API customers.
How the invisible watermark works
The technology is called textGrain, and it changes how a model selects language. Those choices leave a statistical pattern that specialised software can identify later.
Readers will not see a logo or notice any extra characters. Instead, approved researchers will use a dedicated AI detector to search for OpenAI’s signal.
The company is accepting applications for access to that detection tool. Ordinary users cannot access the text detector at launch because its results require careful interpretation. OpenAI also says the system does not identify people or expose their private conversations. The tool will report only if an OpenAI watermark was detected.
Why Europe is pushing for greater transparency
The European Union’s AI Act introduced transparency duties on 2 August 2026. Under Article 50, providers must add machine-readable marks to supported AI-generated material. Some existing systems have until 2 December 2026 to meet that marking requirement.
These rules address concerns about misleading posts and synthetic content shared without clear disclosure. For newsrooms and businesses, the question is increasingly about proving where published material originated.
The European Commission has issued guidance explaining how the transparency duties apply. Its published statement describes the purpose as โreducing the risk of deception and manipulationโ.
Tests show important limits
OpenAI’s own figures show why detection results need caution. In one evaluation involving psychology answers, its detector identified around 80% of watermarked 200-token passages.
For longer passages containing 400 tokens, detection climbed to approximately 95%. Both figures used a target false-positive rate of 1%.
Editing made the results considerably weaker in another test. Replacing 10% of the text with synonyms reduced detection from approximately 92% to 66%.
When researchers replaced 25%, detection fell further to 17%. These results underline why a missing watermark cannot prove human authorship.
A positive result cannot reveal who wrote the surrounding material either. It also cannot establish factual accuracy or determine who owns the text.
What happens to AI-generated images?
OpenAI already supports provenance checks for selected images and audio files. Its public verification service examines Content Credentials and embedded SynthID watermarks.
Someone searching for an AI watermark remover may find tools designed for pictures. Those products should not be confused with methods for detecting patterns in written text.
Image changes can damage provenance information – including metadata saved inside the file. Even embedded marks are not guaranteed to survive every modification.
For publishers, there is a practical lesson about verification. Keep original files and record their sources before distributing material online.
OpenAI says it plans to study detection reliability before expanding access further. The announcement marks a regulatory step but it does not settle the wider question of trust. News organisations should still verify claims and document editorial changes before publishing AI-assisted material.
Leave a Comment