Survey finds 78 percent of developers stop reading an article they believe is AI-assisted or AI-written

Sources: primary, discussion
Summary: Colin Breck, writing 2026-09-20, cites Cynthia Dunlop's survey on AI-scented blog posts: 78 percent of readers stop reading when they believe an article is AI-assisted or AI-authored, 71 percent avoid the author in future, and 98 percent prefer the author's own writing with its flaws over an AI rewrite. His stated mechanism for why AI-generated documents read badly to everyone but their prompter is that the prompter holds the context that produced the output and can skim it, while the recipient has none and must read exhaustively. He reports the inverse split from writing an academic paper, where AI was valuable for verifying claims against source code, configuration, and production logs, for citations, for TikZ diagrams, and for catching a notation error four expert reviewers missed, while asking it to write a paragraph from that same context was never useful.
Why it matters: The figures put a measured readership cost on publishing AI-written engineering prose, and Oxide now mandates that all public writing be reported human-authored by Pangram.