The Wolf Can Be Rented

Every bar has had the same argument. Somebody makes a confident claim, somebody else doubts it, and the conversation arrives at the oldest question in informal epistemology: want to bet?

What happens next is revealing. The person who knows the answer - who has actually read the thing, or was actually there - says yes without hesitating. The person who was only confident starts qualifying. The bet has not changed anything about the world. It has changed the price of being wrong, and a surprising number of opinions turn out not to survive the change.

That is the argument this piece makes, and it is worth stating plainly at the start: a bet is not a claim about the world. It is a claim about the bettor. Everything else follows from that, including why the law cannot see the difference between a bet and a statement of knowledge, why a track record is the only instrument that can, and why AI is about to rearrange who is making the claim.

Wolves, sheep, and the same contract

Market economists have a standard way of describing speculation. Some traders know something and trade to profit from it. Everyone else trades for other reasons - a hunch, a feeling-lucky state of mind, a bit of fun - and knows nothing in particular. The formal literature, following Kyle, calls the first group informed traders and the second noise traders. For the purposes of this piece it is easier to call them wolves and sheep, and the vocabulary is borrowed rather than technical.

Wolves want to trade against sheep, never against other wolves, because two informed traders cancel each other out and nobody profits. So wolves go wherever sheep gather. More sheep, more wolves, better prices. The model is old, simple and, as far as it goes, correct.

It also contains an implication that is rarely drawn out. The same contract is two different things depending on who holds it. A wolf buying a contract on next month's inflation figure is making a claim of knowledge and backing it. A sheep buying the identical contract, at the identical price, on the identical platform, is gambling. Nothing about the instrument distinguishes them. The difference lives entirely in the head of the person placing the trade.

Which is precisely where the law cannot look.

Why the law draws the line in the wrong place, correctly

The previous piece walked through how Croatia's gambling statute catches a prediction market without ever asking whether anyone involved is skilled. The definition reaches any outcome that depends predominantly on "chance or any other uncertain event", and the second limb makes the trader's research irrelevant. Most gambling statutes, one way or another, do the same thing. They regulate the shape of the arrangement (a stake, an uncertain event, a payout) and not the knowledge of the person entering into it.

This looks like a failure of legal imagination.

It is not.

A statute that tried to follow the wolf-and-sheep distinction - gambling if the trader is uninformed, information if the trader is informed - would need to answer, trade by trade, a question nobody can answer from the outside. The structural definition over-includes. It sweeps up the wolves along with the sheep, and occasionally a wheat future along with them. But the alternative is a definition that turns on a mental state, and mental states are what trials are for, not licensing regimes.

There is, however, one way to tell a wolf from a sheep that does not require reading minds. You wait, and you count.

A single bet tells you nothing about the bettor. Two hundred bets, each placed with a stated probability and each later resolved, tell you a great deal. The forecaster who says 70% and is right about seven times in ten is calibrated. The forecaster who says 90% and is right half the time is not a wolf, whatever that forecaster believes (and almost every forecaster does believe it). Nobody opens an account expecting to be the sheep. The feeling of defending your own knowledge is the one thing wolves and sheep have in perfect common.

So a bet becomes a claim of knowledge only retrospectively, and only in aggregate, through a track record. Before that, it is courage.

What a market actually does

It is worth being precise about what markets are for, because the usual answer - "they let people bet on things" - describes the mechanism and misses the function.

A prediction market does three things. It accumulates information: every participant brings a fragment of what they know. It filters it: fragments backed by knowledge tend to make money, fragments backed only by confidence tend to lose it, and over time the money follows the knowledge. And it synthesises it: the whole distributed argument is compressed into a single number, the price, which anyone can read without knowing who contributed what.

Newspapers accumulate and synthesise. So do analysts, research desks, policy briefings and, ever so more increasingly, AI assistants that summarise the day's developments before anyone has finished their coffee. What almost none of them do is the middle step. There is no mechanism by which the confident but wrong contributor pays for being wrong, and the careful but right one is rewarded. The difference between a market and a newsletter is not the quantity of information. It is consequence.

Seen this way, money is not the point of a prediction market. It is the filter: the most effective penalty for error anyone has found, but not the only conceivable one.

A scored, public, dated track record is another. It filters more slowly, and it costs nothing except the possibility of embarrassment. That turns out to be enough of a price for most people to start qualifying their claims.

Now add the machine

Everything above was true twenty years ago. What changes when the trader is, increasingly, not a person?

First: the wolf becomes a service. Wolves earn their edge by knowing things - by reading the filings, following the polls, noticing the regulatory notice before the price does. That is exactly the work that has just become cheap. An agent that reads every public source continuously, at negligible marginal cost, does not need to be brilliant to erase the advantage of a human who reads some of them occasionally. Any sheep with a subscription can now rent a wolf. And when everyone rents the same wolf, the wolves mostly trade against each other - the configuration in which, by the model's own logic, nobody profits.

Second: this moves the only edge worth having somewhere uncomfortable. If public information is priced in almost instantly by machines, what is left to profit from? Information that is not public. The candidate who knows how the count is going. The official who knows what the committee will decide. The person who helped plan the event on which everyone else is guessing. Defenders of prediction markets have long made an honest, uncomfortable argument here - one this site is reporting rather than endorsing: an institution that pays for information will pay most to whoever holds the most, and that is the institution working as designed, the same trade-off journalism makes every time it persuades someone to reveal what they promised to keep secret. The counter-argument, which seems to us the stronger one, is trust: a market everyone suspects is rigged is a market people stop using.

Put those two observations together and the argument sharpens. In a market where machines handle the public information, the marginal trade that still moves the price is disproportionately the one made by somebody who knows something they should not. AI does not create the insider problem. It clears away everything around it until the insider is most of what remains.

Third, and less obviously: the machine may be the first trader whose mind the law can read. Everything said above about the law being unable to see into the bettor was about people. An agent's "knowledge" at the moment of a trade is, in principle, a matter of record - what it retrieved, what it was instructed, what data it was built on, what it logged. The familiar surveillance pattern - a contract drifting at ten percent, a sudden wave of buying, an announcement the next day that moves it to forty - is, with a human trader, where an investigation starts and often where it stalls. With an agent that keeps logs, the question what did the trader know, and when? has an answer sitting on a server. The one test the law abandoned because it required reading minds becomes administrable exactly when the trader stops having one.

Whether anyone will require those logs to be kept, preserved and produced is another matter. That is not a question about prediction markets. It is a question about agent governance, and it will be decided (as these things usually are) by whoever reaches the drafting table first.

What cannot be priced

Enthusiasts sometimes suggest that markets could settle the questions politics cannot; that instead of arguing about whether climate change is a real threat, we could simply ask a market and stop relitigating it every election cycle.

The trouble is that the question contains two questions. Will average temperatures rise by a given amount by a given date? is a belief about the world, and a market can price it. Is that rise a threat worth reorganising an economy around? is a judgement about values, and no price will settle it. Hanson drew exactly this line more than a decade ago, in a paper whose title is its argument: Shall We Vote on Values, But Bet on Beliefs? A market needs a question with a date and a source that can resolve it. "Real or not" has neither.

What markets can do on polarised questions is smaller, and still real. Asking people to back their factual claims with something at stake tends to narrow the gap between what they profess and what they privately know. A market does not change minds. It reveals the distance between the opinion people perform and the one they are willing to pay for.

And for questions that have no resolution at all, there are mechanisms that do not need one. The economist Dražen Prelec has shown that you can often find the right answer to a question nobody can verify by asking two things: what is your answer? and what share of other people do you think will give the same one? The answer that turns out to be more common than the crowd predicted - the "surprisingly popular" answer - is frequently correct even when the majority is wrong. It works because those who know can anticipate the mistakes of those who do not, but not the other way round. Put simply, the wolf knows what the sheep think, and the sheep do not know what the wolf knows.

That asymmetry is also, one suspects, its limit. The method reads a crowd that does not know it is being read; a crowd told in advance that a question is a trap has every reason to start predicting the trap rather than answering the question, and the informational edge it depends on is the first thing to go. A market has the opposite property: when participants get warier, the price moves, and the informed keep correcting it until it is right. One is a photograph of a crowd. The other is a process that punishes error for as long as it runs.

That is the property worth borrowing.

The ledger

This site has now spent three pieces on prediction markets, and it would be a little rich to finish by arguing that the only honest test of a claim to knowledge is a dated, resolvable, scored track record.

And then not keep one...

So, starting with this piece, obliter.ai keeps one.

The ledger will hold every forecast this site makes on AI governance and data-protection questions: a question with a resolution date and a named source that settles it, a probability stated in advance, and the outcome recorded when it arrives. Entries are never edited after publication; a changed mind gets a new entry that links to the old one. In time the ledger will publish its Brier score - the standard measure of forecast accuracy, where 0 is perfect and 0.25 is what you get by answering "fifty-fifty" to everything. Not yet, though: with four entries a Brier score measures luck. It starts meaning something somewhere north of twenty.

No money is staked, which - for readers of the previous piece - is the one design choice that keeps the whole exercise safely outside Article 2.

The first four entries:

Three of those numbers deserve a sentence, because a forecast with no reasoning attached is just a number with a date on it.

The EDPB one is low because the Board's own recent record says so. Consultation on the web-scraping guidelines closes on 30 October 2026, which leaves five months. The blockchain guidelines (finalised at the very plenary that adopted this draft) took thirteen months from the close of their consultation. The DSA–GDPR guidelines took a little over twelve months from draft to final. Web scraping for generative AI will draw far more consultation responses than blockchain did, and the Board has a second major consultation, on anonymisation, closing the same day. Five months would be the fastest turnaround this Board has managed on anything comparable.

The Croatian one is high because the machinery already exists and runs at scale. The Tax Administration has issued work-ban orders against 528 foreign online gambling operators, with web addresses blocked within fifteen days of the order, and it publishes the list. Pending amendments would add automatic blocking and payment blocking. Eight European jurisdictions have already acted against these platforms. The open question is not whether Croatia can or would. It is whether anyone at the Tax Administration gets round to noticing, which is a question about attention rather than law.

The AI Act one is low despite enforcement having started, because starting is not the same as finishing. The Commission's powers over GPAI providers began on 2 August 2026, and the AI Office sent its first requests for information to OpenAI, Anthropic, Google and others on 29 August. A request for information is a long way upstream of a fine: the Commission must reach preliminary findings and give the provider a hearing first. Under the DMA, the gap between opening proceedings and the first fines was about thirteen months; and proceedings are themselves several steps beyond an RFI.

The full resolution criteria are on the ledger page, and they matter more than the numbers: for the first entry, whether "adopt" means the plenary vote or publication of the final text; for the third, whether a work-ban order without technical blocking counts. Two of the four are about prediction markets themselves. That is not a coincidence. It is simply the most natural place to start being wrong in public.

The Clause has no track record. It has never needed one. It defines the categories, and whether anyone inside them was actually right about anything has never once been its concern. That is, on reflection, the most reasonable thing about it: a definition that tried to separate the wolves from the sheep would need rewriting every time a sheep got lucky. The rest of us do not enjoy that exemption. The only method anyone has found for telling knowledge from confidence is to write the number down before the answer arrives. And then to stay in the room when it does.