The Rules Were Not Written By The Robot

[ This piece relies heavily on the interview in Freakonomics Radio episode 687 "Are Prediction Markets the Best Forecasting Tool Ever — or Just Another Casino?" - https://freakonomics.com/podcast/are-prediction-markets-the-best-forecasting-tool-ever-or-just-another-casino/ ]

On the day of the Super Bowl, Cardi B appeared during the halftime show for a clip that ran perhaps seven seconds. She was dancing. Her lips were moving. She was, as far as anyone could tell from the footage, not mic'd. Real money on Kalshi's prediction markets rode on a question the existing contract had not quite anticipated: did she perform?

The rule in force at the time said, roughly: if the person is dancing and singing, they're performing. Nicole Kagan, who runs research at Kalshi and personally writes a good share of its contracts, describes the resulting deliberation without embarrassment. Dancing: confirmed. Singing: "you absolutely cannot tell." Her account of what the team concluded is worth the exact words: "We don't feel comfortable resolving this to 'yes' or 'no,' because we can't tell. We're epistemically a little bit uncertain about this".

The head of research at a prediction market, asked on a podcast why a bet about a halftime show could not be settled, reaches for the word epistemically. The market did not resolve yes or no. It resolved to last fair price, the settlement Kalshi reaches for when a contract genuinely cannot be determined and the honest move is to say so rather than guess with confidence.

Then they rewrote the rule. The new version requires "audible lead or backing vocals". Under the new contract, Cardi B did not perform. The old contract could not have told you that in advance, because the old contract had not yet met a seven-second clip with moving lips and no microphone.

This is, structurally, case law

A fact pattern arrives that the existing rule didn't anticipate, a ruling is issued under protest of its own inadequacy, and the rule is amended so the next tribunal doesn't have to improvise. Kalshi does not call it that. It calls it a mentions-market rulebook, currently seven or eight pages long, which took over a month to draft. It covers what happens if the speaker mispronounces a word. Whether a misspoken word is still the word. What happens if it's a proper noun. What happens if the speech is livestreamed, the stream goes dark, and the recording surfaces somewhere else later.

Bloomberg, covering this in March under the headline Kalshi Faces Growing Problem With Grammar, Language Disputes, looked at eight pages of it and concluded it must have been generated by an AI. The theory has a certain logic to it: this is what mechanical, exhaustive, faintly inhuman precision is supposed to look like. Kagan's answer is worth quoting in full, because it inverts the assumption rather than merely denying it: "We wrote them for user clarity and didn't even consider the fact that if it's clear to a person, it's probably clear to a computer too".

Sit with that inversion...

The reporter's premise was that machine-readability produces this kind of precision, and human-readability produces something looser. Kagan's premise is the opposite, and it happens to be the older and better-tested one: a document written carefully enough for a human to stop arguing over it will, as a side effect, also parse cleanly for a machine. The clarity was never the artifact of automation. The clarity was the cause people mistook for automation.

The machine does get a turn - just not the one the reporter assumed

Reporter - Stephen J. Dubner of Freakonomics - to his credit, asks the obvious follow-up: do you at least run your contracts through an AI? Kagan says yes. They pass drafts through a language model "to see if there's anything that we've missed explicitly or anything that is a logical contradiction that we haven't caught". It is not, she adds, the primary thing they rely on.

That is a more interesting division of labour than either "written by a lawyer" or "written by a robot," and it is worth being precise about. The model is not the author. It is the second reader - the one you hand a draft to after you have stopped being able to see it. And what you ask it for is exactly the two failures a human drafter cannot reliably catch in their own work: the thing that isn't there, and the thing that contradicts the other thing on page six. The model has no idea which questions the contract ought to answer, but... It is quite good at noticing that an answer is missing.

Nobody sat down and reasoned their way to that arrangement from first principles about machine cognition. It is just what happens when someone who takes drafting seriously acquires a cheap proofreader who never gets bored.

Lawyers have known the underlying discipline for centuries, under a less flattering name

Nobody has ever read a well-drafted indemnity clause, a force majeure provision, or the definitions section of a syndicated loan agreement and called it elegant. Every "including but not limited to", every enumerated exception, every clause specifying what happens if a party is renamed, merges, or dies mid-contract, exists because someone sat down before the money moved and tried to list every way the deal could go sideways. Most of those clauses will never be tested. They are written anyway, because the cost of writing them is an afternoon of a drafter's time, and the cost of not writing them is a dispute nobody can resolve after the fact - which is precisely the position Kalshi was in for seven seconds of unamplified lip-sync.

Kalshi's rulebook and a loan agreement's boilerplate are the same discipline wearing different clothes. One prices the omission in basis points of legal fees. The other priced it, briefly, at whatever a few thousand dollars of Cardi B contracts happened to be worth that afternoon. Neither is decoration. Both are insurance against the version of the future nobody thought worth writing down - bought in advance, in billable hours, rather than litigated afterward, in court costs.

The market's other edge case makes the same point with sharper teeth

Kalshi's lifetime ban of George Santos - banned after he posted that he would attend President Trump's State of the Union address, bet on his own attendance both ways, and profited $17,000 when he didn't show - was not a rule amendment. It was a human judgment call, applied once, to a fact pattern the drafters hadn't written a clause for because nobody had thought to ask whether a disgraced congressman would manipulate a market about his own conduct. No line in the rulebook said do not trade on your own broken promises. The rulebook cannot get ahead of bad faith it has not yet met. Something else has to - and at Kalshi, in that moment, the something else was a person deciding a case, exactly the way a court does when a statute runs out before the facts do.

It is worth noting what Kalshi had written down in advance, because it is more than most. Under 17 CFR 40.11, the exchange may not list contracts on terrorism, assassination, war, gaming, activity unlawful under state or federal law, or anything the CFTC deems contrary to the public interest. Kagan says that absent the rule they would still decline those markets, on the grounds that an exchange should never create an incentive for somebody to bring about the outcome. Six categories, enumerated before the fact, precisely because somebody imagined what a market might make a person do. Santos is what the enumeration missed. [The rule itself is not finished being written: the CFTC proposed replacing this framework with a three-step analytical test on 10 June 2026, which was still pending when this piece went up. Even the enumeration gets amended.]

This is where the resemblance stops being cute and starts being useful

obliter spends most of its time on a version of the same problem one register up: AI systems, and the frameworks meant to govern them, described in language that sounds precise but has rarely been stress-tested against a Cardi B halftime show's worth of edge cases. Ask what counts as "high-risk" under the EU AI Act's Annex III categories, or what a company's internal AI usage policy says about an agent that does something nobody explicitly authorized but nobody explicitly forbade either, and the answer is very often: nobody wrote that clause, because nobody imagined the fact pattern that would need it. A prediction market spent a month deciding whether silence counts as singing. Most AI governance documents have not spent an afternoon deciding what counts as an agent acting on its own initiative.

The reporter's mistake was assuming that this level of anticipatory precision must have come from a machine, because no one expects it from a person anymore. The more interesting failure runs the other direction - governance documents that sound machine-generated in their vagueness, because no one made a person sit with the edge cases long enough to write them down. Those documents will also parse cleanly for a computer. There is simply nothing in them to parse.

The Clause has read a great many rulebooks. It has no particular loyalty to the ones written by lawyers over the ones written by exchanges, or to either over the ones a language model could, in principle, draft in an afternoon. It has only ever cared about one thing: whether somebody, anybody, decided in advance what performing means, before the money (or the liability) had already moved.