Passing off ChatGPT’s work as your own is about to get tougher—at least if you live in the European Union. On Monday, OpenAI announced that it would start inserting an invisible watermark, imperceptible to the naked eye but identifiable by machine, into all text outputs generated by its AI models, including ChatGPT and its coding agent Codex.
The watermark will move OpenAI into compliance with the EU AI Act, which requires model-made text to include a machine-readable identifier that flags the content as AI-generated. Per the company, the watermark will be an opt-in feature at first, and users globally can choose to flip that switch starting today. In the coming weeks, the watermark will be added to “eligible ChatGPT and Codex text output in the European Union,” which will not be optional—it’ll appear automatically, with no off switch.
So who’s going to be able to see the watermark? According to OpenAI, researchers and expert organizations will be provided with access to a detector that can spot the invisible indicator. OpenAI opted not to make the identifying tool available to the public at launch due to the risk of false positives. Instead, it said that access will be granted “on a case-by-case basis” to support evaluation.
OpenAI is just the latest company to go the route of watermarking. Anthropic launched a similar, machine-readable marker back in August (which led to people immediately seeking workarounds and ways to scrub their text of the AI indicator). And Google DeepMind has been putting a watermark in its outputs for a couple of years now. (Strangely, Google has gone backwards on this, announcing in August that users could remove watermarks from images, videos, and songs in countries where it’s not required).
Notably, those other companies’ watermarks apply (or applied) to users in the United States. OpenAI’s does not, despite growing concern in the country over the proliferation of AI-generated content. A recent poll from the Rainey Center found that nearly eight in 10 people support AI watermarking requirements, including 82% of Democrats and 75% of Republicans.
But public support does not make policy, and the United States still lags in federal regulations related to AI. With no requirement in place for frontier AI companies to include a watermark that helps users and platforms quickly identify AI-generated content, there is really no motivation for OpenAI to proactively comply with a law that doesn’t exist. Turns out regulations do work if you just pass and enforce them.
