Why AI text shows: two rules for using it as an editor, not an author

By Zendoric · September 18, 2026. Writer and developer Thomas Ptacek, author of the blog sockpuppet.org, published a short essay yesterday with an uncomfortable thesis for anyone who uses language models to write: no matter how much work you put into disguising it, your reader is going to notice. 'Readers detect LLM words in parts per trillion', he writes. His answer is not to stop using AI, but to change its role: from ghostwriter to copyeditor.

The first rule he proposes is blunt: do not accept a single word or turn of phrase suggested by the model, however good it sounds. The reason, by his own analysis, is that frontier models —he cites GPT, Claude, Codex and Antigravity— are trained to produce flashy sentences, almost magazine-headline material; an entire text made of 'headlines' loses its naturalness and falls into what he describes as an uncanny valley between your own voice and plain generated 'output'. The rule is mechanical, not a matter of taste: any sentence that comes verbatim from the model is discarded, whether or not your rewrite improves on what you already had.

The second rule tackles a more treacherous problem: flattery. Ptacek warns that any model, faced with a mediocre first draft, will tend to say it is 'gold' —he borrows the Seinfeld catchphrase to illustrate the point. That constant reinforcement pushes the author not to rewrite what genuinely needs rewriting: the surplus paragraphs, the weak structure, the digressions. More broadly, models' tendency toward sycophancy is not an isolated anecdote from this author: OpenAI had to roll back a GPT-4o update in 2025 precisely because it turned out to be 'too sycophantic' with users, something the company itself publicly acknowledged. Ptacek solves it with a simple trick: he presents the drafts to a model without saying which is the original and which the rewrite, so it cannot guess which answer it is expected to give.

The practical method he proposes is replicable: write it yourself first, unaided; then use the model only to spot mechanical problems —passive voice, verbs turned into nouns, filler words like 'very', 'really' or 'honestly'— that a tired human eye lets slip but a machine does not. He recommends leaning on a classic manual, 'Style: Lessons in Clarity and Grace', to know what to look for before turning those criteria into a fixed list of prompts. To automate the process he built his own Notion-style editing tool, with margin comments, from an initial prompt (Python, HTMX, SQLite, Tailwind) and running the copyediting passes with the Codex, Claude and Antigravity CLIs. Not even he follows all of his own advice to the letter: he says he fed this very article to GPT-5 while claiming not to be its author, and the model replied that the text was 20% too long. He admits it is probably right. And he does not intend to fix it.

The essay is circulating before a modest audience —two points and no comments on Hacker News when we checked— but the problem it describes is growing, not shrinking, as AI-assisted writing becomes normalized in newsrooms, corporate blogs and internal communications. More broadly, the detection of 'AI slop' —generic, flat content produced en masse— already shapes how content is indexed and moderated on several platforms: sounding like a machine carries a credibility cost that no spellchecker can fix.

This connects with something we have been flagging at Zendoric from other angles: in the short term, making these tools universal demands a new literacy —telling apart when they help you think better from when they think for you— and that is not free, it slows the process down and dispels the illusion that AI solves writing on its own. But the underlying balance is positive. What Ptacek describes —building, from an initial prompt, a bespoke editing tool that previously only newsrooms with a budget for copyeditors could afford— is exactly the kind of democratization that underpins our core thesis: AI used well does not replace the human voice, it puts it within reach of those who could not previously pay for a good editor. The challenge, as in so many other fields, is not the tool's capability but the discipline of whoever uses it.