“The commission is looking to use AI to detect patterns that suggest someone has non-public information.”
That is the stated goal, anyway. But let’s be honest: the CFTC is basically trying to play a game of “Spot the Insider” using a tool that is notoriously bad at understanding causality. It is a classic government move—throw a buzzword at a problem they can’t actually solve with existing legal frameworks and hope the optics suffice.
The CFTC is terrified of prediction markets because these platforms move faster than government press releases. When a market shifts 20% in favor of a specific regulatory outcome three hours before the official announcement, it looks like a leak. The solution, according to the agency, is to plug in some ML models and hope they flag the right people.
Detect patterns that suggest
According to Ars Technica, the agency wants to identify anomalies that scream “I know something you don’t.” This is essentially the same logic used in high-frequency trading surveillance, but applied to a wildly more volatile environment. It’s like a referee in a football match who only throws the flag after watching a 4K slow-motion replay of a play that happened three minutes ago. By the time the “pattern” is detected and a human agent reviews the flag, the money has already moved.
Does this actually stop the leak? Probably not. If you have a direct line to a senator or a high-ranking staffer, you aren’t going to bet your entire net worth on a single account that is easily traceable to your home IP address. You’ll split the trade across twenty accounts (or use a proxy).
The tool is a scarecrow.
The agency is betting that the mere existence of “AI surveillance” will deter the casual insider—the mid-level staffer who thinks they’re being clever. But for the professional players, this is just a new variable to optimize for. They will simply find the threshold at which the AI stops flagging “unusual” volume and keep their trades just below that line.
Non-public government information
The real friction here isn’t the AI; it’s the evidence. An ML model can flag a correlation—say, a sudden surge in “Yes” bets on a specific regulatory ruling—but it cannot prove that the bettor received a leaked PDF from a staffer. It provides a lead, not a conviction. (Unless the CFTC plans to use “the model said so” as legal proof, which would be a disaster for due process).
We’ve seen this before with the SEC and their various attempts to automate “market abuse” detection. The result is usually a mountain of false positives and a few easy wins against people who were too sloppy to use a VPN. The gap between a statistical anomaly and a legal crime is a chasm that no amount of compute can bridge.
By Q4, we will see the first high-profile “AI-detected” insider trading case get tossed by a judge because the evidence is purely circumstantial.
The government is trying to police a mirror. Prediction markets don’t create the leaks; they just reflect them. The market isn’t the source of the problem; it’s the thermometer. Trying to stop the “insider” by monitoring the thermometer—or by using a black box to watch the thermometer—is a waste of GPU cycles. If the CFTC actually wanted to stop insider trading, they would focus on the people leaking the information, not the people betting on it. Instead, they’ve chosen the path of least resistance: deploying a tool that looks impressive in a press release but does very little to change the incentives of the people actually holding the secrets.
Leave a Reply