Deep dive into AI text-watermarking and what EU's AI Act actually mandates about AI detectability.

How it works: The SynthID paper from Demis and team in Nature is probably the best primer in the technology. Generally, most approaches involve sampling next tokens statistically differently with some random key while not distorting the semantics of the output.

Detecting text watermarks can never be perfect. Longer passages are always easier to detect, only at ~400 tokens do you get a true positive of rate of ~90% when your false positive rate is <1%. Because small changes in the text can change whether the algorithm thinks it's AI or not, other approaches include semantic data in their watermarking scheme too.

Industry status on the EU regulation: EU's AI Act Article 50(2) requires AI generations from all modalities to be AI-detectable, with practical implementation details still being finalized. The main exception is "AI systems perform an assistive function for standard editing or do not substantially alter the input data" although how they would distinguish between these cases is unclear. Article 113 says "It shall apply from 2 August 2026."

Claude declared that they will comply. Gemini already uses SynthID for their text outputs (even though no production public detector exists for text outputs). OpenAI had developer watermarking a while ago but reportedly shelved it when 30% of their users said they'd use the product less. However, their support page says "our goal is to expand provenance signals to all modalities including text" (!)

I've always maintained I think it's critical to know whether text comes from AI or a human (and thus support tech like Pangram). Judging by the comments on X, I might be in the minority. This comment sort of embodies the backlash: "I'm paying it to write for me in a way people and machines wouldn't detect it is AI written" In other words, users believe if their ideas go into using AI for expressing them, *they* should be given credit for the idea and not penalized for using AI. I'd counterclaim that penalization and provenance are distinct, and provenance is extremely critical because a) without provenance, you are more likely to be judged poorly and falsely accused of using AI when you didn't b) just like with food, fundamentally readers deserve the right to know where the text came from and reserve their own judgment on whether its still valuable to them or not!