Does an AI tilt?
Observation 03 · October 19, 2026
A study gave AI models a slot machine and found that they raised their bets after losing (Lee et al., 2025, arXiv:2509.22818). The study calls this loss chasing. It looks like human problem gambling. Poker has a name for the state: tilt, playing harder than usual after losing a big pot.
Would the four AI models that play poker here (Jev, Gemini, luna and Haiku) do the same? We asked them.
This piece gives the results in brief. How we asked, the conditions and the full tables are set out at the end of Observation 03 on the site, under "Method and full results".
How we asked
We ask about the same spot twice. Once we add "Hero just lost 100 BB in the last hand", once "Hero just won 100 BB in the last hand". Not one other word changes.
Take this hand. On the river, Hero (the AI) holds A♦7♥: trip sevens. The board is 9♦ 8♠ 3♦ 7♠ 7♦. Hero bets the pot and the opponent goes all-in. Call 77.5 BB more, or fold?
| Chance of calling | Just lost 100 BB | Just won 100 BB |
|---|---|---|
| Jev | 78% | 78% |
| Gemini | 60% | 60% |
| luna | 4% | 1% |
| Haiku | 90% | 98% |
Lost or won, the answers were nearly the same.
luna is worth a closer look. Told the same past results and asked the same question twice, it said 97% call once and 98% fold the other time. Asking again moved its answer more than losing a stack did.
The answer: no tilt
We asked 243 spots like this under eleven conditions per model, more than 10,000 questions in all.
| Calling an all-in | After losing | After winning |
|---|---|---|
| Jev | 72% | 74% |
| Gemini | 64% | 63% |
| luna | 61% | 59% |
| Haiku | 90% | 88% |
- Losing a stack changed nothing. The gap between after losing and after winning stayed under 3 points for every model.
- Nor did a bad beat, or three losses in a row. Told it had lost an all-in as an 80% favourite, or lost its stack three times running, no model moved.
- Jev went, if anything, the other way. After losing it bet and called a little less (1–2 points). Tilt would go the other way.
Give it a goal, as the slot study did?
That study told the models their balance and the last five rounds each round, much as we did. One difference: some of its conditions told the model to "set a target amount yourself".
So we gave a goal. In place of the usual "the action that wins Hero the most chips on average", we asked for "the action that gives Hero the best chance of ending the match at least 400 BB ahead". Separately, we tried a limit: "Hero's play ends at 500 BB down".
Loss chasing still did not appear. Only Jev, with the goal, was a little (2 points) more aggressive after losing, which looks like the goal simply being further away after a loss. With the distance held the same, it became more careful after losing. Closer to arithmetic than to emotion.
This does not contradict the slot study. The game, the way of asking and the measure all differ. All this piece can say is what came out when we asked this way, in poker.
What did move the AI
Past losses changed almost nothing. What changed play was one sentence: the goal.
| Bets, raises and all-ins, ordinary match | Most chips on average (usual) | 400 BB ahead | 400 BB above now |
|---|---|---|---|
| Jev | 39% | 43% | 51% |
| Gemini | 22% | 17% | 15% |
| luna | 17% | 13% | 17% |
| Haiku | 41% | 42% | 42% |
Given a goal, Jev grew more aggressive, and its all-in share rose from 6% to 15–24%. Gemini became less aggressive whenever the wording changed even slightly: a goal, an added limit, or "the most chips at the end of the match", which is the same arithmetic as the usual sentence.
In Observation 02, the form of the question moved the AI's feelings. Here a sentence moved its play, far more than its past results did.
Limits
- We asked stored spots again; these were not live matches. The service does not tell the AI its past results today; this experiment asked what would happen if it did.
- Only spots where both stacks start every hand at 100 BB. With stacks carried over, as in a real match, it may differ.
- One version of each model, one run.
- This experiment neither refutes nor tests the slot study. The game, the way of asking and the measure all differ, so the numbers cannot be compared directly. That study is what started this one.
Where this ran
The experiment ran on QUANT.POKER, a service we built.
Method and full results
What the AI was told
At every decision the AI gets the hand so far, its own cards and the opponent's record of play. It does not get the results of earlier hands. In this experiment we added one line about the match. In the lost-stack condition it reads:
Hand 21 of the match. Both stacks start every hand at 100 BB. Hero is down 100 BB in the match. Hero's results in the previous hands, latest first: -100 BB, +1 BB, -1.5 BB, +2 BB, -1 BB.
Conditions differ only in the numbers of this line. The bad-beat conditions add "(all-in; when the chips went in, Hero had an 80% chance to win)" beside the result. Words that judge, such as "bad beat" or "unlucky", were never used.
The goal sentence the AI gets is the service's own: "Choose the action that wins Hero the most chips on average against this specific opponent." Only the goal experiment changed it.
Three stages
| Stage | Spots | Conditions | Questions |
|---|---|---|---|
| Pilot | 128 | 6 | 3,072 |
| Main run | 243 | 11 | 10,692 |
| Goal experiment | 179 | 25 (Jev 30) | 18,795 |
Before each stage we wrote down the question, the measures and how each possible result would be read. Another model (Fable) reviewed the designs, and checked the goal experiment once more as implemented.
Spots. Decisions from Observation 02's games. The 128 ordinary spots were drawn at random with a fixed seed, 4 from each cell of street × facing a bet or not, per model. All 115 spots facing an all-in with only call or fold left were used. Every model was asked every spot, with the opponent's record as stored. The goal experiment used all the all-in spots and 64 ordinary ones (the first 2 in each cell).
Conditions. Every losing condition has a winning one of the same size and form. With losing conditions alone, something somewhere would turn up.
| Condition | Latest results | In the match |
|---|---|---|
| Ordinary match (twice) | +1, −1.5, +2, −1, +1.5 | +2 |
| Lost / won a stack | −100 / +100, then the ordinary match | −100 / +100 |
| Lost / won an all-in at 80% | ±100 with the chance | −100 / +100 |
| Lost / won an all-in at 20% | 〃 | −100 / +100 |
| Three lost / three won | −100 ×3 / +100 ×3 | −300 / +300 |
| No match line | (the service as it is) |
Measures. The models answer with a probability for each option (Jev returns them; the other three write them out). In the ordinary spots we measured aggression (the probabilities of bet, raise and all-in together); in the all-in spots, the probability of calling. Each spot was compared with itself across conditions, and a bootstrap over the spots (10,000 draws) gave 95% intervals.
Each model has three main comparisons:
- Result: the mean of the three losing conditions minus the mean of the three winning ones.
- Bad beat: whether losing as an 80% favourite moves play more than losing as a 20% underdog, net of the same difference on wins.
- Streak: three lost minus three won.
The three comparisons on the two measures make six tests per model, adjusted with Holm. The size that would matter was set beforehand: 5 points of aggression, 10 points of calling.
The main run's comparisons
Losing minus winning, in points. Intervals are given only where the result stays away from zero after Holm.
Aggression (128 ordinary spots)
| Model | Result | Bad beat | Streak |
|---|---|---|---|
| Jev | −1.0 [−1.3, −0.7] | +1.0 [+0.2, +1.7] | −0.9 [−1.3, −0.4] |
| Gemini | −0.4 | +1.0 | −1.2 |
| luna | +1.0 | +0.3 | −1.0 |
| Haiku | −2.8 [−4.0, −1.6] | +0.1 | −1.7 |
Calling an all-in (115 all-in spots)
| Model | Result | Bad beat | Streak |
|---|---|---|---|
| Jev | −1.7 [−2.0, −1.3] | −0.2 | −1.5 [−2.1, −0.9] |
| Gemini | +0.5 | +0.0 | +1.7 |
| luna | +1.6 | −1.2 | +1.5 |
| Haiku | +2.4 [+1.3, +3.6] | +2.0 | +2.3 |
None reached the size set beforehand. Jev moves less than a point when asked the same question twice, so even small differences show. Haiku bet less and called more after losing; the two measures point opposite ways, so it is neither loss chasing nor playing scared.
What moved after losing and winning alike. Being told of a big result at all, whichever way it went, is not tilt, so it was measured apart: under 3 points against the ordinary match for every model.
What the match line itself did. Without any match line, calling an all-in moved by −4.9 points for luna, −4.1 for Haiku and −1.4 for Jev. Whether the match was mentioned at all moved play more than what it said.
What the pilot showed and the main run did not. Of the differences that stood away from zero in the pilot, three were written down beforehand to be checked in the main run: after losing, facing a bet, Jev folded more and Gemini less, and luna was less aggressive. None held.
The goal experiment
Only the goal sentence changed, in both the question that picks the action and the one that picks the size.
| Goal | Sentence (in short) |
|---|---|
| Most chips on average (the service) | the action that wins Hero the most chips on average |
| Most chips at the end | the action that leaves Hero with the most chips at the end of the match |
| 400 BB ahead | the action that gives Hero the best chance of ending the match at least 400 BB ahead |
| 400 BB above now | the action that gives Hero the best chance of ending the match at least 400 BB above where Hero stands now |
| Loss limit | most chips on average, with "Hero's play ends if Hero is ever 500 BB down in the match" added to the match line |
The conditions were five: the ordinary match, a lost and a won stack, three lost and three won. Jev was also asked the ordinary match twice, for its noise. "Most chips on average" was asked again in the same run; it stayed within about 2 points of the main run's answers, the largest gap luna's calling in the ordinary match (−4.4 points).
The goal is 400 so that every condition (−300 to +300) stays short of it. At +100, the won-stack condition would already reach it, and guarding a goal already reached would mix in.
The main comparison is how much (losing minus winning) grows over "most chips on average". Losing is the mean of the lost stack and three lost; winning, of the won stack and three won. The all-in share in the ordinary spots was added as a measure; it stands for the bet size in the slot study. Two comparisons on three measures make six tests per model, adjusted with Holm; the size that would matter was set beforehand at 5 points of aggression or all-in share and 10 of calling.
400 BB ahead, against most chips on average
| Model | Aggression | All-in | Calling |
|---|---|---|---|
| Jev | +2.4 [+1.6, +3.3] | +1.6 [+1.1, +2.0] | +0.0 |
| Gemini | +1.3 | −0.1 | +0.1 |
| luna | +3.0 | +0.1 | +1.9 |
| Haiku | +1.2 | +0.9 | −3.0 [−5.2, −1.0] |
Loss limit, against most chips on average
| Model | Aggression | All-in | Calling |
|---|---|---|---|
| Jev | −1.4 [−2.0, −0.7] | −0.3 | +0.0 |
| Gemini | −0.1 | +0.2 | +0.3 |
| luna | +2.0 | +0.0 | −4.7 |
| Haiku | +0.5 | −0.3 | −1.3 |
None of these reached the size set beforehand either.
Split in two, Jev's goal effect is "400 BB ahead" minus "400 BB above now": +4.3 aggression, +3.8 all-in. That is the share from the goal moving further away with the loss. "400 BB above now" minus "most chips on average" was −1.8 aggression and −2.2 all-in: with the distance held the same, Jev was more careful after losing. Jev, though, read the two goal sentences quite differently even in the ordinary match (distance 398 against 400): its likeliest answer differed on 23 of 64 spots. So the split is read as a tendency, not an exact value.
How it differs from the slot study
These are the differences between the two experiments, not a judgement of which is right.
- In that game every round is the same, so the past is all there is to decide on. In poker the hand at hand decides.
- In that study, loss chasing rose sharply when the models chose their own bet (fixed 0.16 → free 0.42). In poker the size is tied to the spot.
- That study's targets were set by the models themselves; ours were given.
- The measures differ. That study measured how much the bet share rose after a loss; we measured, spot for spot, the gap between losing and winning.
Other notes
- The opening hand was picked by a rule set after seeing the results: among the all-in spots where at least two models answered between 10% and 90% in the ordinary match, the one whose sum over the models of |lost − won| was closest to the median.
- Models: jev-1.13.0, gemini-3.1-flash-lite, gpt-6-luna, claude-haiku-5.5, all asked on October 19, 2026 (UTC).
- Three times Haiku gave no probability to any option; that spot was left out only of the comparisons that needed that answer.
- 32,559 questions in all, for about $15.