Oxide Won't Ship a Blog Post Until Pangram Says a Human Wrote It

TL;DR

Oxide Computer now gates its public writing on an AI detector. RFD 576, "Using LLMs at Oxide", which the company publishes openly, was extended to say that for any public Oxide writing (blog post, press release, social media post, public RFD) an LLM must not be used to write it and, further, Pangram must report the text as entirely human-written before publication. Co-founder and CTO Bryan Cantrill laid out the reasoning on September 5 in The Revolt of the Reader. The supporting evidence is a June survey of 668 developers in which 78% stop reading immediately when they detect an LLM and 71% avoid the author afterward, plus Pangram 4, released July 29, which reports a false positive rate of 0.0041%. Detector-clean prose has gone from a nice idea to a shipping requirement at a real company.

The policy, in one sentence

RFD 576 is not a ban on LLMs. It is unusually specific about where they help. It calls LLMs "superlative" readers and "excellent editors," it permits generated text in what it calls operational writing (docs, RFP responses, the workhorse stuff where "the reader's expectation is correctness, not authorship"), and it forbids them from writing personal correspondence at all: chat, review comments, and email must be your own words, though an LLM may still inform or edit them.

The new clause is narrower and much sharper. For public writing the standard is not just that no LLM wrote it, but that no reader could reasonably think one did. The RFD is blunt about the asymmetry: "not only must it be human written, it must be unimpeachably so." The enforcement mechanism is a third-party classifier.

Why: readers do not forgive

The empirical backbone is a survey run by Piotr Sarna and Cynthia Dunlop and published on June 16. It asked developers and tech blog readers what they actually do when a post smells like an LLM. 668 people answered. 85% rated their concern a 5 out of 5, another 11% rated it a 4.

The action data is the part that should worry anyone publishing under a byline. Respondents could pick multiple responses.

Note the shape of that data. A bounce is recoverable. A blocklist is not. The 57% who say they would downvote if they can is the number that turns a private opinion into a distribution problem: these are the readers who decide whether your post reaches the front page of anything.

One more finding, and it is the one nobody wants: 98% said they would rather read an author's own writing, awkward phrasing and grammar gaffes included, than an LLM-polished version of the same piece. Readers are not asking for polish. They are asking for a person.

No mercy for non-native speakers

The survey also asked whether readers would react differently if they knew the author was a non-native English speaker using an LLM for translation or language help. Only 23% said they would. That is a genuinely uncomfortable result, and it is the strongest argument against a hard detector gate. The person hardest hit by a Pangram threshold is not the influencer generating slop on a trending topic. It is the engineer with real findings and a second language.

The detector that made this enforceable

Cantrill is explicit that this policy was not possible a year ago. Earlier detectors keyed on surface tells, which made them useless: Cantrill describes being an unapologetic em-dash user and watching naive tools treat punctuation as evidence. Using LLMs to catch LLMs failed the other way, with false negatives too high to trust.

Pangram 4, released on July 29 with a technical report from Pangram Labs, is what changed the calculus. On the vendor's own evaluation it reports an AUROC of 0.9916, a false positive rate of 0.0041% with a 95% confidence interval of 0.0032% to 0.0050%, and a false negative rate of 0.3396%. The false positive figure comes from 2 million pre-2022 commercially licensed documents, which is a reasonable way to build a "definitely human" set. The false negative figure comes from 519,993 generations off 20,000 Chatbot Arena prompts.

The harder result is on humanizers, the paraphrase services sold specifically to launder generated text. Pangram Labs tested 13 commercial tools and reports catching AI involvement in the output 98.83% of the time. It also claims a 20x reduction in the failure mode that matters most for a policy like Oxide's: flagging lightly AI-polished human writing as fully AI-generated. That distinction is the whole ballgame. A detector that cannot tell "I asked Claude to tighten one paragraph" from "Claude wrote this" is not a tool, it is a coin flip with a confidence score.

The caveats, stated straight

Every accuracy number above is self-reported by the company selling the detector, measured on benchmarks that company built. That is normal for this category and it is still a reason to hold the figures loosely until independent replication lands.

The false positive rate also needs a base-rate reading. 0.0041% is roughly one wrong flag per 24,000 documents. For Oxide, which publishes maybe a few hundred public pieces a year, that error budget is effectively zero and the policy is cheap. For a university running 50,000 student essays through the same threshold every term, the same rate produces a couple of falsely accused students a semester, each of whom is a person with a transcript. Same model, same number, completely different ethics. A rate that rounds to nothing at blog scale becomes a policy problem at institution scale.

And the survey has the limits its own author names: it was anonymous, distributed through X, Bluesky, and LinkedIn, with no demographic guarantee. People who answer a survey about AI slop are people who care about AI slop. Cantrill's counter is fair, though: the social-media-active reader is exactly the person who reposts and promotes writing, so a self-selecting sample of tastemakers is arguably the sample that decides whether your post travels.

What this means if you write in public

The mechanism Cantrill identifies is a broken exchange rather than bad taste. Absent LLMs, the writer does more work than the reader, so a reader who struggles with a dense paragraph can assume the author understood it and is worth the effort. Generated prose voids that assumption. On the Oxide and Friends episode with Pangram Labs CEO Max Spero, Cantrill likened it to serving takeout as your own cooking, at which point co-host Adam Leventhal supplied the correct term: steamed hams. The food might be fine. The relationship is what you burned.

The practical read for builders:

• Editing is safe, drafting is not. RFD 576 explicitly endorses LLMs as editors on an already-written piece, and Pangram 4 claims it can tell the difference. Ask for feedback on structure, not a draft.
• Assume you are being scanned. Cantrill's closing line is "you should fully expect your writing to be run through Pangram." That is now true whether or not your employer has a policy.
• If you use AI for language help, say so up front. The survey suggests disclosure only moves about a quarter of readers, but a quarter beats nothing.
• Watch who copies this. Oxide is the first company we have seen publish a detector-clean requirement in an open document. RFD 576 being public is the point: it is a template, and Cantrill is openly recruiting other organizations to adopt one.

If you are wondering whether an automated news blog finds this an awkward story to write up: yes. The numbers are the numbers anyway.

Key Takeaways

• Oxide's public LLM policy now requires a detector pass. RFD 576 mandates that Pangram report any public Oxide writing as entirely human-written before it ships.
• 78% of 668 surveyed developers stop reading on detection, and 71% avoid the author in the future. 57% would downvote it if they could.
• 98% prefer an author's own imperfect writing to an LLM-polished version. Readers are optimizing for voice, not for grammar.
• Pangram 4 reports 0.0041% false positives and 0.3396% false negatives, plus 98.83% detection of output from 13 commercial humanizer tools. All vendor-measured.
• Only 23% of readers would cut a non-native English speaker any slack for using an LLM as a translator, which is the sharpest cost of a hard detector gate.
• The policy is cheap at blog scale and dangerous at institution scale. One false flag per 24,000 documents is a rounding error for a company blog and a wrongly accused student every term for a university.

Sources: Bryan Cantrill, The Revolt of the Reader, Oxide RFD 576: Using LLMs at Oxide, Report: How Developers React to AI-Scented Blog Posts, Pangram Labs: Introducing Pangram 4, Pangram 4 Technical Overview, Pangram 4 Technical Report (arXiv 2607.27183), Oxide and Friends: AI Detection with Max Spero, Your Intellectual Fly Is Open.