OpenAI is adding text watermarking in ChatGPT and Codex

Starting over the coming weeks, eligible ChatGPT and Codex users in the European Union will get text output carrying an invisible, machine-readable watermark called textGrain. OpenAI says it is not making this a global default yet. API customers anywhere can opt in, and a detector is opening only to approved researchers and expert organizations.

The trigger is the EU AI Act, which requires generative AI providers to make generated text identifiable by machines. Anthropic made a similar move in August, built on Google DeepMind's SynthID for text, and The Verge notes that not everyone welcomed it. So this is no longer one company's experiment. It is becoming the regulatory floor for anyone selling chatbots in Europe.

We think provenance is worth building, and we give OpenAI credit for how plainly it describes the limits. The company says outright that textGrain does not guarantee reliable detection, and that a watermark does not prove accuracy, ownership, how much a human contributed, or human authorship. That is the right set of warnings. The problem is that warnings live in a blog post, and detection results live in a principal's office, an HR file or a newsroom.

Here is the harm we worry about if this scales carelessly. A teacher, an editor or an employer hears that AI text is now watermarked and starts treating a detector hit as a verdict. OpenAI is keeping its own detector away from the public for exactly that reason, citing the risk of false positives and missed watermarks. But other tools, and other vendors' claims, will not all be so careful. The likelier near-term outcome is a patchwork: some text marked, some not, some edited until the signal fades, and a lot of people drawing confident conclusions from a weak signal.

There is also a two-tier world forming. Europeans get watermarked output by default; most everyone else does not unless a developer chooses it. That makes sense as a regulatory compliance move, and OpenAI frames it as room to learn from real use. Still, it means the same paragraph from the same model carries a hidden label in Paris and none in Phoenix. If the goal is a healthier information ecosystem, regional defaults are a starting point, not an answer.

Our read: useful, overdue and honestly labeled, but small. Watermarking can help platforms and researchers trace patterns of synthetic text at scale. It cannot settle whether any single person cheated, lied or deserves blame, and nobody should let it pretend to.

OpenAI already attaches Content Credentials and invisible watermarks to supported images and audio and offers public verification for those formats. Text is harder because it is easily rewritten, translated or trimmed, which is why the company is treating this as a phased rollout and says it will revisit the approach as standards and evidence evolve.