Aleksei Grebenkin

Both major labs switched on text watermarking within 48 hours at the turn of August. My first thought was that this is very hard to do in a way that actually holds text has no pixel layer to hide anything in. So I read Anthropic's own documentation on how it works.

It does work. It also breaks in a lot of ordinary situations, and they say so themselves: short text leaves too little signal, heavy editing or translation thins it out, screenshots and format conversion strip it entirely.

I build a product mostly on my own, with agents doing a growing share of the mechanical work. This year my output went up and my days got worse, and it took me a while to see why.

Automation ate the typing. It did not touch the checking, and it multiplied the number of things waiting to be checked.

Then I found telemetry that matches: Faros AI looked at two years of workflow data from 22,000 developers across 4,000+ teams. Task throughput per developer up 33.7%. Median time a pull request sits in review up 441.5%. Daily PR contexts per developer up 67.4%.

Both big AI labs quietly switched on watermarking this month Anthropic marks every text Claude touches, OpenAI marks GPT-Live voice audio (EU AI Act, Article 50).

The catch: a detected mark means the model processed the text, not that it wrote it. Fixing grammar leaves the same mark as full generation. And no mark proves nothing either edits, translation or transcription wash it out.