The em-dash tell is finished. Everyone learned it, models were tuned away from it, and writers now strip it by reflex. So the detection moved up a level: to rhythm, to structure, to the little verbal throat-clearing that models do. Ruben Hassid keeps a running file of the patterns that replaced it and posted nine of them on X on 13th August 2026. He writes the How to AI newsletter and, by his own line, hates AI writing while writing about AI for a living.
The post is worth keeping because it is specific. Most "spot the AI" advice is a vocabulary blacklist. This one names sentence shapes, and sentence shapes are what actually leak.
The nine tells
| # | Pattern | Example | Fix |
|---|---|---|---|
| 1 | The negation setup | "That's not compliance. That's stalling." | Write the second half only |
| 2 | Two fragments glued | "Fast. Simple." | Pick one, write a sentence |
| 3 | The empty metaphor pair | "Less a hammer, more a scalpel." | Say what to do instead |
| 4 | Self-applause | "And that matters." "That's the part everyone misses." | Delete. Nothing is lost |
| 5 | The X-of-Y shortcut | "It's the Excel of AI agents." | Only works if the reader knows both |
| 6 | Warming up | "Here's the thing." "Let me be clear." | Start one sentence later |
| 7 | Always three | "Faster, cheaper, smarter." | Reality gives you two, or five |
| 8 | The fake range | "5 to 10 minutes." | A range means you never measured. Say 7 |
| 9 | The closing recap | "In short..." "At the end of the day..." | Stop typing |
Numbers 4, 6 and 9 are the cheapest to fix and the most common. They are all the same fault: the model padding the edges of a paragraph because it has been rewarded for sounding finished.
Number 7 is the sharpest observation in the list. The tricolon is a real rhetorical device, which is why it survives every blacklist. Models reach for it whether or not the world supplied three items. When you see three parallel adjectives, check whether the third one is doing any work.
Number 8 is the one I had not thought about. A hedged range is a confession that nobody ran the thing. It applies well beyond AI text.
Why the em-dash advice expired
The em-dash never proved anything. It became the tell because it was easy to grep and it spread fast enough that everyone patched it. Now the opposite failure shows up: clean, well-balanced prose gets accused of being generated. A single character was never going to hold as a signal.
The nine patterns above are harder to strip because they are habits of structure, not characters. That also makes them a better checklist, and a better thing to keep in a file.
The six-step fix
Hassid's practical half is the more useful half.
Step 1. Put Use ASD-STE100. in the prompt. Two words, and by his estimate they remove most of the bad writing. Works on Claude, ChatGPT, Gemini and Grok.
Step 2. Turn it into a skill so you never retype it. He packaged the full rule set as a free Claude skill: upload the zip under Customize > Skills, then call it with /ste. The download is behind an email opt-in for his newsletter, so budget an inbox for it.
Step 3. Switch it off for anything with a heart. His test: a poem came out unreadable, a LinkedIn post came out sounding like a manual. Use STE for guides, steps, explanations, emails and documents. Not for voice.
Step 4. Add Zinsser to your CLAUDE.md. STE gives you simplicity, brevity and clarity. It strips humanity out on purpose, because a maintenance manual should not have a personality. Adding Zinsser's four principles puts the fourth back.
Step 5. Keep a forbidden-patterns file. Copy the nine tells into a markdown file and tell the model to check every draft against it. Add a line each time you catch a new one. The file compounds.
Step 6. Start fresh chats more often. Long context degrades writing quality: the model repeats itself and drifts back to pretty sentences. When the window feels half full, open a new chat and paste the files back.
What ASD-STE100 actually is
It is a controlled language, not a style guide. ASD (Aerospace, Security and Defence Industries Association of Europe) owns it and the STEMG maintenance group keeps it current.
| Origin | Late 1970s, developed by AECMA with European and American airline input |
| First release | AECMA Simplified English Guide, 1986 |
| Purpose | Make aircraft maintenance manuals readable by non-native English speakers |
| Status | International specification in 2005, international standard in 2025 |
| Current | Issue 9, 15 January 2025: 53 writing rules, 555 dictionary entries revised |
| Cost | Free official PDF from the ASD site |
The logic is worth understanding before you use the trigger word. In a maintenance manual an ambiguous sentence can kill someone, so STE fixes one meaning per word, bans synonym variety, forces the active voice, caps sentence length, and restricts you to an approved dictionary. Every property that makes it safe for a manual is also what makes AI prose readable.
One caveat: telling a model "use ASD-STE100" invokes its idea of the standard, not the standard. It has not memorised the dictionary. You get the discipline, not compliance.
Against my own setup
Four of the six steps are already running here, which is why the post landed.
| Step | Status |
|---|---|
| 1 and 2. STE | Written into my global CLAUDE.md as chat rules. No skill needed |
| 3. Off for creative work | Already scoped: STE governs chat, never deliverables |
| 4. Zinsser | The four principles sit above STE in my config, with a precedence order |
| 5. Forbidden file | anti-ai-writing-guidelines.md, read once per session |
| 6. Fresh chats | The gap. I let sessions run long and the writing does sag |
The precedence rule is the part I would add to his version. STE and Zinsser collide often, so mine settles it up front: clarity beats simplicity beats brevity beats humanity. Never cut a word the meaning needs, never write a warm sentence that hides the fact.
His step 5 also pairs with the de-slop approach: keep the tells in a file, then run a gate over the finished draft rather than trusting the prompt to hold.
De-Slopping AI Output: Ben AI's Skill vs My Guidelines
Where I disagree
The claim that models are "trained to not write with em dashes anymore" is loose. Model behaviour shifted and users strip them by hand, and the two are hard to separate from outside.
The nine tells are also detection heuristics, not writing rules. Every one of them is a legitimate device that a human writer uses deliberately. The negation setup is good rhetoric when the contrast is real. The problem is frequency: the model reaches for the shape when the content does not earn it. Treat the file as a flag, not a ban.
Further reading
- Original X post by Ruben Hassid
- ASD-STE100 official site and free Issue 9
- ASD announcement: Issue 9 becomes an international standard
- Simplified Technical English on Wikipedia
- How to AI newsletter