OpenAI will introduce an invisible watermark to qualifying ChatGPT and Codex texts inside the European Union complete the coming weeks. Their evaluation indicates that replacing certain words alongside synonyms can weaken the watermark’s effectiveness. In one test, substituting 25% of the words in 400-token English passages alongside synonyms lowered finding rates from about 92% to 17%.
From today, API users worldwide can opt to activate watermarking for choose models, although it stays off by default. The content finding tool is not publically accessible at launch; it is currently constricted to approved researchers and expert organizations.
What’s Rolling Out Where
OpenAI states that the ChatGPT and Codex watermark volition attain eligible users throughout all plans inside the EU only. Initially, content watermarking won’t be a default characteristic globally.
The Oct. 5 update does not define which models are part of the API opt-in or whether EU ChatGPT users can disable the watermark. The business says it is operating alongside haze partners to add watermarking to outputs from its models via their services in the coming weeks.
Why The EU, And Why Now
The transparency rules under Article 50 of the EU AI Act began applying on Aug. 2, 2026. The European Commission states that AI systems placed on the market before this date have until December 2 to encounter the marking and finding obligation. The voluntary Code of Practice on Transparency of AI-generated Content provides organizations alongside an optional way to display their compliance. By the end of July, concerning 190 organizations had signed, and OpenAI publicly supported the Code in June.
The business says the EU-only rollout gives it area to study from real-world use and feedback. I covered its decision against content watermarking for ChatGPT in August 2024, following a business study established nearly 30% of ChatGPT users stated they would use the assistance small if watermarking was implemented.
What OpenAI’s Tests Show
OpenAI’s method, called textGrain, adds a statistical indication to the model’s term choices that a detector can appearance for. At a mark false affirmative charge of 1%, the detector established the watermark in concerning 80% of 200-token passages and concerning 95% of 400-token passages, for satisfied specified as psychology. The business reported much lesser rates for satisfied specified as math, anywhere term choice is small flexible.
I noted in August that the Code doesn’t necessitate watermarking of free-form content shorter than 200 tokens. Two hundred tokens is the shortest dimension in the post’s results.
Editing the content weakened the indication in a distinct circumstance alongside 400-token English passages. Replacing 10% of words alongside synonyms reduced finding from about 92% to 66%. Increasing the replacement to 25% additional decreased finding to 17%. Both tests used responses to questions from the ELI5 dataset.
The business reports that textGrain matched or exceeded the achievement of another methods it tested, including SynthID for text. However, they emphasize that successes under ideal conditions do not justify reliability in mundane situations. In their benchmarks, they saw minimal to no difference whether watermarking was enabled or disabled.
Who Can Check A Watermark
Researchers and expert organizations can request access to OpenAI’s content detector, initially on a case-by-case basis according to the Code. The tool reports whether it detects an OpenAI watermark but does not disclose person identities or display their prompts and chats. OpenAI notes that because of the hazard of missed watermarks and possible false positives, it is not accessible to the community at launch. In contrast, depiction and audio verification tools via openai.com/verify and the Content Provenance API remain open to all.
In a 2024 update, the business wrote that equal alongside a low false affirmative rate, “applying it to ample volumes of content would guide to a ample figure of total false positives.”
According to the post, a watermark doesn’t measure how much a individual contributed, and a missing one doesn’t demonstrate a individual wrote the text.
How It Compares With Claude
Anthropic marks content from supported Claude models worldwide, according to its support article. Its explainer says that’s since it doesn’t yet have a durable way to range watermarking by region.
OpenAI vs. Anthropic: How Their Text Watermarking Compares
Different rollout scopes, methods, and admission approaches
| Apps | OpenAIEligible ChatGPT and Codex text, EU only, complete the coming weeks | AnthropicSupported Claude models, worldwide |
|---|---|---|
| API | OpenAIOpt-in for choose models, off by default, models not named in the post | AnthropicMarked on supported models, which the assistance part lists by name |
| Method | OpenAItextGrain, OpenAI’s own | AnthropicA type of Google DeepMind’s SynthID-Text |
| Detector access | OpenAIApproved researchers and expert organizations, by application | AnthropicPrivate preview for eligible organizations, including regulators, media, and researchers, affirmative enterprises verifying their own compliance |
| Published editing results | OpenAIDetection rates following 10% and 25% synonym swaps | AnthropicNo figures in the explainer. Light editing likely won’t completely eliminate it, and a complete rewrite will |
Why This Matters
An agency alongside writers in the two Berlin and Toronto, all sharing the identical ChatGPT plan, power shortly discover that several of their copy has OpenAI’s watermark from one agency but not the other. A agreement clause or AI guideline that considers a finding outcome as evidence of who authored the copy is basically asking the watermark a inquiry that OpenAI has stated it can’t answer.
Looking Ahead
OpenAI says it is preparedness to develop detector admission whenever it feels “results can be interpreted responsibly.” As watermarking starts in the EU complete the next few weeks, teams operating alongside EU personnel on ChatGPT power commencement creating marked satisfied that lone approved researchers and expert organizations volition have the capability to review.
Featured Image: daily_creativity/Shutterstock
