I’ve used Claude quite a bit for both the upcoming new book and the next edition of Managing Humans. I wrote the entire book, but, as I’ve done for years, I’ve used the robots for editing, research, and data wrangling.
When Anthropic announced they were going to do “text and image” watermarking, my eyebrow was raised. When humans I trusted began to analyze the announcement, both eyebrows went up. This morning, as I’ve done my own research, I’m furious.
As of this moment, this is what I’ve researched.
- Anthropic states that models born on or after August 2, 2026, will have watermarking and that older models will be retrofitted by December 2, 2026. Anthropic states these changes are in progress, with no per-model disclosure, which means they can switch marking on silently, with no notice promised.
- There is no way to opt out, no API tier, no plan tier; enterprise-included. Applied server-side, below the model, worldwide.
- No detector currently exists — only Anthropic can check.
As a writer, no sweat, right? I write my own content. I bang on my own keys. Here’s the issue. Each of the following actions watermarks text:
- Synthesis
- Proofreading
- Translation
- Summarization
- Critique1
And most infuriating, the entire burden of their unannounced rollout is on me, and none of it is on them.
Fine. Make it easy. I’ve been a Claude Max user happily from the moment the program appeared. This is what I am doing:
- Building gates to make sure all Claude text or images are never injected into any of my projects. A good practice anyway.
- Researching migrating off Claude for all writing-related work. Stay tuned.
- Getting in the mindset to cancel my Claude Max plan.
As a human who has had to wrangle with EU regulations in the past, I am abundantly clear what’s involved in the laborious bureaucratic process. I can guess what threats Anthropic is facing. However, this is a tone-deaf, clumsy, and alarming opening salvo in their watermark strategy.
Yes, I use Claude Code a huge amount for tool and application development. Go ahead. Watermark away. I am not claiming that I wrote any of that code. My writing is my work, and Anthropic’s current strategy is aggressively writer-hostile.
Update 8/15/26: https://www.anthropic.com/news/claude-text-watermark
- The forthcoming detector (“soon”, “being worked out”) will only say Claude is likely involved with the content at some point. It can’t distinguish “Claude wrote this” from “Heavily edited by Claude”.
- No opt-out. Confirmed.
- Doesn’t work well on small samples and light edits.
- What this means: if I ask Claude to critique a piece and I happen to cut and paste part of its reply into the text, then that could be marked. ↩︎
The number of insignificant bits in short sections of plain text is too small for me to worry about it in a meaningful way marking most of my writing as goblin tainted. The implementation is also too broad, because it must be. They felt in order to be compliant they must add things as benign as proofreading, therefore I’m perfectly willing to admit that 100% of my long-form writing is AI touched.
FWIW knowing your existing corpus and your attention to detail I can’t imagine anyone thinking any word gets into your writing where you do not take personal responsibility for its inclusion, regardless of who or what might have suggested it originally.
Between this and the crackdown on using subscriptions with third-party clients such as opencode, hermes, openclaw, etc…. Anthropic seems to be wanting to throw away their lead in the business space back to OpenAI.
Hopefully at some point there’s a happy medium in AI costs for the consumer that’s in-between the frankly outrageous costs of using pure API interface and what is for a lot a loss-leader in the subscriptions.
If everyone cancels their Claude Max plans over this then I will be glad because people in Memphis will breathe easier. I don’t know how you can be fine with that data centre being the one that serves you.
I don’t see any downside to this measure by the EU. Privacy is not impacted, nor is performance or output quality. If one uses LLMs for any purposes *other than claiming they actually wrote what it produced*, then it’ll be fine. Transparency is good. And in any case, this’ll be rolled out by the other major AI companies.
There’s an obvious course of action here that springs to mind.
Please don’t use AI to write. I’ve followed your writing for years, since I came across Managing Humans, and I want to read your words, human words, not LLM output.
I know this might not make sense to you, but please believe me when I say that to me, and to a large number of people, using an LLM for creative work kills all our interest in it.
_Any_ amount, even a tiny one? In the output, yes. It’s like a pub selling watered down beer. I can see why they might have a financial interest in doing that. And there’s a certain amount of it that they can probably get away with. But if the customers notice, it doesn’t matter how little water it was. Because nobody wants watered down beer.
Do not worry — I never use the robots for writing. However, I have — for years — used them for proofing and have my text accidentally watermarked for doing mechanical work is quite dumb.
Time to go back to a human editor, I’m afraid!
If I find out that AI has barely grazed something I’m consuming – especially a book – then I find myself very very frustrated. I’m devoting the time and energy to buy, read, subscribe to a publication; those people should spend the time to write it. It frustrates me more than reading a grammatical mistake or a typo.
Even using it in minor ways such as editing contributes to the proliferation of AI in media that should be deeply deeply human, so I don’t see any downside in this regulation whatsoever.