LLM Voice

Martin Fowler's "LLM Voice" bliki post delves into the pervasive and often unsettling experience of encountering text that feels, unmistakably, like it was written by a Large Language Model. He describes this sensation as an "uncanny valley" of prose, where the writing is grammatically correct but ultimately lacks a human soul, leading to a sense of revulsion rather than engagement.

• Fowler admits the "tells" of LLM writing are hard to pinpoint precisely, often being a gut feeling rather than specific stylistic quirks. He expresses skepticism about AI detection tools, arguing that by the time one might use such a tool, the human disengagement has already occurred.
• A key warning from Fowler is against using LLMs even for initial drafts, as expelling the inherent "LLM Voice" proves extremely difficult, even for professional editors. He suggests AI might be acceptable for outlines but emphasizes authors must write the prose themselves.
• He passionately defends writing as a crucial tool for thinking, asserting that delegating writing to AI risks delegating comprehension and the very joy of intellectual wrestling with ideas.
• Fowler acknowledges legitimate exceptions, such as writers using AI for accessibility reasons (e.g., Parkinson's disease), translation, or generating content for bureaucratic processes that are less about human connection and more about process fulfillment.
• On his own site, Fowler maintains a policy of disclosing AI use, and personally, he uses AI only as a search tool, never for generating prose or structure.

Ultimately, Fowler concludes that for communication intended for humans, the prose must radiate (or at least ooze) humanity. He posits that while future LLMs might become indistinguishable from human writers, that outcome presents its own set of worrying implications for how we think and connect.