Skip to content
Back/Writing

AI Writing Style Signals

View in Graph
Updated 2026-08-26
2 min read
421 words

AI Writing Style Signals

What it is

AI writing can be detected less by obvious words and more by style signals: stop words, function words, phrase rhythm, perplexity, consistency, and time drift. Human writing is often more variable and less internally consistent than model output. Deft’s DFT v1 tests add a complementary observation: models can also be diagnosed at the batch level by measuring how their output distribution diverges from human writing on n-gram usage, embedding-space distance (MMD), and judge-model quality (JMQ).

Why it matters

If style is partly subconscious, a good AI writing workflow cannot rely only on “write in my style”. It needs real samples, style tests, updated examples, and human editing. The goal is not to remove all variance, because variance itself is part of human voice. Distribution-aware evaluation adds a second layer: even when individual drafts look acceptable, the corpus as a whole can still expose machine habits.

Key points

  • Traditional stylometry finds authorship signals in stop words and word order.
  • Human writing is more variable than machine writing; LLM output also drifts behind evolving human vocabulary after release.
  • Recognizable AI tells include correlative constructions like “not X, but Y”, em dash overuse, parallel triplets, and closing sentences that assert something “matters”.
  • Deft’s experiments suggest that reducing these tells through distribution-level training can make prose less predictable, but not necessarily clearer or better structured.

Evidence across sources

Source Key Claim Relevance
Every: The Science of Why AI Still Can’t Write Like You Stop words and word order are strong authorship signals; human writing is more variable; LLMs exhibit time drift Establishes style-signal framework
Every — I Tried the AI Model Built to Fix AI Writing DFT v1 reduces some AI tells at the distribution level, yet its output remains dense, poorly sequenced, and prone to strict-mode hallucination Shows that removing style signals does not automatically improve writing quality

Open questions

  • Which style signals matter most for Chinese writing and bilingual writing?
  • Can Jean’s own writing samples become a reusable style profile without overfitting into a rigid template?
  • How should wiki output distinguish “AI draft for structure” from “human voice final draft”?
  • When a model reduces obvious AI tells, what new signals should editors watch for instead?

Sources

Synthesized from 2 sources
  • Every: The Science of Why AI Still Can’t Write Like YouSupporting source listed by this page.Whole pagemediumbody
  • Every — I Tried the AI Model Built to Fix AI WritingSupporting source listed by this page.Whole pagemediumbody

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 2 sources.

    From Every: The Science of Why AI Still Can’t Write Like You, Every — I Tried the AI Model Built to Fix AI WritingTo AI Writing Style Signals
    Sources: raw/to-learn/newsletters/every-the-science-of-why-ai-still-cant-write-like-you.md · raw/newsletters/Every/2026-08-26 I Tried the AI Model Built to Fix AI Writing.md

Linked from