You can spot AI-written text in about three seconds now. Not because the grammar is bad or the facts are wrong, but because every LLM writes with the same voice. Developer Tom Yandell published a detailed breakdown of exactly why, cataloging 12 specific anti-patterns that make AI output instantly recognizable.
The Tell-Tale Dozen
The list is uncomfortably precise. "Performative directness" covers those fake-candid openers like "Look," or "Let's be honest" that signal authenticity without delivering it. "False empathy" is the unprompted "Great question!" that nobody asked for. "Throat-clearing" is the paragraph of setup before the actual point arrives. Anyone who has used ChatGPT or Claude for writing will recognize every single one.
Some of the patterns are structural. LLMs default to triple parallel structure (three-item lists produced by pattern completion), trailing summaries that restate what was just said, and compulsive qualification that hedges every claim into meaninglessness. Others are more subtle: "false concession" preemptively addresses counterarguments nobody raised, and "definite article overuse" uses "the" to imply universal truth from limited observation.
The most damaging pattern on the list might be the last: unoriginal ideas. LLMs are trained on existing text, so their default output is a confident restatement of whatever consensus already exists on a topic. If you're using AI to write thought leadership and you don't edit aggressively, you're publishing the average of everyone else's thinking.
The Fix Is a Writing System, Not a Prompt
Yandell's proposed solution is more interesting than the usual "just edit more" advice. He recommends building a personal voice guide document (he uses a CLAUDE.md file, which is Anthropic's convention for giving Claude persistent instructions) that encodes your specific tone preferences, banned patterns, and preferred alternatives. The model reads this document during every session, so it starts closer to your voice instead of its default.
The workflow he describes is collaborative iteration: stepping through sections conversationally with the model, treating each sentence as something you need to genuinely stand behind before moving on. His honest admission is that this process can take longer than writing from scratch, which undercuts the "10x productivity" narrative around AI writing tools but aligns with what most serious writers have found.
Useful for Anyone Publishing AI-Assisted Content
This matters because the bar for AI-assisted writing is rising fast. A year ago, readers gave AI-generated text a pass because the technology was novel. Now, the patterns Yandell describes trigger immediate skepticism. Marketers, content creators, and anyone publishing under their own name should treat this list as a checklist. If your drafts hit three or more of these patterns, readers are clocking it as AI slop whether they consciously realize it or not.
The practical takeaway: AI writing tools are best used as a starting point that requires significant human editing, not as a finished product pipeline. The people getting the most value from them are the ones spending the most time rewriting the output.