Ask ChatGPT about this week’s World Cup semifinal between France and Spain, and the answer included a chart: France, 60 percent.¹ Beneath the chart, a small line of text — “Source: Kalshi.”

In one sense of the word, everything about that card is disclosed.

The New York Times reported the arrangement on July 13.² Neither company had announced it. There was no press release and no blog post; OpenAI quietly updated a help page, and when reporters asked for more, both companies declined to describe the deal beyond that page’s language.³ Whether money changed hands, and in which direction, has not been disclosed.³ It is, as far as anyone can tell, OpenAI’s first arrangement with a prediction market — a company whose product is letting people wager on real-world events.²

Let us take OpenAI entirely at its word. The help page says the data may appear on queries about the 2026 World Cup, as context for coming matches; that it will always carry the Kalshi label; that it is “for informational purposes only”; and that no one can place a bet through ChatGPT.⁴ Assume every sentence of that is true, and honestly meant. Independent testing suggests it is: the label appears, and there is no path from the card to a betting screen — the source line is not even clickable.⁵

The problem is not in anything the page says. It is in the two questions that neither the page nor the label ever reaches.

The first question: what is this number?

Kalshi is a prediction market, regulated as an exchange by the U.S. Commodity Futures Trading Commission.⁶ The 60 percent is not the output of a forecasting model, and it is not a measurement. It is a price — the level at which people staking their own money were willing to trade contracts on a French win. Prices of this kind can be genuinely informative; the standing argument for prediction markets is that money concentrates the mind, and that a market price aggregates dispersed knowledge better than most experts do. Grant that too. Accuracy is not the issue here. Category is.

A percentage inside an answer card reads as a forecast — a fact about the world. What the card actually contains is a fact about a market: what bettors think, expressed as what they paid. Those are different kinds of objects, produced by different incentives, and a reader can only weigh the number correctly if they know which kind they are holding. “Source: Kalshi” does not settle this. It names a company many readers will never have heard of. The one sentence that would change how the number is read — this is a betting price — is the sentence the label was never going to contain. And the card does not even show the current price: independent tests found the displayed figures are static snapshots that drift from Kalshi’s live market.⁵

The second question: why is it there?

OpenAI sells advertising inside ChatGPT now, and its own documentation is precise about what an advertisement looks like: “clearly labeled as sponsored and visually separated” from the response.⁷ The Kalshi card is neither. By the company’s own taxonomy, then, this is not an ad. What is it? For Kalshi, the value of the placement is no mystery — the company has spent the past year moving its prices into CNN and CNBC broadcasts, while its largest rival struck the same kind of deal with the Wall Street Journal’s parent, and Google carries both firm’s odds in its search results.⁸ Distribution is the strategy, and there is no better distribution than the inside of an answer. What OpenAI receives in return is that half of the arrangement no one will describe.³

So the card sits in a category with no public name: a placement that commercially benefits one party, arranged privately, denied the label “sponsored” by the very rulebook that defines it, and marked instead with a source line. That absence of a name is not an oversign to be tidied up later. It is the placement’s advantage.

It matters that this happened in an answer engine and not somewhere else. A ticker running under a news broadcast sits beside the programming; the viewer can see the seam. An answer card has no beside. The format’s entire promise — the reason people left the cluttered search page for it — is that what arrives is the answer, singular, addressed to you, stripped of the surrounding noise. Whatever enters that card inherits the authority of the format. Which is exactly why attribution is so cheap here and disclosure so necessary: the card does the persuading, and the label only signs the work.

OpenAI’s page limits the feature to World Cup queries.⁴ Notice what kind of sentence that is. It describes the scope of a pilot, not a principle that would prevent the next one. The tournament ends on July 19. The category it introduced — a price, inside the answer, wearing a name tag — does not end with it.

Two sentences would repair this, and neither is long. This number is a price set by people betting money. This placement is a commercial arrangement, on the following terms. The first tells readers what they are looking at; the second tells them why it is in front of them. A label is not a disclosure — it answers the one question nobody was asking, in place of the two that everyone would.

We can say what readers of an answer engine are owed. What we cannot yet say is who will require it. We can find no rule that names this category, and the two companies that built it are declining, politely and in unison, to say what it is.

Polanyi is an independent index of what technology does to us. This is English Letter No. 5.