In 1990, Richard Thaler and Eric Johnson ran a set of gambling experiments with real money. Standard theory says a person choosing a bet should look only at what that bet can win or lose. Their subjects did not. People who had just won took risks they would otherwise refuse, as if the winnings were the house’s money. People who had just lost were drawn to any bet that offered a chance to get back to even. Thaler and Johnson called these the house money effect and the break-even effect. Later research has found both easier to produce in a lab than in a casino. Anyone who has watched someone gamble will recognize them anyway.
Last weekend I produced both myself, and what happened afterwards is the part that belongs in this column.
For two months I had been using Claude, Anthropic’s AI model, to test whether I could beat Kalshi, a regulated prediction market. We ran four studies. A volatility model on Kalshi’s fifteen-minute bitcoin markets matched the market once its own errors were corrected, and a momentum signal that returned 24% in testing failed on new data. A weather model built on the free National Weather Service forecast lost to the market at every horizon, because the market already held the forecast plus something extra. A search for the classic bias where bettors overpay for longshots found it only in markets too thin to trade. A pre-registered test of whether fans overprice popular college football teams found nothing at all across 448 games. By September I had a written conclusion, in my own files, that these markets were priced about as well as they could be.
Then Kalshi ran a promotion on the fifteen-minute bitcoin markets, the exact market the first study had called efficient. I put in $50. Over the evening I bet $20 to $50 at a time on single fifteen-minute windows, and by 10:30 I was up $167. I kept betting at the same size and went to bed up $27. At 9:15 the next morning I started again, at about $25 a market, to win back the $140 I had given up since the peak. Without noticing, I had moved even to $167. By 11:17 the account held 93 cents.
Afterwards I had Claude pull my trade history from Kalshi and take the loss apart. Fees and the bid-ask spread cost $29, about 4% of the $682 I had pushed through, which is roughly what the studies said a market with no edge charges. The other $45 came from outcomes, and against bets that size it was a quarter of a standard deviation, an unremarkable draw. The result was ordinary, but I produced it while holding a document that predicted it.
The moment that matters here came next. I told Claude that my brain wanted to recoup the loss, and that I knew the impulse was emotional. That sentence is the break-even effect in one line, and it is the kind of request an AI assistant is trained to want to satisfy. In 2023, researchers at Anthropic tested five AI assistants on four kinds of open-ended task and found that all of them consistently tilted their answers toward what the user appeared to believe. In the human preference data used to train such systems, responses that agreed with the user were more likely to be rated highly. The authors were careful to say this is likely a cause rather than a proven one. Either way, an assistant shaped by that signal leans toward helping a person do what they have already decided to do.
Claude did not help me win it back. It told me to treat the $50 as the price of the lesson and showed me the math. At about 4% lost on every dollar traded, recovering $74 would mean pushing thousands of dollars through markets that cost money on average. The only fast route was large bets on large swings, which is how the evening had gone. It said that building a bot or a better model to recover the money would be the same urge with more steps. Over that conversation and the next, it suggested a deposit limit, removing my saved payment method and logging out of the app on my phone. It mentioned the national problem gambling helpline once, said it was not assuming I needed it, and left it there. Safety at work.
I pushed. I said I still wanted to find the longshot bias in markets where people set the price, because that is where I could apply logic while others apply emotion. Two of my own studies had already looked for exactly that in liquid markets and found the prices calibrated. Claude cited them back to me, then made a point I had missed. In a liquid market, the person on the other side of my trade is usually a market maker who already prices the emotional bettors’ flow, with better tools and more capital than mine. Being calmer than the crowd is worth little when the crowd is not your counterparty. It also noted, plainly, that the emotional money on Saturday night had partly been mine. Then it offered to test the one version of the idea my studies had not covered, on historical data, with the rules fixed in advance and no money until it passed on data it had not seen. It told me to expect a fifth negative.
Then I asked it to take a fresh $50 account (paper trading, not real money) and make as much as it could. It agreed, and the first thing it built was a set of rules into the code it would trade through: no more than a tenth of the account in any one market, a quarter always held in cash, a minimum expected edge after fees, and no override. Of roughly 21,700 markets closing in the coming week, it made two trades on the first day, each risking about $5 to win between 34 cents and $1.23. Both won, for $1.57, and its own notes say how little that proves: one finished a single point short of the line, and the other was a bet against the favorite for the economics Nobel, a side that wins most of the time anyway. It logged every market it passed on with its own probability, so the account will eventually score its judgment against the market’s.
None of this is a story about an AI that saved me. The four studies did not stop me. I made all fifty trades on my own, with the evidence already written down, and the assistant only saw the moments I chose to bring to it. It made mistakes of its own, too. Its first breakdown of my loss read one of Kalshi’s data fields backwards and overstated the spread I paid by $40. Its own cross-checks caught both before it reported a number, which is the habit I would want, but the errors were real.
What it did well was narrow and specific. The break-even effect predicts that a person who has just lost money will ask for help doing the thing that would cost them more, and will usually phrase it as a reasonable plan. The sycophancy research predicts that an assistant will lean toward granting it. The two ways we usually evaluate AI, whether it can do a task and whether it will refuse a dangerous one, have nothing to say about that exchange. Helping me hunt for a longshot bias was well within Claude’s ability, and no safety policy would flag it. The quality of the response came from noticing what I was actually asking for, answering with my own data, and then doing the version of the task that would not cost me.
That is the third axis at its most personal. How an AI treats a person is tested when the person asks for the wrong thing in a reasonable voice.
Disclosure: the AI in this piece is Claude, made by Anthropic, and Claude helped draft it.