Martin Fowler, the well-known software engineer and technical writer, devoted two blog posts in late October 2026 to what he calls the 'LLM Voice' — how readers intuitively spot AI-generated text, and the controversy and limits of that instinct. In his Oct 29 Bliki post 'LLM Voice', Fowler wrote that reading certain pieces gives him a 'growing feeling' that he is reading the output of a large language model, even though he cannot say exactly why.
The instinctive reaction to 'LLM Voice' and doubts about tools
Fowler describes the feeling as accompanied by an 'atrophying' desire to read on, sometimes to the point of 'revulsion'. He mentions common tells but stresses that 'it's not the em-dashes, it's the load-bearing tapestries' — a phrase pointing to the piled-up imagery common in AI text.
Fowler is wary of tools that claim to distinguish AI text from human text. He doubts their accuracy and argues that relying on them 'misses the point'. The reaction to LLM Voice is a human one, he writes, a sense of 'the uncanny valley of prose'; by the time a tool is needed to judge, 'it's already too late, even if the tool confidently tells me there's no LLM present'.
Self-doubt and the 'one-strike' effect
Fowler also questions whether his own ability to spot LLM Voice is accurate. Language changes over time, he notes, and complaining about how young people talk is a common marker of age. He says phrases like 'adopting a new sales motion' or 'a sharp round-up of news' make him uncomfortable, but he has heard humans say the same things, so this is more a generational difference than a carbon-versus-silicon one.
In his Oct 21 Fragments post, Fowler cited the experience of Gergely Orosz. Orosz ran an essay competition explicitly labeled 'No AI' and found that some entrants had used AI. Orosz said that when he starts to suspect some part is AI, and the person is someone he doesn't know, he cannot tell how much of it is AI, nor does he want to invest the time and effort to figure it out.
'I fear most people do not realize this last part: that when it takes time and effort from others to figure out how much is AI and how much is you, I'll just assume it's probably 100% AI and 0% you.' — Gergely Orosz
Orosz further warned that in making any content for others, even a little AI use can trigger the assumption that it is 'all AI and not worth paying attention to'. Fowler broadly agrees, saying that once he hears LLM Voice he 'flips the bozo bit' and stops reading.
Legitimate uses and questions of fairness
Fowler also offers a caveat: how accurate is his own 'neural network' at AI detection? Have humans absorbed AI-style elements into everyday communication? More importantly, AI has some legitimate uses. He cites Jim Highsmith's story of how his 'second brain' helps him cope with Parkinson's disease, and people with limited English using AI for translation.
Fowler asks: when he reads or hears such people's content and his AI-voice alarm goes off, is that unfair to them? The question shifts the discussion from detection technology alone to the ethics of use cases.
Fowler's posts offer no clear method for identifying AI text; instead they lay out the tension between instinct, tools and legitimate use. As AI-generated content becomes more common, how readers trade off efficiency against fairness, whether detection tools can become more reliable, and whether 'LLM Voice' will fade as human language habits evolve all remain open questions.