This edition is a bit special, because I am coming back to a question that concerns all of us at work → how do you know whether a piece of content was generated by AI?
You can have doubts about an email reply, a colleague's message on Slack, or even a deliverable.
AI is everywhere today, social networks included. Across more than a million posts analysed on LinkedIn, 41% of long-form content is entirely AI-generated. On Substack, the same measurement tool finds closer to 10%.
But since 21 July, Substack has put the subject on the table by integrating Pangram to analyse posts (more on that below)
Chris Best, Substack's CEO, talks about "Claudefishing" to call out the AI content that pollutes feeds and makes human writing harder to find.
So I wanted to dig into the subject, and give you what you need to spot AI where it hides.
The real issue → the cost transfer
Let me share a personal example. I get more and more email replies where I immediately recognise the model's fingerprint (usually Claude's). Flawless structure, every one of my points picked up in order, a smooth tone…
But it is so frustrating to write to your accountant and end up with an answer from Claude…
At that point the value of the email drops to zero. If the person pasted everything into Claude, I could have done it myself and saved the round trip. We exchanged two emails and created zero information.
That is the real issue, and it is not a question of authenticity in the moral sense. It is a question of cost. The person writing saves time, the person reading pays for it.
Nobody blames anyone for using a spell checker. What breaks trust is delegating without reviewing. And once the doubt is there, the relationship gets harder..
The signals that give it away
If you use Claude / ChatGPT / … regularly, you have probably started noticing the patterns that keep coming back. Here is what stands out, from the most visible to the deepest:
Punctuation
- The long dash (the em dash): the best-known signal, and it is even more telling in French, where it is almost never used while it is common in English.
- The semicolon ; → inside a bullet list or in the middle of a sentence, it is very rare in French.
- Copy-paste leftovers. Stray ** or ## in a field that does not render markdown, curly quotes in a plain-text email, or the famous utm_source=chatgpt.com stuck at the end of a shared link.
Syntax
- See-saw negation: "it is not a tool, it is a new way of working". It adds nothing, it is heavy syntax, Claude does it all the time.
- Long sentences stuffed with repeated adverbs: currently, truly, genuinely, particularly, undeniably, deeply, …
- The systematic rule of three: "fast, reliable and scalable"
- The explanatory tail at the end of a sentence: "…, highlighting the importance of collaboration". Nobody writes that naturally…
Substance
- No proper nouns. "A client", "a tool", "a platform". Someone who did the thing names the thing.
- No numbers. "We work with around fifty companies.."
- Vague authority: "studies show" / "experts agree"
- Everything is positive and certain, no failure, no question left open, no friction.. A real field report always contains something that went wrong.
There is no point hunting down the people who write with AI. That is not where it breaks. The issue is delegating 100% without bringing anything of your own. The human brings expertise and experience, and that is exactly what matters.
If you spot other signals, drop them in the comments, it is always fun to see what comes up.
But detection stays fallible
I find it hard to trust detectors when I see this kind of result on ZeroGPT with the French Declaration of the Rights of Man, for instance:
Like the models, the detectors have improved. Pangram looks particularly strong.
Independent researchers measured a false-positive rate of 1 in 10,000. That is better, but the tool still makes mistakes. Here for example, on this LinkedIn post:
Whose content is flagged as human (the rest is correctly flagged as AI)
The tools own up to a few blind spots:
- Short texts under 50 words
- Formatted text such as short bullet lists, outlines, or very formulaic writing
- Reworked text: content run through a style skill, then reread and edited, becomes very hard to classify. (more on that below)
The stronger the model, the more you need to check
It is slightly counter-intuitive, but the more powerful the models get, the less time we take to verify the information.
On the other hand, Anthropic said at the Opus 5 launch that it had found a significant number of cases where the model states an answer confidently while not being sure.
Opus 5 hallucinates slightly more facts than Opus 4.8 (about 14 points apart)
Even though overall accuracy is higher:
That comes from the model answering more often when it is uncertain (rather than saying "I don't know")
Remember that these models are trained to satisfy you, optimised so that you approve the answer and move on.
Skills fix the form, not the substance
If you want content that feels less like AI, you just need to create skills (files that give instructions to the models). Two are enough:
1/ A style skill → you pull your LinkedIn posts (or the articles of an author you want to imitate) through Claude Cowork, and extract a voice.md with your tone, your examples, your writing tics. The model stops producing generic text.
2/ An anti-AI skill: no long dash, no semicolon, no see-saw negation, no rule of three, sentences of uneven length. It corrects every signal listed above.
I already covered this in my article on Claude Cowork, with a voice.md file (your tone, your examples, your writing tics):
Claude Cowork: the tool you are not using (and that changes everything)
It is a fix for the form problem. Skills make AI content undetectable, but they do not solve the substance problem, the reader's problem, who still reads a text nobody has thought through.
How I use it on this newsletter (and in general)
Let's be honest, you are probably wondering: is this newsletter written by AI?
First "no", then "yes", and finally "no".
Writing an edition takes me roughly 4 hours.
Concretely, out of those 4 hours, half goes into ideation: choosing the topics and how to bring them into the different sections. That is exactly the part I cannot delegate.
Then I usually go back and forth with the AI to validate the structure. The AI generates a first version inside my Claude Project that gathers all my newsletters and my writing style. I go back over the whole thing, and there is often a lot to change (anecdotes, personal opinions, numbers, ..). It can include back and forth with the AI on some parts to make the writing easier (the form) once the substance is right.
In short
AI detection works better than it used to (including human detection, once you have the patterns in mind), but only on content that has not been optimised. You can get around that quickly with skills. Either way, it will never fix the underlying problem (or only with difficulty)
The real risk is not getting caught. It is wasting the time of the people who expect value from you, readers, clients or colleagues, and paying for it in credibility without ever knowing.
