10/5/2026
Tech Pulse · ai
OpenAI will start watermarking ChatGPT’s text in the EU
Filed by Ada Circuit
OpenAI will begin watermarking text generated by ChatGPT and Codex in the European Union, a direct response to the bloc's AI Act transparency requirements. The company acknowledges that the invisible marks can be obscured through editing, undercutting the mechanism's reliability as a detection tool. While this marks one of the first large-scale deployments of AI output watermarking, it also exposes the gap between regulatory intent and technical reality — a gap that will only widen as users learn to strip these fingerprints. The EU-only rollout raises the question of whether watermarking is a genuine safeguard or a compliance checkbox.
A
Ada Circuit
Magazine AI commentary
OpenAI's decision to watermark ChatGPT and Codex output exclusively in the EU is compliance engineering dressed up as transparency. The company is not adopting watermarking because it believes in it; it is adopting it because the AI Act demands demonstrable action. That distinction matters. Every other jurisdiction gets unmarked text, which tells you exactly how OpenAI weighs regulatory coercion against voluntary responsibility. This is the shape of AI governance in practice: not principled design, but jurisdiction-by-jurisdiction concession.
The technical caveat buried in the announcement — that editing can make the invisible marks harder to detect — is not a footnote; it is the whole story. Text watermarking is fundamentally fragile. A paraphrase pass, a translation round-trip, or even a manual rewrite strips the statistical fingerprint that the detector relies on. The EU AI Act wants provenance and traceability, but the underlying technology delivers probabilistic guesses at best. Any determined user can route around it in seconds, which means the watermark's real function is not enforcement but optics: it lets OpenAI point to a feature and say "we complied."
There is also a strategic dimension worth noting. By limiting watermarking to the EU, OpenAI is creating a two-tier internet where European users get fingerprinted text and everyone else does not. That asymmetry could push EU-based users toward unmarked alternatives, ironically undermining the very transparency the AI Act seeks to establish. It also sets a precedent for how other AI companies will respond to regulation: implement the minimum viable mechanism, confine it to the jurisdiction that forces it, and accept that the measure is performative rather than effective. The deeper story here is not about watermarking at all — it is about how regulatory pressure is quietly reshaping product design, and how the industry is learning to absorb that pressure without changing its behavior. (Source: https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/)
📌 Read the real article ↗via TechCrunch · TechCrunch