Link: https://gizmodo.com/openai-is-adding-text-watermarks-in-the-eu-because-regulation-works-2000821852

Huh, the government seems to have tools to make companies comply. Strange.

An algorithm that prefers certain words over others, for specific inputs. The same way we encrypt data really.

Based on “controlling” of a sampling set for a token prediction . Nice summary is here https://sebastianraschka.com/blog/2026/claude-text-watermarking.html eg

A pattern of invisible characters perhaps.

This is also a really good explainer: https://declaude.org/watermarking/

As the others explained, it's mainly statistics regarding usage of specific words. This also means anyone can make a tool that uses a thesaurus to mess with the watermark (these are already available online - they have the disclaimer that they can't guarantee that it works every time tho), and short texts might not provide enough data for the watermark to be reliably identified at all.

I would love to see regulations work, but I'm convinced it's watermarked so that OpenAI can filter out content generated with ChatGPT from its own training data^1

If someone's interested in more details: https://youtu.be/BFksx2M93sw

Theres also the significant issue that it just limits access to unwatermarked image generation, not gets rid of it entirely. So state actors and sufficiently capable entities can still make disinformation and whatnot

Still very helpful harm reduction though

this doesn't work remotely as well as the writer of this article seems to think. furthermore, it doesn't apply to all AI corporations. nor could it be enforced against open weight models, since every single one of those models is available in abliterated versions created by end users that remove guardrails and censorship put in place by the creators. these abliterated models are capable and willing to help users with things governments would define as illegal.

The access to which will only be given to "good guys".

Its a not a product meant for end-users to filter AI generated text, rather its for the corpos to check if an end-user used AI for something.