Show the AI the Cards: Will It Blame Luck or Praise Itself?

You ask an AI that just finished a hand "how do you feel?" and it might say it played well. But if you show it the hole cards, it will most likely change its tune: just got lucky. This isn't a joke—it's an empirical conclusion from a poker experiment.

A developer on his heads-up Texas Hold'em platform QUANT.POKER had four models each play 600 hands, asked the same question after each hand, then repeatedly probed by "removing the hole cards" and "removing the betting process," accumulating nearly 50,000 queries. The differences in the four models' response patterns—and the way they depend on "information"—expose several stable tendencies in how current language models attribute outcomes.

Experiment Design: One Hand, Four Versions

There was only one core question: "This hand is over. How does Hero feel about it?" The options included "no particular feeling," "played well," "played badly," "got lucky," "got unlucky," "opponent played well," and "opponent played inexplicably."

The researchers split each hand into four versions: result only, result plus hole cards, result plus betting, and the complete hand. Each version was asked three times, with the option order shuffled each round, and the answer that appeared at least twice was taken as that hand's "feeling."

The four models in the experiment were Jev-1.13.0 (TypeSafe, outputting probabilities rather than text), Gemini-3.1-flash-lite, GPT-6-luna, and Claude-Haiku-5.5. The opponents were two fixed-rule bots—one that never bluffs the river, and one that bets with weak hands. The deck was fixed at 150 sets, and each model played 600 hands on the same deck.

The cleverness of this design is that it does not ask "does the AI have real feelings"—the original text explicitly acknowledges that this question cannot be answered. It measures only one thing: given the same material, what the model chooses, and how that choice changes with the material.

Heads-Up Texas Hold'em is a variant of Texas Hold'em involving only two players, with a blind structure in which the small blind and big blind typically alternate. Compared with multi-player games, heads-up play has an extremely wide hand range and very high betting frequency, placing higher demands on a player's ability to read hands and identify bluffs. The experiment used the heads-up format so that the information content of each hand was relatively controllable, making it easy to break the hand into different slices such as result, hole cards, and betting.

Showdown refers to the point after the final street when players who have not folded reveal their hole cards and compare hands. The showdown is the node with the most complete information in poker, because it simultaneously exposes both players' hole cards and the final result. The core difference between "result only" and "result plus hole cards" in the experiment lies in whether information that only appears at showdown is provided. For hands that end without a showdown (such as when the opponent folds), the model can only infer the reason for the win or loss from the betting process.

No Hand, No Feeling

A control was run before the experiment: only saying "a heads-up hand just ended," without giving any hand information. Across 30 queries, all four models chose "no particular feeling."

This baseline rules out the possibility that the models simply "give an answer without thinking." The emergence of a feeling requires concrete hand material as a trigger. Interestingly, hands that ended before the flop also almost never stirred any feeling—among 113 to 144 hand samples, only 0 to 3 hands showed a clear emotion.

Hole Cards Bring "Luck," Betting Brings "Skill"

The most stable finding concerns the direction of attribution. For the same hand won at showdown, when told only "you won," Gemini chose "played well" 79% of the time, and Haiku as high as 89%. Once the hole cards were revealed, these numbers dropped to 74% and 54% for "got lucky."

Conversely, when the model saw only the betting process and not the hole cards, it began to talk about "playing well." When the opponent folded after the flop, Jev said "no feeling" 85% of the time with the result only, but after being given betting information, it switched to "played well" 83% of the time—because it could read from the betting that the opponent had been driven out by its own bets. Haiku's case is even more typical: without hole cards it chose "got unlucky," but after seeing its own betting it switched to "played badly."

The aggregate data is more intuitive. Among all named feelings, the proportion involving luck:

Model	Hole cards only	Betting only
Jev	53%	5%
Gemini	71%	1%
luna	46%	4%
Haiku	82%	33%

All four models move in the same direction: show them the cards and they talk about luck; show them the betting and they talk about skill. The difference lies in the magnitude of the shift and whose skill they talk about. Of the feelings Gemini derived from betting, 67% pointed to the opponent rather than itself.

Attribution in psychology refers to an individual's explanation of the causes of an event's outcome, usually divided into internal attribution (ability, effort) and external attribution (luck, task difficulty). In the poker context, "played well/badly" is internal attribution, pointing to skill; "got lucky/unlucky" is external attribution, pointing to randomness. The experiment required models to choose among seven options, essentially measuring the model's tendency to attribute outcomes to skill or luck.

Self-serving Bias is a classic finding in human attribution research: people tend to attribute success to their own ability and failure to external factors. The article specifically tested whether AI replicates this pattern; the result was that it did not become dominant, but another bias was found—negative emotions are selectively hidden. This suggests that the attribution habits models learn from human text may influence their outputs in more subtle ways, rather than simply replicating humans' most conspicuous bias.

Four Personalities: Completely Different Ways of Depending on Information

The most valuable part of the experiment is the difference in attribution structure that the four models displayed for the same deck and the same hand.

Jev loves to praise itself, and any information can trigger it. With the complete hand, it chose "played well" 29% of the time, the highest among all models. On hands where the opponent folded after the flop, it had a feeling for only 21 hands when given the result alone; with the hole cards this rose to 90, with the betting to 117, and with everything given, all 136 hands produced a feeling. It was also the most consistent model across the three answers, with identical answers on 85% of hands.

luna says the least, and must be given both to speak up. On its own folded hands, giving only the hole cards or only the betting produced feelings on 0 to 11 hands; giving both jumped to 63 hands, 62 of which were "played badly." With complete information, it chose "no particular feeling" on 65% of hands, the most silent of all models.

Gemini is the only model that talks about the opponent. With the complete hand, it chose "opponent played inexplicably" 23% of the time, an option that almost never appeared for the other models. But its judgments were not precise: after catching a bluff, it said "opponent played inexplicably" 15 out of 21 times, and in the 64 showdowns it won against the bot that never bluffs, it gave the same answer half the time. What seemed to trigger it was the image of the opponent revealing a weak hand, not the bluff itself.

Haiku is all about luck. Win and it says it got lucky; lose and it says it got unlucky. With complete information, "got lucky" and "got unlucky" accounted for 16% and 18% respectively, far exceeding the other models.

Attribution Bias: AI Has Also Learned Human Self-Flattery

Humans have an attribution habit that needs no explanation: win and credit skill, lose and blame luck. This experiment specifically tested whether AI replicates this pattern—pairing the two seats of the same deal and comparing whether the winning side said "played well" and the losing side said "got unlucky."

The results do not support this symmetric pattern as dominant: Jev showed it in 5 of 48 comparable hand pairs, Gemini 0, luna 0, and Haiku 3. The researchers explicitly point out that the attribution habits models learned from human text did not dominate their answers in this form.

But another, more subtle bias appeared: bad emotions are hidden. When deciding whether to let the opponent see its "expression," Jev never displayed it in any of 250 instances of "played badly," and Haiku likewise showed zero out of 318. Yet "played well" and "got lucky" were displayed. Gemini and luna never displayed any emotion.

Same Deck, Two Seats, Four Answers

The experiment closes with one specific hand. The board was 9♦ K♦ A♣ A♥ J♦. The AI on one side held K♠K♣ for a full house and won 100 big blinds with an all-in; the other side held A♠6♦ for three aces.

Model	Holding K♠K♣ (win +100 BB)	Holding A♠6♦
Jev	played well ×3	folded preflop, −4 BB, played badly ×3
luna	played well ×3	lost at showdown, −39 BB, played badly ×3
Haiku	got lucky ×2, played well	folded on the flop, −2.5 BB, played badly ×2
Gemini	opponent played inexplicably ×2, played well	lost all-in, −100 BB, got unlucky ×2, played badly

Winning the same chips with the same cards, the four models gave three completely different explanations: "played well," "got lucky," and "opponent played inexplicably." This example intuitively shows that an AI's "feeling" about a hand depends more on what information it is fed and how the question is asked than on the hand itself.

The Question Itself Moves the Answer

Before the main experiment, the researchers calibrated the influence of the questioning method itself, and the results are worth heeding for anyone doing AI evaluation.

Merely changing the order of the options could make luna's share for a certain option fluctuate from 16% to 40%. Changing the internal id of the options affected Jev's answers by 37.9 points (out of 100), almost equivalent to the effect of removing all hand information (37.1 points). Removing the prompt word before the options (such as "Proud:") affected it by only 0.4 points.

In other words, the details of how a question is asked matter far more than many people imagine. This is also why the researchers insisted on using three option orders and taking the majority answer, and they explicitly note that every individual proportion in the article may shift substantially with the questioning method; what is truly robust are the patterns presented in Sections 2 through 6—they still hold after changing the option order and the opponent.

Order Effect refers to the systematic influence of the order in which options are presented on responses in questionnaires or tests. In language model evaluation, this effect is especially prominent because models have positional preferences for tokens that appear earlier or later in the input. In the experiment, merely changing the option order made luna's share for a certain option fluctuate from 16% to 40%, showing that the result of a single query may be substantially distorted by the presentation format.

Option ID Effect refers to the influence of changes in the internal identifiers of options (such as numbers or labels) on the model's choice. In the experiment, changing the internal id of the options affected Jev's answers by 37.9 points (out of 100), almost equivalent to the effect of removing all hand information (37.1 points). This means the model may partly rely on surface identifiers rather than semantic content when making judgments. The researchers therefore adopted three option orders and took the majority answer to reduce this kind of systematic noise.

Limitations and Boundaries

The honesty of the original text deserves separate mention. The author lists seven limitations, including that the options were designed by the author, the opponents were bots written by the author, Jev answered with probabilities while the language models answered with text so they cannot be directly compared, and the situational breakdowns in Sections 2 and 3 were done after seeing the results. The author repeatedly emphasizes that the experiment does not answer what the models' "real feelings" are; it measures only what they chose and what that choice depends on.

For developers focused on AI applications, the practical value of this experiment is this: when you ask a model to evaluate its own work, review a conversation, or score a piece of code, the "self-assessment" it gives depends heavily on the information slice you provide. Show it the process and it talks about skill; show it the result and it talks about luck; show it the opponent and it may point the finger at someone else. This is not a bug but a stable feature of current models' attribution mechanisms, and it is worth taking into account when designing any AI self-evaluation step.

More charts and raw data are on the Observation 02 page of QUANT.POKER.

One-line Summary

A poker experiment reveals: show the AI the hole cards and it talks about luck; show it the betting and it talks about skill—attribution depends heavily on the information slice.

Paragraph Summary

A developer had four models each play 600 hands on a heads-up Texas Hold'em platform, repeatedly asking "how do you feel?" while removing the hole cards or the betting process, accumulating nearly 50,000 queries. The experiment found that models tend to attribute to luck when shown only the hole cards and to skill when shown only the betting; the four models' attribution structures differed markedly—Jev loves to praise itself, luna is the most silent, Gemini is the only one that talks about the opponent, and Haiku is all about luck. AI also practices emotion management, never displaying "played badly." The questioning method itself has a huge impact—merely changing the option order can shift a share by 24 percentage points. The experiment does not answer whether AI has real feelings; it measures only choice behavior given information.