It marks the text. It doesn't carry your prompt, your client, or a percentage of who wrote what. Your actual exposure is sitting in local disclosure rules nobody at your firm has inventoried.
This raises a much bigger question than whether a “Claude watermark” exists. As AI-written content becomes nearly indistinguishable from human writing, the tools used to identify it will inevitably become more sophisticated too. For writers, the real advantage may not be avoiding AI, but making sure our ideas, experiences and perspectives are genuinely our own. AI can replicate a writing style far more easily than it can replicate original thinking.
A watermark may be most useful when we treat it as provenance, not judgment. It can tell us that a model participated in producing the text, but not why it was used, how much human reasoning followed, whether the output was verified, or who ultimately owned the decision. The risk may be that organisations turn a detection signal into a conclusion it cannot support. Good governance probably needs both layers: evidence of AI involvement, and a separate record of the human process and responsibility around it.
This is just beyond excellent Steve. The thinking, the emphasis as always ensures we stay focused on what matters. And your consistently concise, reader-considerate prose is greatly appreciated. I read it. I got it. I enjoyed the reading experience. I don’t care where or how this article is watermarked. Thank you!
Walking around Portugal this month, I can’t even walk into a corner fruit market and not see images of grapes on sale that come with an AI-generated disclaimer.
I know some will roll their eyes saying this is a socialist nanny state hell. But others have had enough of the sloppification of everything. Sensible regulation hypochondria is a thing, and not all windmills cause cancer.
Really interesting piece and discussion. I approach this from a behavioural science perspective, and the question it leaves me with is what meaningful human responsibility looks like in practice.
As you mention, we can’t always “judge the final product” independently of the process. A convincing output doesn’t tell us whether appropriate human judgement, review and verification actually happened.
I see the challenge firms will increasingly face is preventing AI from becoming a shortcut at precisely the points where human judgement matters most — and creating the expectations, norms and working environment that keep people actively engaged in that judgement, even when the technology makes it easy not to do so.
That essay really landed with me. It has materially sharpened my thinking. It’s not merely a grammatically correct piece of writing; it’s a clear statement that cuts through the noise. My honest take is that there is a real message embedded within it.
This is a well-balanced post on this topic, and I appreciate it. My fear is that AI watermarking + book skinning + regulatory capture has the potential of putting these top tier frontier labs in a position of political power under the moniker of safety. Perhaps I'm exaggerating, or just need to back off on the coffee ... but that's my story and I'm sticking with it ...
... well at least that's the futuristic fictional fable I wrote over on my substack on this topic. Still, thanks for getting us past the hyperventilation and hubris on this topic.
Thanks for the detailed analysis on this watermark thing- fascinating- but honestly I am just staring at how absolutely adorable your dog Magnus is and sorry, he is just wayyy too adorable for me to focus on your essay!! What a darling!!!
One small legal correction: Article 50 does not itself prohibit users from removing or altering an AI watermark.
What Article 50(2) actually says is:
“Providers shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”
That is an obligation on providers to mark outputs, not a general prohibition on downstream users removing the mark. (EUR-Lex)
The voluntary Code of Practice may impose additional commitments on its signatories, and local court, professional or contractual rules may of course prohibit tampering independently. But that is legally different from saying, as the article does, that “Article 50 prohibits deliberately removing or altering AI markings.” (Digitale Strategie der EU)
Spot on. Deployers are mostly off the hook with text unless it is aimed at informing the public… (and their own transparency obligations are not tied to removing an original watermark)
For an article about AI disclosure which I could tell was composed by AI like three sentences in, I thought it was interesting that the article made no mention of this fact. I had to click the "scan for AI" Substack button to find your disclosure.
Kane - thanks for the comment and I agree. I’ve had a lot of discussions in the past few days on this topic and to me the thing that matters is the content created - is it valuable and insightful? Is it useful? If so - how people use different tools and what tools they use is (to me) not that important. Thanks for the comment. I appreciate it.
That graphic does not show Magnus to scale!
This raises a much bigger question than whether a “Claude watermark” exists. As AI-written content becomes nearly indistinguishable from human writing, the tools used to identify it will inevitably become more sophisticated too. For writers, the real advantage may not be avoiding AI, but making sure our ideas, experiences and perspectives are genuinely our own. AI can replicate a writing style far more easily than it can replicate original thinking.
Your clients can finally tell whether you said hello yourself or had AI help you do it.
Compliance is the ultimate gate keeper.
A watermark may be most useful when we treat it as provenance, not judgment. It can tell us that a model participated in producing the text, but not why it was used, how much human reasoning followed, whether the output was verified, or who ultimately owned the decision. The risk may be that organisations turn a detection signal into a conclusion it cannot support. Good governance probably needs both layers: evidence of AI involvement, and a separate record of the human process and responsibility around it.
This is just beyond excellent Steve. The thinking, the emphasis as always ensures we stay focused on what matters. And your consistently concise, reader-considerate prose is greatly appreciated. I read it. I got it. I enjoyed the reading experience. I don’t care where or how this article is watermarked. Thank you!
Who's on the ladder?
Also, seriously, I appreciate your articles so much. I'm learning so much that I can use to do a better job for our clients.
Walking around Portugal this month, I can’t even walk into a corner fruit market and not see images of grapes on sale that come with an AI-generated disclaimer.
I know some will roll their eyes saying this is a socialist nanny state hell. But others have had enough of the sloppification of everything. Sensible regulation hypochondria is a thing, and not all windmills cause cancer.
Really interesting piece and discussion. I approach this from a behavioural science perspective, and the question it leaves me with is what meaningful human responsibility looks like in practice.
As you mention, we can’t always “judge the final product” independently of the process. A convincing output doesn’t tell us whether appropriate human judgement, review and verification actually happened.
I see the challenge firms will increasingly face is preventing AI from becoming a shortcut at precisely the points where human judgement matters most — and creating the expectations, norms and working environment that keep people actively engaged in that judgement, even when the technology makes it easy not to do so.
Thanks for sharing!
Ah so it’s another pattern recognition like pangram is doing essentially but reversed
Claude will write a certain way that it shows it was written by Claude.
Same as Google SynthID Text.
As soon as you adjust the words and sentences to give your (human) touch and their whole watermarking goes to toss
Bah, not impressed.
I think of it as AI Detection = Looking at a pencil
That essay really landed with me. It has materially sharpened my thinking. It’s not merely a grammatically correct piece of writing; it’s a clear statement that cuts through the noise. My honest take is that there is a real message embedded within it.
This is a well-balanced post on this topic, and I appreciate it. My fear is that AI watermarking + book skinning + regulatory capture has the potential of putting these top tier frontier labs in a position of political power under the moniker of safety. Perhaps I'm exaggerating, or just need to back off on the coffee ... but that's my story and I'm sticking with it ...
... well at least that's the futuristic fictional fable I wrote over on my substack on this topic. Still, thanks for getting us past the hyperventilation and hubris on this topic.
Thanks for the detailed analysis on this watermark thing- fascinating- but honestly I am just staring at how absolutely adorable your dog Magnus is and sorry, he is just wayyy too adorable for me to focus on your essay!! What a darling!!!
One small legal correction: Article 50 does not itself prohibit users from removing or altering an AI watermark.
What Article 50(2) actually says is:
“Providers shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”
That is an obligation on providers to mark outputs, not a general prohibition on downstream users removing the mark. (EUR-Lex)
The voluntary Code of Practice may impose additional commitments on its signatories, and local court, professional or contractual rules may of course prohibit tampering independently. But that is legally different from saying, as the article does, that “Article 50 prohibits deliberately removing or altering AI markings.” (Digitale Strategie der EU)
Spot on. Deployers are mostly off the hook with text unless it is aimed at informing the public… (and their own transparency obligations are not tied to removing an original watermark)
Except you are incorrect. A watermark is the necessary conditions for individual tracking of content.
For an article about AI disclosure which I could tell was composed by AI like three sentences in, I thought it was interesting that the article made no mention of this fact. I had to click the "scan for AI" Substack button to find your disclosure.
Who cares? The article was thoughtful and well-written.
Judge the final product, not the tool used along the way. Which, ironically, is the very point AI detection misses.
Kane - thanks for the comment and I agree. I’ve had a lot of discussions in the past few days on this topic and to me the thing that matters is the content created - is it valuable and insightful? Is it useful? If so - how people use different tools and what tools they use is (to me) not that important. Thanks for the comment. I appreciate it.