OpenAI Watermarks ChatGPT Text by Default in the EU — But It's Easy to Remove

OpenAI is turning on text watermarking by default for ChatGPT and Codex outputs in the EU — a compliance move under the AI Act — but the company itself acknowledges the system is not especially reliable and is easy to circumvent.\n\n## Key takeaways\n\n- OpenAI is enabling invisible text watermarks by default for EU ChatGPT users, a requirement tied to the EU AI Act's transparency rules.\n- The watermarking technology, called textGrain, embeds signals in generated text that are invisible to readers but theoretically detectable by a verification tool.\n- OpenAI and Ars Technica both note the system is not especially reliable and can be defeated by minor edits — copy-paste, paraphrasing, or reformatting.\n- The default-on rollout applies only in the EU; users in the US and elsewhere are not affected.\n- No equivalent mandatory watermarking is in place for ChatGPT's image outputs under this rollout.\n\n## What textGrain actually does — and doesn't do\n\nOpenAI's watermarking system, called textGrain, encodes invisible signals into the statistical patterns of generated text — word choices, punctuation, subtle phrasing shifts — rather than appending any visible tag. The idea is that a detector tool can scan a piece of text and flag it as likely AI-generated, even when it reads identically to unwatermarked output.\n\nThe catch is substantial. According to Ars Technica, the system is acknowledged to be not especially reliable, and circumventing it requires nothing more sophisticated than lightly paraphrasing the output, reformatting it, or running it through a second tool. That's a low bar for anyone motivated to strip the signal — which includes bad actors using AI to generate misleading content, the exact use case regulators are most concerned about.\n\nFor AI-art creators and writers using ChatGPT as part of a creative workflow — drafting captions, writing character backstories, generating prompt variations — the practical impact is close to zero day-to-day. The watermark is invisible and doesn't alter the text's usability. But it does mean EU-based users are now generating text that carries a hidden provenance signal, whether or not they know it or want it.\n\n## EU-only, by design\n\nThe default-on status is explicitly a response to the EU AI Act, which requires providers of general-purpose AI systems to mark AI-generated content in ways that allow detection. OpenAI is complying in the jurisdiction that mandates it and not extending the requirement elsewhere. US users see no change.\n\nThat geographic split matters for creators working across borders. If you're an EU-based creator submitting AI-assisted text to a platform, publication, or client that runs AI-detection checks, your ChatGPT outputs now carry a signal — however weak — that a detector might pick up. Outside the EU, the same output carries none.\n\nThe broader regulatory picture remains fractured. As Charmloop has covered, meaningful AI content regulation in the US remains gridlocked in Congress, leaving the EU as the de facto rulemaker for watermarking standards right now.\n\n## The reliability problem no one has solved\n\nText watermarking is genuinely hard. Unlike image watermarking — where metadata or pixel-level signals can survive many transformations — text watermarks degrade the moment the output is touched. A single round of light editing, translation, or summarization can strip the signal entirely. OpenAI is not claiming otherwise.\n\nThis puts textGrain in an awkward position: it satisfies a regulatory checkbox while providing limited real-world detection capability. Researchers studying AI writing tells have found that statistical fingerprints — like the word "dependable" appearing 23 times more often in Anthropic's Opus 5.5 than in human writing — can be more durable than engineered watermarks, precisely because they're harder to deliberately remove.\n\nFor creators who use ChatGPT to assist with writing alongside image generation in tools like Charmloop's generator, the takeaway is practical: don't treat the presence or absence of a watermark as a reliable signal of anything. The technology isn't there yet, and OpenAI isn't claiming it is. What's changed is the default setting — not the underlying capability.