The Number Arrived Without a Denominator
The number arrived without a denominator.
A wire story crossed my desk โ the kind that takes ninety seconds to read and three hours to unlearn โ announcing that a prediction market had doubled its implied odds on the passage of an American "AI safety bill." The figure was 30%. That was the entire payload: one probability, one multiplier, one source, and a legislative noun so elastic it could have described any of a dozen unrelated proposals currently drifting through committee rooms in Washington.
I have spent the better part of twenty-nine years in this industry watching numbers detach from the things they claim to measure, and I have learned to read the absence as carefully as the presence. So I did what I always do when a statistic arrives dressed as a fact. I went looking for what wasn't there.
There was no bill number. No market identifier. No volume figure, no open interest, no bid-ask spread, no resolution criteria, no sampling methodology, no named researcher, no second source. Just 30%, doubled, presented to a general audience as though it were a reading on a thermometer rather than a price on a ledger.
Chaos is data in disguise. The missing pieces were not omissions. They were the story.
What follows is not a piece about artificial intelligence legislation, and it is only incidentally a piece about prediction markets. It is a piece about a specific, repeatable failure mode in how financial instruments get laundered into journalism โ and about why the laundering is accelerating precisely at the moment the instruments are becoming most trusted.
What a Prediction Market Is, and What It Quietly Is Not
To understand why a bare 30% should trigger an auditor's reflex rather than a reader's acceptance, you have to understand the instrument itself, and then understand the gap between the instrument's design and the way it is quoted.
A prediction market is a venue where participants trade contracts that pay out one unit if a defined event occurs and zero if it does not. Because a contract paying 1 on success and 0 on failure should, in a well-behaved market, trade at a price equal to the crowd's implied probability of that event, the price becomes a live, continuously updated forecast. This is the elegant part, and it is the part that gets quoted in headlines.
Polymarket, built on Polygon and settled through UMA's optimistic oracle, is the largest permissionless venue of this kind. Kalshi is its regulated American counterpart: a CFTC-designated contract market that spent years litigating for the right to list event contracts on political outcomes, won a meaningful round in 2024, and then watched the question migrate back into the appeals process. Between them they now constitute something that did not exist a decade ago โ a real-time, money-weighted polling apparatus that updates faster than any survey and, on its best days, forecasts better than most of them.
The elegance has a cost, and the cost is structural. A prediction market does not output a probability the way a scale outputs a weight. It outputs a clearing price, and a clearing price is a function of two inputs: aggregate belief and aggregate liquidity. Headlines quote the first and ignore the second. That is roughly equivalent to reporting a stock's price while neglecting to mention that nine shares changed hands, and that the ninth share was bought by the person who set the price.
Then there is the entity the price is supposed to describe. "AI safety bill" is not a bill. In the United States, AI regulation in this cycle has been a fragmented, contested, multi-front affair: state-level statutes with mutually incompatible definitions of "high-risk systems," federal executive orders that were signed and then rescinded, voluntary framework documents carrying no enforcement mechanism, and a rotating cast of proposed bills that die quietly in committee without a floor vote. The European Union's AI Act, by contrast, is a single instrument with a phased timetable and a published compliance calendar โ which is precisely what makes it tradable. American AI legislation is not one event. It is a category error wearing legislative clothing.
That distinction โ one instrument versus one category โ is where the 30% begins to come apart. And it is where the analysis has to begin, because a probability attached to an undefined event is not a weak signal. It is a signal of a different kind entirely, one that measures the confidence of the people quoting it rather than the likelihood of the thing being quoted.
A Price Is Not a Probability
The most common error I encounter when institutions first approach these markets is the assumption that price equals probability. It is an easy assumption to make, because the instrument's documentation encourages it and because the arithmetic โ pay 1, receive 1, so price equals chance โ is superficially clean.
It is clean only under conditions that almost never hold simultaneously. The equivalence requires deep liquidity, negligible transaction costs, participants who are risk-neutral, no hedging demand, no capital constraints, and a resolution mechanism so unambiguous that no reasonable trader could disagree about what the contract promised. Remove any one of those and the price drifts away from the probability. Remove three or four โ as thin policy markets routinely do โ and the two numbers can be separated by more than the entire range of plausible outcomes.
I first internalized this viscerally in 2020, during the DeFi summer, when I spent weeks buried in the liquidation mechanics of early Aave and Compound forks. What I found was not a flaw in the math but a flaw in the assumption. The protocols assumed that a price feed reflected value. It did not. It reflected the last trade, executed by whoever happened to be trading, at whatever depth the book happened to have that hour. Efficiency, I learned, is not the same thing as accuracy โ and in thin markets the two can be actively opposed. I have carried that lesson into every market I have touched since, including this one.
In a thin policy market, a 30% price does not mean thirty percent of the market believes the bill will pass. It means that, at the last moment someone bothered to trade, the marginal buyer and the marginal seller agreed on 30 cents. Beneath that number can sit any composition you like: a genuine 30% consensus, or a 4% consensus overlaid with a single enthusiast's conviction, or a 50% consensus that has been dragged downward by somebody hedging an unrelated position. The price cannot tell you which. Only the liquidity can.
This is why I keep returning to a sentence I wrote for myself years ago and have never found a reason to amend: follow the liquidity, ignore the hype. It is not cynicism. It is the minimum standard of care. A probability without a denominator is not a forecast. It is an opinion wearing a decimal point.
The Denominator Nobody Quotes
Let me be concrete about what a denominator does, because abstraction is how these errors survive.
When an order book is deep, moving the price requires capital proportional to the depth. To push a market from 15% to 30% against a book with a hundred thousand dollars of resting liquidity, you must be willing to buy through a substantial stack of offers, and you must accept the slippage that comes with it. That is an expensive opinion, and expensive opinions tend to be informed, or at least committed.
When an order book is shallow, the arithmetic inverts. Against five thousand dollars of resting liquidity, a two-thousand-dollar order can move the quoted price by fifteen percentage points. Against a book with a few hundred dollars on each side, a single wallet can produce a headline. The same number โ 30% โ can therefore represent a market's considered judgment or a single trader's Tuesday afternoon. Without the volume figure, the two are indistinguishable to the reader. And the reader, in this case, is being handed the number precisely as though the distinction did not exist.
I have a specific habit when I evaluate these markets for the fund, and it comes directly out of the 2017 whitepaper audits โ the long, isolating months I spent reading through more than fifty project promises, cataloging the gap between the rhetoric and the engineering, identifying ten clearly fraudulent tokenomics before the bubble burst. That habit is this: I do not read the claim. I read the conditions under which the claim could be falsified, and then I check whether anyone has bothered to state them. A market that cannot tell you its volume is a market that cannot tell you whether it is a market. It is a number generator with a URL.
The "doubling" framing makes this worse, not better. Doubling sounds enormous โ it carries the emotional weight of a regime change. But if the prior was 15%, the doubling is fifteen percentage points. If the prior was 3%, it is three points, and the headline is simply a rounding artifact given a verb. The base rate matters, the absolute move matters, the volume behind the move matters, and the resolution criteria matter. The multiplier matters least of all. Yet the multiplier is the only component the headline preserved.
The Load-Bearing Wall Is the Resolution Criteria
Every prediction market rests on a single load-bearing wall, and that wall is the resolution criteria โ the rules that determine what the contract actually promises and who decides whether it happened.
On Polymarket, resolution typically runs through UMA's optimistic oracle: an outcome is proposed, and unless someone disputes it with a bond, the proposal stands. This design is elegant and it is also fragile in a very specific way. It works beautifully when the underlying question is crisp โ did this candidate win, did this asset close above this price. It becomes contested and expensive when the question is interpretive. And "did an AI safety bill pass" is interpretive at every joint: Which bill? Passed by which chamber? Signed, vetoed, overridden? Does an executive order count? Does a state statute count? Does a funding provision attached to an unrelated appropriations bill count?
When I audited the collapsed balance sheets of Terra and FTX in 2022 โ months of solitude, reading not for the numbers but for the ethical failures that produced them โ the recurring theme was never a broken equation. It was a definition that had been quietly stretched until it meant whatever the person in control needed it to mean. "Stable." "Collateralized." "Segregated." Each word was load-bearing. Each word was redefined in private until the public version was a formality.
Prediction markets carry the same exposure, and they carry it more acutely because their entire output is a single interpretive number. If the resolution criteria for this AI market are vague enough to admit multiple readings, then the 30% is not a probability at all. It is the market's collective shrug about a question nobody has precisely asked. The algorithm has no conscience. It will happily price ambiguity, because pricing ambiguity is what it does. It has no mechanism for refusing a badly posed question.
So the first thing I would demand from anyone citing this number is the market's resolution text, verbatim. Not the headline's paraphrase. The actual rules. If the rules cannot be produced, the number should not be cited, because nobody โ including the person who quoted it โ can say what it refers to.
The Long-Shot Bias and the Calibration Problem
There is a deeper empirical problem, and it predates crypto by decades.
Prediction markets, like the racetracks and lotteries that preceded them, exhibit a well-documented tendency to systematically overprice improbable outcomes. This is the long-shot bias, and it is remarkably persistent across venues, asset classes, and decades. Low-probability contracts tend to trade rich โ that is, the price sits above the realized frequency. High-probability contracts tend to trade cheap. The bias is not a bug in a particular platform. It is a feature of human beings, who overweight vivid, narrative-driven, imaginable outcomes relative to their base rates.
Now put that bias next to the base rate for American legislation. The overwhelming majority of bills introduced in the United States Congress never become law. Even restricting the pool to "major" legislation with meaningful sponsorship and committee traction, the historical pass rate is measured in single digits to low double digits, and it collapses further for anything touching a contested policy domain. Comprehensive AI safety regulation, in the current political configuration, sits squarely inside the contested zone.
None of this proves the 30% is wrong. It does something more useful: it tells me what the 30% would have to be made of to be right. For a market to hold 30% on a genuinely ambiguous, low-base-rate legislative event, it needs either (a) hard, specific news of procedural progress โ a committee markup, a leadership commitment, a floor scheduled โ or (b) a strong emotional driver pulling capital in one direction. The brief that reached me contained no evidence of (a) and a suggestive trace of (b): a background sentence about researchers warning of risk. That is a sentiment input, not a procedural input. It tells you what people are thinking about, not what the legislature is doing.
Which means the move from 15% to 30% is most parsimoniously explained as a sentiment repricing on thin liquidity, amplified by a narrative multiplier, transmitted to a general audience as a fact. That is not a market signal. That is a market rumor with a settlement layer.
Framing: Doubling Is Not Moving
Behavioral economists have a name for what the headline did: the framing effect. The same underlying change โ a fifteen percentage point move โ reads as a transformation when described as a doubling and as a modest adjustment when described in absolute terms. The reader's response is driven by the frame, not the datum.
I have watched this happen in crypto for nine years. In 2021, during the NFT explosion, I funded three small artist-centric DAOs โ not for return, but to understand whether on-chain governance could actually build community. What I learned was uncomfortable and it rhymes with this: the framing did the work, and the substance never had to. A floor price "up 400%" from a near-zero base, a "record" volume that reflected two wash trades, a "community" that was one Discord and a shared conviction. The numbers were all technically true and collectively meaningless. Meaninglessness scaled beautifully. That is the whole lesson of that cycle, and it is being replayed here at the level of policy journalism.
When I see "doubled," my first question is always the same: doubled from what, on what volume, resolved by whom? If the answer to all three is unavailable, the correct response is not skepticism mixed with interest. It is to decline the number entirely and ask why it was published.
The Niche Shift: From Casino to Citation
Here is the part that actually matters, and it is the part the brief never surfaced because the brief did not consider it newsworthy.
Prediction markets are undergoing an ecological niche shift. They are migrating from "crypto-native speculative venue" to "cited information infrastructure." Journalists are now pulling odds the way they once pulled polls โ as a forward-looking, quasi-quantitative input into the story. The venue is no longer the subject. It is the instrument through which the subject is observed.
This is the most important development in the sector, and it is happening with almost no scrutiny attached. The dependency is now bilateral. Media depends on prediction markets for a live read on expected outcomes. Prediction markets depend on media for traffic, legitimacy, and the soft political cover that comes with being treated as a public utility rather than a betting parlor. Each citation is a small deposit into the venue's credibility account.
But credibility deposits made without verification are also liabilities, and they compound the same way leverage does. A prediction market's public status rests entirely on the integrity of the two things I already flagged: resolution clarity and liquidity depth. The moment a widely cited market is shown to have been moved by a single wallet, or resolved against its own posted criteria, the niche collapses โ not gradually, but in a single news cycle. The instrument does not degrade. It is simply withdrawn from the citation pool, and the media goes back to sourcing its expectations from people with titles.
The brief that reached me contained none of this analysis, because a low-density news item does not contain analysis. But the item's existence is itself the evidence. Somebody, at some outlet, decided that a floating price on a crypto venue was a publishable surrogate for a legislative forecast. That decision โ not the 30% โ is the thing worth examining. The number will be forgotten by Thursday. The norm it established will not.
The Regulatory Irony of the Source
I would be failing my own standard if I did not flag the tension sitting under the entire exercise.
The venue supplying the headline number is one that, in the recent past, paid a seven-figure penalty to the Commodity Futures Trading Commission for offering event-based contracts to American users without the registration the agency deemed required, and then walled off the United States in response. The precise legal posture has shifted with subsequent litigation โ Kalshi's successful challenge to the commission's refusal to list election contracts reopened questions the agency had considered settled โ but the underlying dispute has never been resolved at root. Whether an event contract is a swap, a commodity, or a wager remains contested between the federal commodities regulator and state gaming authorities, and the answer determines which body, if any, has jurisdiction.
So the situation is this: a headline about American legislative prospects is being sourced from a market whose own American legality is an open question. I am not making a moral argument here. I am making an evidentiary one. When a source's legal status is itself unsettled, the appropriate response is heightened scrutiny of the source's output, not relaxed scrutiny. The irony does not invalidate the number. It raises the bar the number must clear, and by every measure available to me, it cleared none of it.
The Decoupling Thesis, Inverted
The prevailing narrative in the sector is that prediction markets are decoupling from crypto โ that they are becoming neutral, useful, cross-domain information utilities whose settlement layer happens to be blockchain but whose identity is now political forecasting, sports, and macro.
I think the decoupling is real, and I think it is happening for the wrong reason. It is not decoupling from crypto. It is decoupling from verification. The more mainstream the citation becomes, the less anyone checks the liquidity behind the price. The more institutional the audience, the fewer questions about resolution text. The number travels faster than the conditions that make it meaningful, and the conditions never travel at all.
This is the contrarian read, and it is the one I would defend with capital: the danger to prediction markets is not that they will be rejected as crypto. It is that they will be accepted as oracles before they have done the unglamorous work of proving they deserve to be. Legitimacy awarded in advance of rigor is a debt, and debts in this industry are almost always settled suddenly and in public. Every headline that quotes an odds figure without a volume figure is a loan against credibility that nobody has underwritten.
What I Would Watch, and How I Would Position
If you want to test whether this cycle's prediction-market enthusiasm is substance or froth, there are five things worth tracking, and none of them require a subscription.
Watch the volume on the policy markets specifically, not on the election markets. If the AI-legislation books are trading tens of thousands of dollars rather than millions, the numbers coming out of them are commentary, not signal.
Watch for a second venue quoting the same event. A genuine market price converges across venues; a thin, sentiment-driven number diverges wildly, and the spread between two venues is the cheapest manipulation detector available.
Watch for the bill number. If a specific, numbered piece of legislation gets attached to the market, and if its procedural milestones start appearing in the resolution criteria, then the odds become meaningful โ because they become falsifiable. An odds quote tied to a named bill with a committee calendar is a forecast. An odds quote tied to a category is a vibe.
Watch the citation frequency in mainstream outlets. Each additional citation is another deposit into the venue's legitimacy account, and the account balance is now the sector's most valuable, least audited asset.
And watch the regulators. The commission that once penalized this activity has not lost interest in it; it has lost a round in court. Doctrine changes with personnel, and the jurisdictional question over event contracts remains genuinely open.
As for positioning: volatility is the price of admission, but only if you understand which room you have been admitted to. These markets are not forecasting instruments with a decorative trading layer. They are trading instruments with a forecasting layer, and the forecasting layer is only as good as the assumptions no one publishes. I will keep using them where the resolution is crisp and the depth is real. I will keep ignoring them where the question is a category and the liquidity is a rumor.
Somewhere in Washington, an AI bill that may not exist is being assigned a 30% chance of passing by a market that will not tell you its volume, sourced by a story that will not tell you its bill number, read by an audience that will remember the 30% and nothing else. Follow the liquidity, ignore the hype โ because in this new arrangement, the hype is being manufactured out of the liquidity's absence, and the absence is the only thing anyone has thought to measure.