The revolt of the reader: why LLM-written prose loses its audience

Engineer Bryan Cantrill, one of the authors of DTrace and a co-founder of Oxide Computer, has published an essay on his blog arguing that readers are in open revolt against text written by large language models. In his account, LLM-authored prose is easy to spot, and the practice breaks what he calls the social contract between writer and reader.

The essay describes the structural tells of LLM writing as jarring enough that readers bail out of a piece mid-sentence. Cantrill points to a survey by Cynthia Dunlop, in which 668 developers described how they react when they encounter AI-generated blog posts.

What the survey found

According to the figures Cantrill cites, 78% of respondents stop reading immediately once they detect an LLM, and 71% go further, avoiding the author in the future. He stresses that readers are not chasing linguistic perfection: 98% said they preferred an author's own, imperfectly written text over one polished by a model.

Cantrill warns against dismissing those respondents as a self-selected group. Active readers on social media, he notes, are exactly the people most likely to share writing they like — the early adopters and tastemakers of online prose.

Pangram and the Oxide policy

Earlier attempts at detection, he writes, were unreliable: models used as detectors missed too much, and checking superficial tell-tales proved too broad to be useful. In his assessment, Pangram Labs' Pangram 3, launched late last year, was a major improvement, while Pangram 4, introduced about a month ago, is a step-function better still, with low false positive and low false negative rates.

Oxide has extended its internal RFD 576 so that public writing issued under the company's banner must be reported by Pangram as human-authored. Cantrill encourages organisations that value an authentic institutional voice to adopt a similar policy.

The spam parallel

Cantrill compares LLM-authored text with email spam. In the early 2000s, he recalls, there were real fears that spam would destroy email; by the late 2000s filtering had improved enough to undermine the economics of spam, and being labelled as spam became genuinely damaging for legitimate businesses.

He argues the same dynamic may now apply to AI-written prose: if readers choose to ignore an author, that author has undermined the entire purpose of publishing. His advice is blunt — use an LLM as an editor, not as the writer, because a prompt is not an article.