This essay does something most AI-writing critiques don't: it splits the problem cleanly in two. The complaint isn't that AI exists or that using it is cheating. The complaint is that people are shipping AI output as communication and expecting other humans to do the cognitive work the writer skipped. Design documents that summarize already-built systems. Pull request descriptions that inventory changes without explaining why. Personal messages workshopped into blandness. The pattern is consistent — the writer had context, the reader doesn't, and the AI can't bridge that gap. The strongest section is the analysis of context asymmetry. When you prompt an AI, you hold the constraints, the source material, the intent. You can skim the output and spot what's wrong because you already know what's right. Send that same output to someone else and they're reading a wall of plausible-sounding text with no map. The essay names this precisely: 'statements outside of context.' That's the diagnosis, and it's correct. The positive case — AI as verification engine, citation assistant, grammar cop, diagram generator — is grounded in a specific experience writing an academic paper. The author describes asking AI to check paragraphs against source code and production logs while continuing to write. The reversal is telling: asking AI to write paragraphs from the same context produced consistently worse results. The one thing AI wrote well was the abstract — the most mechanical, context-free part of the paper. The essay draws on Cynthia Dunlop's survey data (78% of readers stop reading AI-detected articles, 71% avoid the author afterward) and Bryan Cantrill's work on Pangram, an AI-detection model now mandated for public writing at Oxide. These aren't abstract worries — they're behavioral data showing readers are already revolting against AI prose. The Bjarne Stroustrup quote about hearing an author's voice is well-placed and earns its spot. Two forward-looking threads get space: ASD-STE100 Simplified Technical English, an aerospace-origin standard for controlled vocabulary now adapted for AI models, and Pangram as an enforcement mechanism. Neither is presented as a solution — more as experiments worth tracking. The essay is honest about the limits of both. The closing move — writing as vulnerability, as relationship, as something that can't be optimized without being destroyed — lands because it follows specific craft observations rather than leading with sentiment. Simon Sarris's 'Resist Summary' and Murat Demirbas's observation that LLMs 'lack an active mental model of a specific human reader' provide the intellectual scaffolding. The essay practices what it preaches: it's clearly written by a person with opinions and idiosyncrasies, not generated. What holds it back is scope. This is a practitioner's essay, not a systematic argument. The academic paper anecdote is one data point. The survey data is self-reported. The essay doesn't engage with counterarguments from people who genuinely write better with AI assistance (non-native speakers, people with disabilities, etc.). It's a well-argued position paper from one professional's vantage point, and it knows it.
