You can recognize AI-generated text by dashes in every sentence, one idea spread across three paragraphs, and phrases that don't convey anything. This can be fixed: there are clear rules and ready-made tools.
1. "Write, Shorten"
A book by Maxim Ilyakhov and Lyudmila Sarycheva about information style. It wasn't written about AI, but its advice definitely applies to machine-generated text:
- Remove introductory phrases like "it's important to note." Get straight to the point.
- Replace judgments with facts. "The service is good and useful" tells you nothing. "The service gives 50 requests per day without a card" says everything.
You can present the book's rules to the model as instructions: what to strike out and what to replace it with.
2. Wikipedia: Signs of AI writing
Wikipedia editors faced a stream of generated edits and compiled signs of such text on one page. The list is long, here are two of the most notable points:
- Dashes instead of commas. Models use them where a comma or period would suffice.
- Repetition. One thesis is repeated from paragraph to paragraph in different words.
The page is convenient to use as a checklist when proofreading.
3. Ready-Made Tools
- humanize-writing-skill — a skill for an agent that edits text in three passes. First it removes marker words, then template constructions from the Wikipedia list, finally adds natural turns of phrase. The word list is more geared toward English.
- AIWords — JSON with words and phrases that models overuse. Serves as a stop-list.
- gptslop — a similar list that the community maintains.
How to Put This Into a Working Process
- Give the model rules before writing text, not after.
- Include two or three of your own texts as tone examples.
- Run the finished draft through a checklist: dashes, repetitions, empty judgments.
- Read it aloud. Where you stumble, rewrite it yourself.
Tools remove the external signs. The author still brings specifics, numbers, and personal experience to the text.