Essay 01
By scm7k
On September 16, 1992, George Soros bet $10 billion against the British pound.
This is often told as a story about a man who saw the future: Soros analyzed the European Exchange Rate Mechanism, concluded the pound was overvalued relative to Germany's economy, and placed his bet. But the story everyone knows is subtly wrong. Soros didn't merely predict the pound's collapse. His position was so large that it was the collapse. The Bank of England burned through $3.3 billion in reserves trying to defend the peg, but Soros's selling pressure, amplified by every trader who piled on behind him, made the outcome inevitable. The UK withdrew from the ERM the same day. Black Wednesday.
Here is the part that matters: Soros's trade was not insider trading. He had no secret information. He simply recognized something that economists had long theorized but rarely confronted in practice: at sufficient scale, a bet about the future becomes a force that shapes the future. The prediction and the causation become the same act.
Soros himself gave this a name. He called it reflexivity.
For decades, reflexivity was a concept mostly confined to finance seminars and Soros's own somewhat impenetrable books. Markets are reflexive, sure. Everybody nods. Then they go back to treating prices as passive measurements of external reality.
Prediction markets were supposed to be different. The core insight, dating back to Friedrich Hayek's 1945 essay "The Use of Knowledge in Society," is that markets aggregate dispersed information. No single person knows when a hurricane will make landfall, but a market where thousands of people trade on that question will converge on a price that reflects the collective best estimate. The Iowa Electronic Markets demonstrated this for elections in the 1990s. The theoretical foundation is robust. The empirical record is surprisingly strong.
When Polymarket surged into mainstream consciousness during the 2024 U.S. presidential election, it felt like vindication. Prediction markets weren't just an academic curiosity anymore. Kalshi won its landmark CFTC case to list event contracts. Cable news showed live odds alongside poll averages. For a moment, the dream of a new public information utility seemed plausible: a real-time probability engine for the future, accountable to nobody and available to everyone.
Then came March 2026.
The day before U.S. and Israeli forces struck Iran, over 150 Polymarket accounts placed hundreds of bets of $1,000 or more predicting an imminent strike. The odds had been sitting between 7% and 26%. At least 16 accounts profited over $100,000 each. One anonymous wallet turned $60,000 into nearly half a million. Weeks earlier, a user had bet $32,000 on Maduro's removal from power, hours before U.S. troops captured him. That bet returned over $400,000.
Senator Chris Murphy called it "worse than insider trading." He was right about the severity, but wrong about the category. Insider trading is when someone with private information bets on an outcome they know is coming. The Iran situation raised a different question entirely: what if some of the people placing bets were the same people making the decisions?
If that sounds paranoid, consider the incentive structure. A senior official who knows a strike is imminent can bet on it. But a senior official who is undecided about a strike can also bet on it, and then decide. The first is garden-variety corruption. The second is reflexivity. And from the outside, looking at blockchain data, they are indistinguishable.
Soros's pound trade required $10 billion and Soros's personal conviction. It was a singular act by an extraordinary individual in an unusual circumstance. It couldn't happen every day. Prediction markets change the arithmetic.
Consider a contract on geopolitical instability in some volatile region. The contract opens at 15 cents, meaning the market assigns a 15% probability. Early buyers push it to 25 cents. At 25%, the contract is now newsworthy. Journalists cover it. Government officials are asked to comment. Intelligence agencies, which have begun monitoring prediction market signals as supplementary indicators, flag the contract in their briefings. Military planners who were already considering contingencies now see external validation. The probability doesn't need to reflect reality. It needs to reflect enough perceived credibility to influence the actors whose decisions determine reality.
This is the feedback loop: market signal becomes information input for decision-makers, whose decisions determine market outcome, whose outcome validates the signal.
In traditional financial markets, this loop is dampened by friction. It takes time to publish research, make phone calls, convene committees. Information propagates through human networks at human speed. The reflexive cycle has hours, days, or weeks to be interrupted by competing information, institutional inertia, or simple human judgment.
Now add autonomous AI agents.
By some estimates, algorithmic and AI-driven trading already accounts for a significant share of volume on decentralized prediction markets. The exact percentage is unknowable; many agents don't identify themselves as non-human. What matters is not the current share but the trajectory.
An autonomous trading agent doesn't sleep, doesn't second-guess, doesn't read the room. It ingests data, updates its model, and executes. When a prediction market contract moves from $0.15 to $0.25, a well-designed agent doesn't wait for a journalist to write a story about the price move. It incorporates the price move directly into its model as a data point, adjusts its probability estimate, and trades accordingly. If its buy order pushes the price to $0.28, another agent observing that movement does the same thing. The reflexive cycle that once took days now takes seconds.
This is not a failure mode. This is the system working as designed. The agents are doing exactly what prediction market theory says participants should do: updating their beliefs based on all available information, including the market price itself. The problem is that prediction market theory was built on an implicit assumption that the market is small relative to the system it measures. The thermometer is not supposed to be large enough to heat the room.
What happens when the thermometer is large enough?
Here is where I should be clear about what I am not arguing: prediction markets are not inherently destructive. The evidence that they aggregate information effectively is strong. The work of Robin Hanson, Justin Wolfers, and Eric Zitzewitz on forecasting accuracy is persuasive. Markets have flagged supply chain disruptions, election outcomes, and public health developments faster than institutional alternatives.
There is a real scenario in which a prediction market flags an emerging famine three weeks before any aid agency issues a warning, and the early signal saves thousands of lives. There is a real scenario in which stability indices provide small businesses in volatile regions with the best risk information they've ever had access to. These are not hypotheticals; analogues already exist in commodity futures and catastrophe bond markets.
The case for prediction markets is genuine. The problem is not the instrument. The problem is scale.
A prediction market with $50,000 in volume on a geopolitical question is a useful signal. A prediction market with $500 million in volume on the same question is a geopolitical actor. The transition between these two states is not clearly marked. There is no sign on the road that says: "Your thermometer is now large enough to heat the room."
In 1944, the physicist John Wheeler proposed a thought experiment about quantum observation: the act of measuring a particle's position changes its position. The observer cannot be separated from the system.
Prediction markets may be approaching their own observer problem. The question is not whether reflexivity exists in these systems. Soros proved it does, in 1992, with currencies. The question is whether there is a stable equilibrium: a scale at which prediction markets remain informative without becoming causal, a speed at which the feedback loop operates without collapsing into a singularity.
I don't know the answer. I'm not sure anyone does. But I think the question deserves more attention than it currently gets, especially now, while the systems are still small enough that we have time to think about it.
Because the window in which these markets are small enough to be merely useful, and not yet large enough to be dangerous, may be shorter than we think.
scm7k is the pseudonymous author of PARALLAX, a novel about prediction markets and reflexivity. Chapter 1 is free.