The reader revolt: audiences are tuning out LLM-written prose — and detecting it

Bryan Cantrill argues that readers can reliably spot LLM-authored writing and increasingly resent it. He points to a survey by Cynthia Dunlop of 668 developers: 78% stop reading the moment they detect an LLM, 71% will avoid that author in the future, and 98% would rather read a human’s imperfect prose than an LLM-polished version. The takeaway is that audiences aren’t demanding stylistic perfection — they want authenticity, and outsourcing the writing to a model breaks the implicit contract between writer and reader (the argument Cantrill formalized in Oxide’s RFD 576).

He frames the moment as analogous to the early-2000s email spam crisis: the tide turned once spam could be identified at scale, which wrecked its economics and made being flagged as spam reputationally toxic. Cantrill sees the same dynamic emerging for machine-written text. Naive detectors that key on surface tells (like the em-dash) are unreliable, but he credits Pangram Labs’ newer models — Pangram 3 and especially Pangram 4 — with very low false-positive and false-negative rates, making detection genuinely usable.

Acting on that, Cantrill extended RFD 576 to require that public Oxide writing register as human-authored under Pangram, and he urges other organizations that care about their institutional voice (calling out the Rust Foundation) to adopt similar standards. His bottom line for writers: expect your work to be run through a detector, and note that LLMs make excellent editors even if they make repellent authors. Beyond ethics, he contends that using an LLM to write public pieces is becoming self-defeating — it drives away the very readers and tastemakers a writer is trying to reach.