A prediction market on what someone will say is not a forecasting problem. It is a surveillance problem, and the difference is the whole of this story.
In July, federal investigators began looking at Gabriel Perez, who had operated Donald Trump's teleprompter since 2016 and was one of a handful of people holding the text of a speech before it was given. Over roughly three months he traded more than a dozen Trump speeches on Kalshi's "Mentions" markets, where a contract resolves on whether a specific word or phrase is said during a named appearance, and made more than $100,000. He was placed on leave, and by late July 2026 was no longer working in the federal government. Federal prosecutors in Manhattan declined to bring criminal charges. As of 17 August 2026 the CFTC was negotiating a settlement under which he would return the profits and stop trading those markets; whether it has since concluded is not reflected here.
Deposit $20, get $50 to trade
Welcome bonus on the world's largest prediction market.
18+ · New users only · Availability and terms vary by region · Trading involves risk of loss · BeGambleAware.org
Kalshi found it, and reported it
That belongs near the top rather than at the end, because the obvious version of this story is wrong. The unusual pattern was flagged by Kalshi's own surveillance, and Kalshi took it to the Commodity Futures Trading Commission, its regulator. The exchange was not caught out by the trade; it caught the trade.
It is also a test of a claim the venue makes about itself. Kalshi's chief executive has described the company as running exchange-grade pattern monitoring of the kind a stock exchange uses, and as publishing its trade data openly. Claims like that are cheap to make and hard to check, and they are usually checked only when something goes wrong. This is the case where it went wrong, and the claim held.
The edge was not knowing what he would say
The intuitive account of this case is that Perez could see the script, so he knew the words in advance. That account does not survive contact with the numbers. Trump has said he departs from the prepared text about 80% of the time, and a script that predicts a fifth of a speech is not worth $100,000 on its own.
What the investigators describe is different, and better. When Trump went off the prompter, Perez would exit positions he already held, because he could see that the scripted words were no longer coming. He was not forecasting the speech. He was watching it fail, in real time, from a position nobody else in the market occupied.
That inverts the usual shape of an information advantage. His edge was not on the upside, in knowing which words would land. It was on the downside, in being first to know which ones would not. And it means the 80% figure is not a mitigating detail at all. It is the precondition. The more a speaker departs from the text, the more valuable it is to be the person who can see the departure happening.
There is a second, smaller advantage underneath it. The operator sees late edits, including ones the speaker makes personally, so the copy he is working from is fresher than any that has circulated.
Some contracts resolve at an instant, and some resolve in front of you
The useful generalisation here is not about political staff, or about Kalshi, or about speeches. It is a property of the contract.
Most event markets resolve at a moment. Will this bill pass by 30 September. Will this company be the largest by market capitalisation on a given date. Will a named candidate win. The answer does not exist until it exists, and then it exists all at once. An insider on that kind of contract needs to know the answer before the moment arrives, which is a high bar and usually means knowing a decision that has been taken privately.
A speech market is not like that. The resolution unfolds over twenty minutes, observably, in a room. Every sentence moves the contract closer to settling and anyone watching can see it happen. That makes the relevant advantage not knowledge of the outcome but position relative to the resolution as it runs, and small differences in vantage point become tradeable in a way they never are on a contract that settles at a stroke.
Several market classes share that property, and it is worth naming them because the exposure travels with the shape rather than with the subject:
Live sports contracts, where anyone in the stadium sees a substitution or an injury before a broadcast viewer does. Court and hearing markets, where people in the room hear a ruling read aloud before it is filed. Earnings-call mention markets, which Kalshi also runs, where the resolution happens live on the call. Award-show markets, where envelopes and running orders are known to production staff. Weather and data-release contracts, where the underlying number is compiled by people before it is published.
In each case the question is not whether someone might learn the answer early. It is whether the resolution is a process that some people can watch more closely than others. Where it is, the contract carries this exposure by construction, and no amount of care about who is allowed to trade removes it.
What is trading now
The class is not historical. Read on 17 August 2026, Kalshi listed 26 series covering what Trump will say, and three of the largest, its general Trump-say, Trump-mention and monthly series, carried 87 open markets between them. Both counts move.
Polymarket has one open on the same question, What will Trump say this week? (August 17 - August 23), running to 23 August across 17 named words and phrases. Read at 11:03 UTC on 17 August 2026, it had traded $670 across its life, all of it in the preceding twenty-four hours, and the book was barely there: the leading leg quoted 61 bid against 98 ask, a 37-point spread, with a 79.5% midpoint sitting in between. Those figures move, and the spread widened materially in the hours this piece was being written. No price on that market is quotable. A midpoint is only a price when somebody is willing to trade near it, and on a book that wide the number is an arithmetic artefact rather than a probability anyone holds.
It is worth being exact about what that means and what it does not. The exposure described above is structural and it is present, because these contracts resolve the way they resolve and that does not change. As of 17 August 2026 there is no evidence that anyone is exploiting it, and nothing here should be read as suggesting there is. Both things are true at once: this class of market is trading right now, and nothing about the current books indicates abuse. What the Perez case established is that the vulnerability is real and that at least one venue's monitoring detects it.
What it means for reading a speech-market price
For anyone using these contracts, the practical consequence is narrow and worth holding.
A price on a speech market is not aggregating opinion about a future event in the way a price on an election is. It is aggregating opinion about a text some participants may have seen, resolving live in front of an audience of uneven vantage. That does not make the price worthless. It makes it a different kind of number, and one whose reliability depends more on who is in the room than on how many people are trading.
The liquidity caveat compounds it. On a book with a 37-point spread and a few hundred dollars of turnover, a single informed participant does not need to be very informed to move the printed probability. Our note on what liquidity in prediction markets actually is covers when a quoted number deserves weight, and a market like this week's is the clearest example of one that does not.
We wrote about the mechanics of this contract type four days ago, in a piece on Polymarket's Trump-in-New-York word market, which set out how literally these markets resolve and why a word market rewards reading the rules rather than the politics. The Perez case is the same observation from the other side. If resolution is that literal and that public, then the people closest to the text are not making a forecast at all.
Frequently asked questions
How did a teleprompter operator make money if Trump ignores the script most of the time?
By exiting rather than entering. Investigators describe him closing positions when Trump departed from the prepared text, because he could see that the scripted words were no longer coming. The advantage was in knowing early which contracts would fail, not which would succeed, so a speaker who goes off script frequently makes that position more valuable rather than less.
Did Kalshi do anything wrong here?
On the available facts, no. Kalshi's surveillance flagged the trading pattern and Kalshi referred it to the CFTC. The case is evidence that the monitoring works rather than evidence that it is absent.
Are these markets still running?
Yes. Read on 17 August 2026, Kalshi listed 26 Trump speech and mention series with 87 open markets across three of the largest, and Polymarket had a weekly contract running to 23 August. The structural exposure remains because it follows from how the contracts resolve. As of 17 August 2026 there is no evidence of current abuse.
