The Uncanny Valley of LLM-Generated Prose
Why delegating writing to AI risks losing the human connection and critical thinking.
1Recognizing the LLM Voice
Readers increasingly detect a distinct, artificial tone in content, often described as the uncanny valley of prose. While automated detection tools exist, they are unreliable and often miss the point. The visceral reaction to this synthetic style suggests that human readers instinctively crave authentic, biological expression rather than polished, machine-generated output.
2The Trap of AI Drafting
Authors should avoid using LLMs to generate initial drafts, even if human editing follows later. The synthetic voice is notoriously difficult to fully remove once established in the text. Even accepting minor rephrasing suggestions from AI can compromise the unique human quality that readers value in professional and personal communication.
3Writing as Cognitive Process
Writing is not just output; it is a fundamental tool for thinking and achieving coherence. By wrestling with words, authors overcome confusion and clarify complex ideas. Delegating this process to an LLM risks outsourcing the actual thinking, effectively removing the intellectual rigor that makes the act of writing personally rewarding.
4Exceptions for Accessibility
The rule against LLM-generated prose is not absolute. For individuals like Jim Highsmith, who battles Parkinson's disease, AI tools are essential for maintaining the ability to share wisdom. In such cases, the value of the content justifies the presence of the synthetic voice, prioritizing accessibility over stylistic purity for the reader.
5Translation and Language Barriers
For non-native English speakers, LLM translation serves as a vital bridge for communication. While some prefer the clunky but authentic prose of a human, AI-assisted translation enables collaboration across linguistic divides. It remains a practical trade-off, balancing the need for clear understanding with the desire for a genuine human voice.
6Contextual Use of Automation
Using AI for low-stakes tasks, such as drafting form letters to officials or reports unlikely to be read by humans, is reasonable. When communicating with a process rather than a person, the human element is less critical. Future documentation might be written primarily for LLMs, rendering the uncanny valley concern irrelevant.
7Transparency and Best Practices
Authors using AI for structural outlines or research should declare their methods in acknowledgments. This promotes transparency and helps readers understand how tools are being utilized. For the author, AI should ideally function only as a search tool, with all retrieved information verified by returning directly to the original source.
8Preserving Human Connection
Ultimately, writing remains a skill that requires consistent practice to master. Surrendering the craft to AI undermines long-term progress and diminishes the author's voice. To effectively communicate with other humans, authors must ensure their humanity shines through their prose, resisting the temptation to delegate the essential act of thinking.
Original article: Martin Fowler — The Uncanny Valley of LLM-Generated Prose