Oxide requires posts pass an AI-detector before publishing

Oxide changed its public writing policy on September 5, requiring every post pass Pangram's AI detector as human before it publishes, per Bryan Cantrill's essay. The rule doesn't ban using an LLM. It requires a passing score, so a human sentence that trips the detector can't run either. That's a stricter bar than most AI-content policies, and Oxide adopted it after a survey of 668 developers reacting to LLM prose.

Cynthia Dunlop's survey found 78% of developers stop reading the moment they detect LLM authorship. 71% avoid that writer afterward. 98% preferred an author's imperfect prose over an LLM-polished version of the same piece. Dunlop's 668 respondents opted into a study about LLM writing, a group already primed to care, so 78% is likely a ceiling, not an average.

Oxide says it accepted the false-positive risk because Pangram's rate has been low in its own testing. Cantrill hasn't published that sample, so there's no way to check the claim independently yet.

Cantrill frames the shift as economics, not ethics. He points to email spam, where filtering made mass mail unprofitable well before senders developed scruples, and argues reader filtering will do the same to polished LLM prose.

I write daily and use Claude Code every day in my personal projects, so this is personal for me. Cantrill's data suggests the research a model does for you doesn't show up in the prose. Smoothing your sentences into an even, frictionless rhythm trips the detector and costs you readers. My own unedited drafts run too uniform in sentence length. Breaking that up is something I do on purpose, every time.