OpenAI said Monday it’s going to start hiding a watermark in ChatGPT’s writing. In Europe, anyway.
It’s not a little logo, and it’s not hidden characters. It’s a pattern in the word choices themselves. The model gets nudged toward certain words using a secret key, hundreds of tiny nudges add up across a passage, and a detector holding that key can spot it later. It’s rolling out over the coming weeks to ChatGPT and Codex users in the EU. Everywhere else, developers using OpenAI’s API can switch it on, but it’s off by default. And for now the detector only goes to approved researchers and expert organizations. The reason is the EU AI Act, whose transparency rules kicked in August 2 and require AI-generated text to be marked so machines can identify it. Anthropic did the same thing with Claude two months ago, worldwide.
So the machine is signing its work now. Surprise.
Here’s the part I didn’t expect. OpenAI was pretty honest about the limits. In its own test, swapping 10% of the words for synonyms dropped detection from about 92% to 66%. Short passages and translated text are harder to catch. A missing watermark, the company said, “does not prove human authorship.” And then this line, which I’ve read about five times now: a watermark “can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.”
Read that again. The people who built the thing are telling you it answers one question: was our model in the room? It doesn’t answer the question anybody in publishing actually fights about. Who did the work? Who’s standing behind it?
And look at what happens when a human gets involved. The more you edit, the fainter the mark gets. Rewrite enough and it’s basically gone. Think about that for a second. In watermark world, the human hand is literally what erases the machine’s signature. That’s not a bug for writers. That’s a description of revision.
Meanwhile, over in publishing, detectors that give out percentage scores keep showing up in book-deal fights, and the people getting hurt are often writers who never got a say in having their manuscript scanned. At least a watermark is honest about what it is. It’s still not a judge.
So here’s the little world.
The above was all AI draft. It checked out pretty good and I left it alone, finding it fairly interesting. The rest of this is mostly mine, as it required a lot of editing. My bot helper, I realized, has grown a bit careless in expressing my views, and so due diligence requires me to re-write rthis and explain to my Grok bot the problems I find with his wording. GAPOZ Press and the associated functions we perform, such as the Circle, are not nearly as invested in AI hair splitting as this post was starting to sound. I like AI. I’m old, and this shit lets me do all kind of things I could never dream of doing in my days. I love it. I don’t see it as a stealer of someone else’s skills or talents. I see it as a great helper in doing all kinds of things for me. So much so that my ideas don’t have to just be dreams due to cost and practicality. They can be done, by me. The way I want them. I have a hard time finding a problem with that. Where there is a problem, IMHO, is in people whose only true creative sense is in making money as fast and easy as possible. If all you really put into the project is spent as a 2 minute thought, then I don’t think you can really call it yours. Here in the Circle, we don’t define exact lines. We just want you to be honest, so others know how you judge yourself, and how they should judge your work. When that doesn’t work itself out inherently, we have a Hall of Justice that will review and resolve specific issues on a case by case basis. Those same people that contribute, and review, and create alongside you. We encourage those with like interests in writing and other creative pursuits to drop by gapoz.com and look around. Registration is currently open but subject to approval.
Circle is being built around the question the watermark can’t answer. Before a member of the Circle submits their first work for review by membership, they acknowledge their responsibility in being honest about their use of AI, and the role it plays in their specific works. They accept, acknowledge responsibility and stake their word on it. That means something here. It’s self-attested. There’s no detector gate at the door. There’s no repetition of their promise. It’s your word. We believe it and accept it not because nobody ever fails, but because a signal buried in a word tells you about a machine. A review here is about a person and their work.
A member requests an assignment, reads somebody’s work, and writes a structured and prompted review: what worked best for them. What didn’t. Reviewers are encouraged to provide thoughtful and real feelings about the work, and guided in understanding the various lenses through which they can provide constructive and informed feedback. Writers gain insight and are free to revise and resubmit their work. You earn credit by doing it for other people, and you spend that credit getting it done for yours, or others.
So use the tool if you want. Just don’t try to deceive. You put your name on it. Treat it as you do your name. GAPOZ Press is the people’s publisher, and the people sign their own work.
Also on the beat
- Japan’s book publishers association has asked distributor Nippan whether large book sales went to Anthropic for AI training, after US court papers referenced “Nippan Japan.” Here’s the Sankei report (in Japanese).
- Kathleen Schmidt asks who AI detector Pangram really serves, pointing to conflicts of interest and harm to Black authors, in Who Does Pangram Really Benefit?.
- YouTube is cutting Shorts reach for re-uploads that add “nothing of your own,” and voiceover and template edits don’t count, according to AI Weekly, via 9to5Google.
- The Goncourt named its eight-book second selection today, nearly two weeks after the Orélien withdrawal, and the finalists come October 27, per franceinfo.
- Anthropic settlement payments are expected on or before November 15 for works with no split dispute, and the dispute window has been extended to 60 days, according to this roundup of AI rules for authors.
How this was made
AI assisted research and a first draft of this Circle View. A human reviewed, edited, and is accountable for the judgment and the final take. Disclosure by role — not undetectable.
- A score is not a confession → Page 1 Still think a percentage is proof? Read Do AI detectors work on fiction?
- A Detector is NOT a Judge → Page 1 The short version of why a score can’t judge a book: Do AI detectors work on fiction?
- Prove you didn’t use a machine → Page 1 If someone’s asking you to prove a negative, start here: Do AI detectors work on fiction?
- Disclosure that never faces the reader → Page 2 Here’s how we keep it simple in Circle, no formulas: How Circle handles AI
- How did you use AI? → Page 2 We ask once, and your word does the rest: How Circle handles AI
- Put it on the Imprint Page → Page 2 Why I’d rather skip the label tiers altogether: How Circle handles AI
- Watermarks don’t sign your name → Page 1 Watermarks, detectors, same problem: Do AI detectors work on fiction?
- They knew the shelf would get crowded → Page 2 Slop is the problem, not the tool. Here’s where Circle stands: How Circle handles AI
- Label the voice → Page 2 How we handle AI in Circle, without the labels: How Circle handles AI
- Byline is no Costume → Page 2 Your name, your word, your work. That’s the deal in Circle: How Circle handles AI
