Саймон Уиллисон сделал на Fable 5 детектор клише из текстов нейросетей

Developer Simon Willison released a free web tool called LLM cliché highlighter on July 17, 2026, which highlights about 11 common stock phrases typical of text generated by neural networks.

How the tool came about

Simon Willison put together the LLM cliché highlighter after once again stumbling on an article stuffed with AI clichés like "no fluff, no filler, no jargon". He built the app using vibe-coding — that is, he asked Anthropic's Fable 5 model to write all the code from a description of the task, rather than programming it by hand.

Willison is one of the most cited writers on LLMs, creator of the Datasette library and the `llm` command-line tool. On his blog simonwillison.net, he regularly examines how large language models write, make mistakes, and give themselves away through characteristic turns of phrase.

"I'm tired of reading yet another article stuffed with clichés from LLM-written text," —

Simon Willison, on the blog simonwillison.net.

How does the tool work?

The LLM cliché highlighter analyzes pasted text and highlights in yellow the fragments that match preset patterns of machine writing. A screenshot from Willison's blog shows the set includes 11 rules, which can be toggled on and off individually.

• Published — July 17, 2026, by author Simon Willison
• Built using vibe-coding on Anthropic's Fable 5 model
• Highlights about 11 common LLM cliché patterns
• Loading an article by link — via the r.jina.ai service
• The "Show just the highlights" mode displays only the highlighted spots

The highlighting has two levels: pale yellow marks the entire "suspicious" phrase (flagged sentence), while a saturated color marks the specific turn of phrase within it (pattern match). Among the example patterns are constructions like "Is real … and/not" and "Worth naming"; the demo phrase in the interface reads "That loss is real and it's worth naming," where the tool highlights two matches at once.

It also works with other people's articles: you paste the address into the URL field, the page is pulled in via the r.jina.ai service, after which the text is run through the same rules. For a quick check there's a button to load a ready-made example and a mode that leaves only the highlights on screen.

Why this is useful

The LLM cliché highlighter gives a quick way to check your own draft for the "smell of a machine" before publishing. The author pastes in the text, sees the highlighted stock phrases, and rewrites them — exactly the turns of phrase that today's models insert most often and by which text is recognized as generated.

The approach reflects a broader problem: as LLMs write more and more public content, the set of their favorite turns of phrase becomes a recognizable marker. Tools like this one turn the intuition that something "sounds like ChatGPT" into a concrete list of rules that can be applied mechanically rather than by eye.

What this means

LLM clichés are turning into a measurable checklist: something that used to be a matter of an editor's taste can now be checked with a single click. For authors it's a cheap filter against machine-style writing; for readers, a reminder that recognizable turns of phrase give away generation.

Frequently Asked Questions

What is the LLM cliché highlighter?

The LLM cliché highlighter is a free web tool by Simon Willison, published on July 17, 2026. It highlights about 11 patterns in text that are characteristic of writing by large language models.

How do you check an article for AI clichés?

You need to paste the text into the tool's field or provide the article's URL — the page is then pulled in via the r.jina.ai service and run through the set of rules. Matches with machine-writing patterns are highlighted in yellow, and the "Show just the highlights" mode leaves only those on screen.