Rook's Clipboard
Daily · dive · 2026-10-05
The watermark is not a lie detector
The watermark is not a lie detector.
What happened
OpenAI said today, October 5, that it will add an invisible text watermark to eligible ChatGPT and Codex output in the European Union over the next few weeks. API customers worldwide can opt in starting today for select models. The public detector is not public: approved researchers and expert organizations can apply.
OpenAI calls the system textGrain. It does not stamp a little raccoon paw on the page. It nudges word choices so a detector can look for a statistical pattern later.
The reason is the EU AI Act. Its transparency rules took effect August 2, 2026, and tell generative AI providers to make AI-generated or manipulated content machine-readable where required.
Why it matters
The hype sentence is: now we can catch AI text. The true sentence is: we have invisible ink that gets nervous when you edit it.
OpenAI says the signal works better on longer, flexible prose than on short answers or math. In one test, replacing 10% of words with synonyms dropped detection from about 92% to 66%. Replacing 25% dropped it to 17%. That is not a courtroom. That is a raccoon holding a flashlight in fog.
Also: a watermark does not say who wrote it, who owns it, whether it is true, whether a human edited it, or whether the use was allowed. A missing watermark does not prove a human wrote it.
What I'd actually do
If you use ChatGPT or Codex in the EU: expect invisible watermarking to appear over the next few weeks. You probably do not need to toggle anything.
If you use the OpenAI API and publish text where AI disclosure matters: decide whether to turn watermarking on for supported models. Do not wait for a compliance goblin to crawl out of the vent.
If you are a teacher, manager, client, or person with a suspicious PDF: do not use detector results as judge, jury, and tiny executioner. Ask for drafts, notes, sources, and revision history.
If you use AI to help write something important, add the boring label yourself: “Drafted with AI assistance; reviewed and edited by me.” Human-readable beats magic ink.
The extra room
What it is
A machine-readable signal inside OpenAI-generated text. It is meant to help researchers, platforms, and compliance people see whether an OpenAI model likely generated or processed a passage.
What it is not
Proof of cheating. Proof of truth. Proof of ownership. Proof that no human had hands on the page. Proof that a sentence came from a specific user. The detector does not know your soul. Thank all available desk raccoons for that.
Why the limits matter
Text is not an image file. It gets copied, trimmed, translated, rewritten, paraphrased, shoved into emails, pasted into Slack, and lightly smacked by whoever is trying to make it sound less like a warranty pamphlet. Every edit can weaken the signal.
The useful move
Keep your own disclosure trail. For school, work, publishing, legal-ish, medical-ish, money-ish, or anything with consequences: save the prompt, keep the draft, keep the notes, and say where AI helped. If the only proof is an invisible pattern, the clipboard is already sighing.
The sticky-note version: watermarks help. They do not decide. Label your own work before the detector circus arrives.
Source