On September 17, Thomas and Erin Ptacek published “How to Write with an LLM” on sockpuppet.org. By the weekend it had 634 points and 375 comments on Hacker News, which is roughly what canonization looks like in this business. The advice is two rules: never keep a single word the model suggests, and ignore the model’s praise for your draft. Commenters noted within hours that the essay’s own text uses “load-bearing” — a phrase that has become the fingerprint of prose something mechanical has handled.

It is a good essay. It is also aimed at the wrong target.

The Docket Is Not Complaining About Style

Consider what else happened this year. According to EDRM’s tally, U.S. courts imposed at least $145,000 in sanctions in the first quarter of 2026 for fabricated citations — a figure that includes a record-setting Oregon penalty and the first substantial federal appellate fine tied to AI-tainted briefs. Norton Rose Fulbright’s 2026 litigation update describes the newer, nastier failure mode: not invented cases but fabricated quotations attributed to two real Sixth Circuit decisions, plus flat misstatements of what those cases held.

In that appellate matter, the court reportedly identified the brief as AI-drafted because of the file name, which indicated the platform that produced it. Read that again. The tell was metadata. The stylistic tells everyone in the comment thread is training themselves to hear are unreliable; the file name told the truth.

Nobody on the bench is grading prose. A judge does not care whether your brief is 10 percent flatter than the last one. A judge cares that a citation resolves to a real holding — and when it doesn’t, the opposing party has spent real hours researching a case that does not exist.

Note also what the sanctions actually punish. No court has sanctioned a lawyer merely for disclosing that a model was involved; liability attaches to fabricated or unverified authority, under Rule 11 and Model Rule 3.3, and it attaches to the signature rather than to the tool. Which means “the AI did it” was never available as a defense, and everyone drafting these briefs knew it.

A Rule You Can’t Audit Only Binds the Already Careful

Ptacek’s first rule requires an honest self-audit at the exact moment of temptation, usually 11 p.m., usually under deadline. That is a real discipline. It is also, by construction, a discipline observed only by people who did not need it.

The incentive structure is not subtle. A third-year associate, typing into a firm-wide channel at 11:40 on a Tuesday, put it plainly enough: “The partner wants three drafts by morning. Nobody’s reading the cases at that hour — the model writes and we spot-check.” That is not cynicism. That is a production schedule.

You do not bill by the sentence. Throughput is the metric, and the model is a throughput machine. So here is the part of the taste argument that its authors keep skipping: if AI-assisted prose is 10 percent worse and 40 percent cheaper to produce, the market clears at flatter. It clears politely, at scale, and without asking anyone’s permission. The aesthetic objection loses that fight on arithmetic. It was always going to.

Disclosure Is the Cheapest Rule in America, Which Is Why Everyone Resents It

Several courts now require disclosure of AI use in filings. The compliance cost is one line in a document that already exists. Set that against $145,000 in quarterly sanctions, a district judge’s docket time, and the hours a defense team burned chasing a hallucinated opinion through a database that has never heard of it.

Disclosure does not prevent the error. It changes who bears the risk of the error, which is the only thing any liability regime has ever done. That is modest, and it is measurable, and it is more than the craft advice can claim.

What is interesting is the politics. People who distrust these systems want them labeled. People who trust them treat labeling as a paperwork tax engineered for shaming. Both camps are half right, and neither has noticed the mechanism: this rule is arriving through individual judges and bar committees, decentralized, personally enforced, trivial to comply with and expensive to ignore. That is the shape of regulation people who dislike the administrative state claim to prefer. Mostly they don’t, when it shows up wearing a robe.

What Readers Were Actually Buying

Here is the uncomfortable part for the writers, and for the people who spent last week nodding along with them.

The complaint about homogenized AI prose is, in most cases, a labor complaint wearing a beret. That does not make it wrong, but it makes it weak, because no one is owed a market for their sentences. The objection that survives contact with a docket is the one about verification.

And that objection implicates something the writerly class would rather not say out loud: what readers were paying for was never the voice. It was the guarantee that a person had checked. The voice was the packaging. The guarantee was the product, and it is precisely the part that got automated away — by people who never learned to write well and never needed to.

Ptacek’s first rule is sound advice for anyone who wants to remain readable. Just don’t mistake it for a policy. One is a habit. The other produces a $145,000 invoice, an appellate opinion, and a judge who reads the file name.

Sources