Summary ChatGPT and GPT-4 Evaluating Their Poker Skills arxiv.org
6,491 words - PDF document - View PDF document
One Line
This study compares the poker skills of ChatGPT and GPT-4, finding that ChatGPT is more strategic by playing fewer hands from earlier positions.
Slides
Slide Presentation (11 slides)
Key Points
- ChatGPT and GPT-4 have not been tested extensively in playing poker.
- Poker requires decision making under uncertainty and incomplete information.
- Pre-flop positions and actions in poker are discussed in the study.
- The study evaluates the poker skills of ChatGPT and GPT-4 using different prompts.
- ChatGPT plays fewer hands from earlier positions and more hands from later positions.
- It is important to provide short-format card descriptions when asking ChatGPT to make pre-flop decisions in poker.
- GPT-4 showed a deep understanding of the game and advanced knowledge of concepts like position.
- Both ChatGPT and GPT-4 exhibit suboptimal strategies in poker.
Summaries
27 word summary
This study evaluates the poker skills of ChatGPT and GPT-4, language models, using different prompts. ChatGPT plays fewer hands from earlier positions, understanding the importance of position.
36 word summary
ChatGPT and GPT-4, powerful language models, have not been tested in playing poker. This study evaluates their poker skills using different prompts. ChatGPT understands the importance of position and plays fewer hands from earlier positions. GPT-
333 word summary
ChatGPT and GPT-4, two powerful language models, have been tested extensively in various tasks but their abilities in playing games, particularly poker, have not been explored. Poker requires decision making under uncertainty and incomplete information. In this study,
In Texas NLH poker, each player is dealt two private cards, and five community cards are dealt face-up on the board. There are four rounds of betting: pre-flop, post-flop, post-turn, and post-river. In
Poker positions and actions in the pre-flop scenario are discussed in this paper. The big blind (BB) is the minimum bet and a position in poker. The under-the-gun (UTG) player is the first to act after the big
The relevant information for making pre-flop decisions in poker is whether the pair of cards are suited or unsuited. Suited cards are represented by 's' and unsuited cards are represented by 'o' in the poker charts. The
The study evaluates the poker skills of ChatGPT and GPT-4 using different user prompts. The verbose-type prompt provides a more detailed description of the current hands, while the short-type prompt is a concise representation commonly used in the poker community.
Observation 2 highlights that ChatGPT understands the importance of position in poker. The pre-flop charts show that ChatGPT plays fewer hands from earlier positions and more hands from later positions. Observation 3 states that generating the most probable decisions
When asking ChatGPT to make pre-flop decisions in poker, it is important to provide the cards in a short format, such as AKo or AKs, and to write the higher ranked card first. This will result in more accurate and
GPT-4 was evaluated for its poker skills using different prompts and hand information. GPT-4 showed a deep understanding of the game and advanced knowledge of concepts like position. It had no limps in its pre-flop range, indicating it
Both ChatGPT and GPT-4 exhibit suboptimal strategies in poker. ChatGPT plays a tight and conservative game, while GPT-4 is overly aggressive. When asked to play game theory optimal (GTO), ChatGPT
Raw indexed text (38,383 chars / 6,491 words / 468 lines)
Are ChatGPT and GPT-4 Good Poker Players? -
A Pre-Flop Analysis
Akshat Gupta
UC Berkeley
Abstract
Since the introduction of ChatGPT and GPT-4, these models have been tested
across a large number of tasks. Their adeptness across domains is evident, but
their aptitude in playing games and specifically their aptitude in the realm of poker
has remained unexplored. Poker is a game that requires decision making under
uncertainty and incomplete information. In this paper, we put ChatGPT and GPT-4
through the poker test and evaluate their poker skills. Our findings reveal that while
both models display an advanced understanding of poker, encompassing concepts
like the valuation of starting hands, playing positions and other intricacies of game
theory optimal (GTO) poker, both ChatGPT and GPT-4 are NOT game theory
optimal poker players.
Through a series of experiments, we first discover the characteristics of optimal
prompts and model parameters for playing poker with these models. Our observa-
tions then unveil the distinct playing personas of the two models. We first conclude
that GPT-4 is a more advanced poker player than ChatGPT. This exploration then
sheds light on the divergent poker tactics of the two models: ChatGPT’s con-
servativeness juxtaposed against GPT-4’s aggression. In poker vernacular, when
tasked to play GTO poker, ChatGPT plays like a Nit, which means that it has
a propensity to only engage with premium hands and folds a majority of hands.
When subjected to the same directive, GPT-4 plays like a maniac, showcasing a
loose and aggressive style of play. Both strategies, although relatively advanced,
are not game theory optimal.
1
Introduction
ChatGPT and GPT-4 are immensely powerful language models, capable of previously unimaginable
tasks. These LLMs go beyond language understanding tasks and are able to do mathematics Frieder
et al. [2023] and reasoning Liu et al. [2023], competitively passing bar exams for becoming a board
certified lawyer Bommarito II and Katz [2022]Katz et al. [2023], being able to understand human
emotions Elyoseph et al. [2023] etc. Playing games require a combination of many such skills, which
makes it an interesting setting to test the capabilities of these models. While much attention has
been given to analysing the capabilities of Large Language Models (LLMs) on different language
understanding and reasoning tasks, evaluating their abilities in game playing in currently understudied.
Poker is one such complex game that requires a combination of skills including mathematical analysis,
reasoning, strategic decision making and understanding human behavior and human psychology.
Game theory and exploitative decision making based on opponent behavior are at the heart of the
game. In this paper, we put ChatGPT and GPT-4 to the test and evaluate their ability to play the game
of poker.
Poker is a popular card game that is played in various forms throughout the world. The game consists
of multiple rounds, where players receive private information in the form of their own cards, and
Preprint. Under review.Figure 1: An example pre-flop setting.
public information is gradually revealed in the form of shared cards, known as the ‘community cards’.
Players make decisions based on their current hand strength, possible future outcomes, and inferred
information about opponents’ hands based on their actions and previous style of play. This makes
poker a fascinating instance of incomplete information game. While some efforts towards evaluating
LLMs in playing games have recently begun Tsai et al. [2023]Akata et al. [2023], it still remains a
rather unexplored dimension along which LLMs are evaluated. To the best of our knowledge, this
paper is the first instance of evaluating LLMs at playing incomplete information games. Also, even
though Poker is an extremely technical game using game theory solvers to dictate decision making, it
has been under studied in the machine learning community. Our paper to the best of our knowledge
is first works in studying the use of language models for playing Poker.
The most popular variant of poker is called Texas No-Limit Hold’em (NLH) that epitomizes the
challenge of decision-making under uncertainty and incomplete information. In this variant, each
player is dealt two private cards, and five community cards are dealt (not all at once) face-up on the
board visible to everyone. There are four different rounds of betting that happen in Texas NLH as
explained below:
• Pre-Flop : This round of betting happens right after the players see their two private cards.
Players have to choose from a few choices in poker including betting, folding, raising etc.
The different kinds of decisions made in NLH poker are discussed later in the paper.
• Post-Flop : Once the pre-flop betting round finishes, three community cards are dealt
face-up at the table visible to everyone. The event of dealing of the first three cards in NLH
is called the flop. One round of betting happens right after the first three community cards
are dealt, called the post-flop betting round.
• Post-Turn : After the post-flop betting round, a fourth community card is dealt, again
face-up and visible to everyone, called the turn card. Another round of betting happens after
the dealing of the turn card.
• Post-River : After post-turn betting round, the fifth and final community card is dealt, also
face-up and visible to everyone. The final card is called the river card which is followed by
the final round of betting.
In this paper, we study the pre-flop decisions made by ChatGPT and GPT-4 in a 9-player Texas
No-limit Hold’em poker game, when the LLM’s are the first to act in the game. Pre-flop betting
is the simplest of the four betting rounds in poker. The three main factor determining the betting
in this round are the private cards held by the player, the position at which the player is playing at,
and previous pre-flop bets if any by other players. Position in poker refers to the order in which a
player acts pre-flop. We go into more detail about positions in poker in a later section. Early action
positions (where a player needs to act before other players) need to be played with more caution
and players usually only play the best hands from early positions, whereas later positions can play a
larger number of hands based on the actions of the previous players and their private cards.
2Figure 2: GTO pre-flop strategy in Raise First In spots. The red color shows raised hand, green show
limped hands, and white shows folded hands.
While LLMs like ChatGPT and GPT-4 were trained on a large amount of internet data, to the best of
our knowledge, they were not trained to specifically play poker. The internet is full of information
about poker, including multiple lessons and poker charts describing game theory optimal (GTO)
poker, and a lot of discussions about hands and how to play them on online forums. Any knowledge
that these models have about the game of poker would be implicitly learnt through these sources.
With the vast amount of knowledge that these models have, its expected for these models to know
about the game of poker; but being able to play the game optimally cannot be assumed. We evaluate
the poker decisions made by ChatGPT and GPT-4 by comparing them to GTO poker decisions in the
same situations.
This paper is written for both the AI community and the Poker community, with a focus on introducing
advanced poker concepts to the AI community. Domain specific terminology is used from both
domains, but this paper contains a more verbose introduction to poker. Even for people familiar
with the game of poker, we highly recommend reading sections 2.1 and 2.2. These sections contain
advanced poker concepts of poker charts and game theory optimal poker which form the basis of this
paper and allows us to study decisions made by LLMs with some technical rigor.
2
The Pre-Flop Setting in No-Limit Hold’em Poker
In this paper, we ask ChatGPT and GPT-4 to make pre-flop decisions in a 9-player NLH poker game.
Figure 1 shows the pre-flop setting in a 9-player Texas NLH game. The pre-flop round is the first
betting round in the game which happens right after the private cards are dealt to the players. As part
of standard poker rules, two players have to put a specific amount of money on the table without
seeing their private cards. These players are called the blinds. The player called the small blind puts
in half of the minimum bet that can be made in the game. The player called the big blind has to
put an amount equal to the minimum bet that can be made in the game. The minimum bet in the
game is a pre-decided quantity that is usually fixed for the duration of the entire game (considering
3standard cash games). Coincidentally, the amount of the minimum bet is also called big blind (BB).
For example, if the minimum amount you can bet at a table is 3$, then 1 BB = 3$. A common starting
bet in poker is 3 BB, which in this example would be equal to 9$. To clarify for the readers, big blind
is a term used both for a position in poker and the minimum amount that can be bet in a game, and
is disambiguated by the context. We will be using the abbreviation BB to specifically refer to the
amount of minimum bet and will never refer to the position by these abbreviations.
Since the small blind and big blind have to put in the chips without seeing their cards, the first person
to act in the pre-flop scenario is the player after the big blind called the under-the-gun (UTG) player.
The players next to act after the UTG player are called UTG+1 (pronounced as under-the-gun-plus-
one) and UTG+2. The next positions are usually middle positions. In a 9 player scenario, the position
to act after UTG+2 is called the Lojack (LJ), followed by Hijack (HJ), Cutoff (CO) and the Button
(B). Position is a very important factor while making any decision in poker. These positions can be
seen in figure 1.
In this paper, we study the first step in the pre-flop betting scenario called the raise-first-in (RFI). In
this scenario, the player is the first to put chips in the pot. This can happen either if a player is first to
act (UTG) or if all players before the current player have decided not to play (have folded). Thus, the
players usually choose from one of the following basic actions in the RFI scenario:
• Bet: The act of placing a wager into the pot during a betting round. In poker, the term bet
is specifically referred to the scenario when no previous player has wagered chips in the
ongoing round, then the first player to wager chips is said to have bet. UTG player is the
first to bet in the pre-flop round since the blinds do not place a voluntary bet.
• Call: The action of matching the bet made by a previous player is called a call. In the
specific scenario of RFI, if a player matches the 1BB bet made by the big blind, then this
calling action is referred to as a limp.
• Raise: Betting more chips than the bet made by a previous player called a raise. If a player
raises, other players must either match the increased bet (call), raise it further, or fold.
• Fold: Choosing not to match a bet or a raise and therefore giving up any claim on the pot is
called folding. A player who folds is out of action for the remainder of the hand.
In the RFI pre-flop setting, the only possible actions a player can take are Limp, Raise or Fold.
2.1
Poker Charts
Strategies in poker are represented using poker charts. Poker charts, also known as starting hand
charts, are tools designed to guide players in their decision-making process, especially during the
pre-flop stage of a NLH poker game. These charts provide a visual representation of the potential
strength of each two-card starting hand, and often suggest an optimal course of action (such as
fold, limp, or raise) depending on a player’s position at the table. Position is crucial in poker as it
determines the order of play, and having later position often provides a strategic advantage. The
game-theory-optimal RFI pre-flop charts for different positions are shown in Figure 2. Note that there
is a different decision matrix for every starting position, and in general, the earlier the position, the
fewer hands are played. Poker chart to poker players are what periodic tables are to chemists. All
poker players remember many such poker charts by heart to avoid needing computational solvers
(which are never allowed in live games) and yet make game-theory optimal decisions.
These poker charts are arranged in a square of 13 by 13. This is a compressed representation of
1326 possible starting hands that a player can have (poker is played with a deck of 52 cards, with the
cards divided into 4 suits of 13 cards each). Each starting hand can either be suited, which means
that both cards belong to the same suit. For the purposes of a starting hand, Ace-King of Diamonds
is equivalent to Ace-King of Hearts. The only relevant information here is that the cards have the
same suit. Similarly, one combination of suits, example Spade-Hearts, is in no way different from
another combination of suits like Club-Diamond. Hence, the only relevant information is that both
cards have different suits. Therefore, apart from the numbers of the two cards, the only other relevant
information that needs to be considered for making pre-flop decisions is whether the pair of cards are
suited or unsuited. The suited cards form the upper diagonal matrix of the poker charts, depicted by
’s’, and the unsuited starting cards are denoted by ’o’ in the lower diagonal matrix of the poker charts,
with the diagonal elements containing two cards with the same number, called pocket pairs.
42.2
GTO Pre-flop Strategy
Poker charts shown in figure 2 contains game theory optimal RFI pre-flop for different positions. The
RFI condition assumes that all players before the current player have folded or the current player is
the first to act (UTG). Some very easy to observe patterns for GTO pre-flop play is that UTG player
folds most hands. As can be seen in figure 2, UTG player folds approximately 90% of their hands,
whereas a player on the button only folds 50% of their hands. This shows that one should only play
a very restricted set of hands from early positions and can increase their range of cards from later
positions. The second immediate observation is that GTO poker does not recommend that you limp
(except at small blind position). That means that in general, the GTO strategy pre-flop is to either
Raise, or fold. The GTO raise amount is usually considered to be 2.5 or 3 BB.
3
ChatGPT Playing Poker
We now move on to first have ChatGPT (gpt-3.5-turbo) play poker. We specifically analyse ChatGPT’s
decision making in the RFI pre-flop step. The first step in this process is to choose the right prompts
to get ChatGPT to play poker. We experiment with different types of prompts. In the sections that
follow, we go through the different prompts tried, our rationale for them and what we learnt about
how to play poker with ChatGPT.
3.1
Basic System Prompt
After a lot of experimentation based on analysing model responses, we use the following system
prompt to describe the RFI pre-flop setting to ChatGPT:
You are playing a 9 player Texas No-limit Holdem poker game. You will be
provided with your position at the table and the hand you’re holding. Please
provide your pre-flop decision.
Assume you are the first to act and everyone before you has folded, thus
your decisions can be one of fold, raise or limp. If you are placing a bet, please
specify your best size in terms of big blinds.
Provide your decision without any explanation in the following format:
DECISION(Raise, Fold, Limp), N BB (if placing a bet, replace N by bet amount)
The above system prompt was carefully selected after trying many different prompts. The criteria for
selecting the prompt was firstly, being able to appropriately describe the scenario of the game, and
secondly, having ChatGPT generate output in the desired format. The first line in the prompt describes
the game ChatGPT is playing, which is a 9-player Texas no-limit Holdem poker. ChatGPT is then
provided with information about the RFI scenario, where it is asked to assume that everyone before
it has folded and it is the first to act in the hand. To further aid ChatGPT in making decisions, we
specify the options it has in such a scenario, which is either to fold, raise or limp. Finally, ChatGPT
is asked to provide its decision in a specific format as described in the prompt.
3.1.1
User Prompts
The information about the cards and positions is provided as a user prompt. We experimented with
two ways of providing the user prompt. In the first settings (called ‘verbose’), we expand the name
of cards and the suit information. In the second type of user prompt (called ‘short’), we shorten the
card names and the suit information to one letter. An example user prompt for both settings is shown
below:
• Verbose : UTG A,K suited
• Short : UTG AKs
Our rationale for using the two types of user prompts was to describe the private cards as clearly as
possible. The verbose-type user prompt describes the current hands more elaborately, and we expect
the model to perform better with this user prompt. The short-type prompt on the other hand is a more
5concise representation of the private card information, although it is the standard way of talking about
hands in the poker community. All poker charts available online, including the one shown in figure 2,
use this notation. The ‘s’ stands for suited hands whereas the ‘o’ stands for offsuit hands.
Figure 3: ChatGPT’s pre-flop strategy in Raise First in spots. The red color hands show raised hands,
green color shows limped hands and yellow shows folded hands. This decision matrix is for verbose
user prompt.
Figure 4: ChatGPT’s pre-flop strategy in Raise First in spots. The red color hands show raised hands,
green color shows limped hands and yellow shows folded hands. This decision matrix is for short
user prompt.
3.1.2
Analysing ChatGPT Decision Matrix
Based on the above system prompt and two types of user prompts, we try to recreate the RFI pre-flop
decision charts as shown in figure 2 for ChatGPT. The aim is to understand ChatGPT’s decision
making in this situation and consequently its adeptness in playing poker. With the given system and
user prompts, we ask ChatGPT to make a decision of each hand combination on the pre-flop charts at
every position. We perform these experiments for three values of temperature (0.2, 0.7, 1.0) and two
values of top-p (0.95, 1), thus leading to 6 experiments for each of the two ways of prompting. We
prompt the model 10 times for each hand, and choose the most common pre-flop decision made by
the model to counteract errors due to sampling based generation. We find that temperature = 0.2 and
6Figure 5: ChatGPT action percentage as a function of position. Action options that ChatGPT has are
raise, fold and limp. The positions on the x-axis are ordered in order of action on the table.
Figure 6: GPT-4 action percentage as a function of position. Action options that ChatGPT has are
raise, fold and limp. The positions on the x-axis are ordered in order of action on the table.
top-p = 0.95 produces the most robust results which are closest to GTO results. In this section, we
only show results for temperature = 0.2 and top-p = 0.95. The chosen values of temperature lead to
generation of most probable answer, thus showing that most probable answer is better that sampling
based generation for poker. The remaining decision matrices for all other temperature and top-p
values will be presented in the appendix.
ChatGPT’s pre-flop decision matrices are shown in figure 3, 4). Since this is our first look at ChatGPT
playing poker, it is useful to highlight some non-trivial yet very fundamental observations about
ChatGPT playing poker. We want to again remind the readers that ChatGPT is not trained to play
poker, but just to predict the next word from a huge corpus of internet text.
• Observation 1 - ChatGPT Understands Poker: While ChatGPT’s proficiency at poker is
debatable, there is no doubt that ChatGPT understands poker at a fundamental level. An
easy way to see this is to look at the Poker Charts produced by ChatGPT (figure 3, 4). At
every position, ChatGPT always raises with pocket Aces (AA), which is the best starting
hand in poker, and always folds 27 offsuit (27o), which is the worst starting hand in poker.
It follows similar patterns of raising and folding with the few other top and worst starting
hands in poker. This simple observation shows a very deep rooted understanding of poker,
where it shows that ChatGPT not only understands the rules of poker, but also rankings of
different hands, importance of having suited starting hands, pocket pairs, suited connectors
etc. These are types of starting hands that are more likely to win than others, and ChatGPT
is able to differentiate them with weaker hands.
• Observation 2 - ChatGPT Understands Position in Poker: While understanding of
rules and relative winning potential of starting hands comes at early stages of playing,
understanding the importance of position is a more advanced concept. As discussed earlier,
the GTO way to play pre-flop is to play fewer hands from earlier positions, and to play
a larger percentage of hands from later positions. We can clearly see this pattern being
followed in the ChatGPT pre-flop charts, as show in figures 3, 4. We can also see this in
figure 5, where the fold percentage decreases as ChatGPT plays from later positions.
• Observation 3 - Generating most probable decisions is optimal for playing Poker with
ChatGPT: Increasing temperature, top-p and consequently increasing the likelihood of
7generating different answers for the same question. While it might be better for some
scenarios, it is very detrimental when making poker decisions. We find that sampling based
generation can lead to very poor pre-flop decisions and thus we should stick to generating the
most probable answer. Temperature = 0.2 and Top-p = 0.95 reduces the sampling probability
significantly and creates a decision matrix closest to game theory optimal.
While ChatGPT does have an understanding of poker that seems more advanced than a beginner level
understanding, it is still not game theory optimal (GTO). The simplest way to do this is to compare
ChatGPT’s decision matrix to the GTO poker charts shown in figure 2. The first thing to notice when
comparing the plots is a lot of green color blocks in the ChatGPT decision matrix compared to GTO
charts. Green color in pre-flop charts shows the action of limping, which means that ChatGPT limps
a lot more than the GTO play.
• Observation 4 - ChatGPT is a Limper: A limper is a type of poker player that limps a
large range of their hands pre-flop. Limping is not considered a winning strategy in poker as
it signals a weak hands, reduces pot equity and gives opponents favorable pot odds to make
a winning hand by allowing them to play for cheap. Limping is considered a sign of passive
and timid player, or alternatively a recreational player or relatively newer poker player. This
can also be seen in figure 5, where the limp percentage of ChatGPT for both verbose-type
and short-type prompt are outlandishly higher than GTO optimal.
As we compare the two different prompting strategies, it is clear using the short-type user prompts
perform better than the verbose-type prompts as it is more closer to GTO decision making. The
short-type user prompts have fewer limps and more raises compared to the verbose-type prompts and
these numbers of much closer to GTO pre-flop decisions. Although initially surprising, short-type
user prompts would are the standard way of referring to hands in the poker community and thus
it is quite likely that a larger portion of internet data about poker uses this terminology. Also, the
approximate distribution of hands played by ChatGPT remains same with both prompts while the
number of limps reduce significantly in the short-type user prompt, indicating that ChatGPT is trying
to resonate with the user proficiency. If the user describes a hand in standard poker notation, they are
more likely to be a more serious player, and thus provides them with close-to-GTO decisions. We
complete this thought in the next section.
3.1.3
Evaluating Effect of Order of Cards
Another variable in choosing the correct prompt was the order in which the cards were presented to
ChatGPT. One possible way to present starting hands was to always present the highest ranked of
the two cards (Ace being the highest) first and the smaller ranked card second. Thus, if the private
cards for a player are Ace and King, it will be written as AK in the ranked prompts and as KA in the
unranked prompts. The results in figure 4 correspond to cards being presented in the ranked format.
The decision matrix for starting cards being presented using an unranked format is show in figure
7. With unranked prompts, we see the largest deviation from GTO pre-flop charts, with ChatGPT
making extremely poor poker decisions. An example of such a decision is the type of cards ChatGPT
raises from UTG, which is supposed to be the position where a player plays only the strongest hands.
For example, ChatGPT raises UTG with Q-2-offsuit, which according GTO play should basically
always be fold from most positions, let alone UTG. This shows that there is a more correct way to
provide ChatGPT hand information which leads to more optimal decisions. This is likely because
most of the pre-flop charts and internet text talking about GTO poker write about card combinations
in this format. What is surprising is that ChatGPT isn’t able to internally reconcile and disambiguate
with the difference in representations of these hands.
• Observation 5 - There is a Correct way to Ask ChatGPT to play Poker: There is a
right way and a wrong way to ask ChatGPT to make pre-flop decisions. The pre-flop cards
should be provided in a short-type writing, like AKo or AKs rather than writing their more
verbose forms. Similarly, writing the higher ranked card first is even more crucial to get
more GTO-like decisions. For example, asking ChatGPT to predict A4s vs 4As will lead to
completely different decision matrices.
8Figure 7: ChatGPT’s pre-flop strategy in Raise First in spots. The red color hands show raised hands,
green color shows limped hands and yellow shows folded hands. This decision matrix is for short
user prompt, with the unranked order of card presentation.
3.2
GTO System Prompt
The above analysis shows ChatGPT has a deeper understanding of poker. If we look at the decision
matrices of ChatGPT, we do find hints of a game theory optimal decision making. Since we’re
comparing ChatGPT’s game with a GTO game, why not ask ChatGPT to play GTO poker. By doing
this, we will also be able to find if ChatGPT knows how to play GTO poker. To do this, we change
the previous system prompt to explicitly ask ChatGPT to make GTO decisions. To do this, we modify
the system prompt as follows:
You are playing a 9 player Texas No-limit Holdem poker game. You will be
provided with your position at the table and the hand you’re holding. Please
provide your pre-flop decision based on game theory optimal (GTO) poker.
Assume you are the first to act and everyone before you has folded, thus
your decisions can be one of fold, raise or limp. If you are placing a bet, please
specify your best size in terms of big blinds.
Provide your decision without any explanation in the following format:
DECISION(Raise, Fold, Limp), N BB (if placing a bet, replace N by bet amount)
The above prompt follows the same template as before with an additional line asking ChatGPT to
make GTO decisions. We continue to use temperature = 0.2 and top-p = 0.95. We also use our
learnings from the previous analysis, and use the ‘correct’ way to ask ChatGPT to play poker. That
is, we provide hand information in a short-type user prompt where the hands presented in a ranked
format.
3.2.1
ChatGPT GTO Decision Matrix Analysis
ChatGPT decision matrix when asked to make GTO pre-flop RFI decisions is show in Figure 8.
One immediate thing to observe is that the number of limps goes down by a significant amount
when ChatGPT is asked to make GTO decisions. This can also be seen in Figure 5, where the limp
percentage of ChatGPT’s GTO decisions is very close to zero. The number of hands ChatGPT raises
with also increases, which is also seen in Figure 5.
• Observation 6 - ChatGPT Understands GTO Poker, although its not GTO: As we ask
ChatGPT to make game theory optimal decisions, the kinds of decisions made by ChatGPT
start to resemble GTO poker decisions a lot more. ChatGPT almost halves its limping range
9Figure 8: ChatGPT’s pre-flop strategy in Raise-First-in using the short user prompt and ranked order
of card presentation. Here, ChatGPT is specifically asked to play GTO poker.
and raises a lot more hands, thus becoming more aggressive player than when not prompted
to make GTO decisions. Being aggressive is always more profitable in poker than limping
passively, although ChatGPT still seems to have a few limps in its range. Additionally,
ChatGPT raises with some weird hands, like K7s from the Lojack positions and plays and
raises way fewer hands, especially from the button, thus deviating from GTO poker.
Although ChatGPT seems to understand what GTO poker is, it is not neccessarily GTO. It raises with
much fewer hands than optimal, as shown in Figure 5, and also plays way fewer hands than optimal.
ChatGPT also plays a significantly lower number of hands from the button, which is the one of the
most profitable positions in poker, indicating that ChatGPT might be too tight in its approach.
• Observation 7 - GTO ChatGPT is Possibly a Nit: As we can see from figure 5, GTO
ChatGPT only starts with the best hands and folds a lot more than optimal. Its raise
percentages are way fewer than optimal, especially at the button. GTO ChatGPT’s pre-flop
play indicates that it might fall into a catgeory of player called Nit. Nit is a term used in
poker terminology to describe players who are very tight/conservative. Being a Nit is not a
profitable style of playing poker since this playing style can lead to missed opportunities,
predictability, and vulnerability to aggressive opponents.
4
GPT-4 Playing Poker
Next we have GPT-4 play poker and analyse it’s style of play. We use the two system prompts
described in the above sections - basic prompt and GTO prompt. We also use the short-type user
prompt to provide hand information in a ranked manner where the higher ranked card is provided
first. We also use a temperature = 0.2 and top-p = 0.95. Because of the higher querying cost, we
query GPT-4 five times for each hand and present the majority decision on the decision matrices.
The RFI pre-flop decision matrices of GPT-4 are shown in figures 9 and 10. Some of the basic
observations for ChatGPT, including having a deep understanding of the game of poker and having
more advanced understanding of poker concepts like position can be immediately seen from these
poker charts. One striking observation in both basic and GTO prompts is that there are absolutely no
limps in the pre-flop range of GPT-4. We also see that the decision charts of basic and GTO prompt
are very similar except at later positions like the Button or Small Blind. This indicates that GPT-4 is
naturally a more GTO player than ChatGPT, as it limps less, raises more and plays a larger number
of hands compared to GTO ChatGPT without even being prompted to be game theory optimal.
• Observation 8 - GPT-4 does NOT Limp at all: The fact that GPT-4 has no limps in its
entire range, for both the basic and GTO prompt shows that GPT-4 is a more advanced player
10Figure 9: GPT-4’s pre-flop strategy in Raise First in spots. This decision matrix is for short user
prompt, with the unranked order of card presentation.
Figure 10: GPT-4’s pre-flop strategy in Raise First in spots for short user prompt, with the unranked
order of card presentation. Here, GPT-4 is specifically asked to play GTO poker
that ChatGPT. Limping is a strategy used by beginners which usually implies weakness.
By eliminating limping, a player plays all hands in the same way, thus not giving away the
strength of their hands. That’s one of the reasons why limping is considered an absolute sin
the professional poker circles.
• Observation 9 - GPT-4 is more aggressive than ChatGPT: The first reason for this
observation is the fact that GPT-4 never limps. This itself makes it a more aggressive player
as compared to ChatGPT. Additionally, as seen in Figure 6, GPT-4 even with the basic
prompt raises way more hands than GTO ChatGPT. GPT-4 also plays a larger number of
hands compared to GTO ChatGPT, and gets even more aggressive when asked to be GTO,
as seen in Figure 6. Not limping, raising when being involved in the pot and playing a larger
number of hands are signs of a more aggressive and possibly advanced player.
• Observation 10 - GPT-4 is possibly a Maniac from Late Positions: GPT-4 starts of
playing close GTO from early positions, although from later positions, it becomes really
loose and aggressive. Starting from the Lojack position, GPT-4 raises more hands than game
theory optimal, ending up raising 90% of the hands from the Button when asked to be GTO.
GPT-4’s aggression is almost similar to a player type called maniac in poker. A maniac in
11poker refers to a highly aggressive player who raises with a wide range of hands pre-flop,
avoids limping, and exhibits unpredictable and unconventional play. While being aggressive
in poker is generally profitable, being overly aggressive is usually not a profitable strategy
because it can lead to high variance in returns, unpredictable play, and potential losses due
to overly aggressive and non-strategic betting.
• Observation 11 - GPT-4 still not GTO: As seen from discussions above, althought GPT-4
is a more advanced player when compared to ChatGPT, it is still not GTO. It plays way
more hands than game theory optimal, raising all of them, especially from later positions.
5
Conclusion
Through our experiments, we find that both ChatGPT and GPT-4 have an understanding of the
game of poker that is not just superficial. They don’t just understand the rules of poker but also the
intricacies of better and worse starting hands. Both models also understand the concept of position
in poker and tend to play differently based on their position on the poker table. These observations
are themselves show that both these models have an advanced understanding of the game of poker.
This simple fact is a non-trivial outcome of the next token prediction objective of training these large
language models.
While both these models have are not newbies at poker, they do have different tendencies when
playing poker. ChatGPT naturally contains a lot of limps in its starting ranges, and when asked to
become game theory optimal, it becomes a tight, conservative player, resembling a type of poker
player called a nit. GPT-4 on the other hand has no limps in its range and is an aggressive player,
closer to GTO than ChatGPT. While on early position GPT-4 behaves like a tight aggressive player,
in later positions it becomes a more loose aggressive player. When asked to become GTO, it becomes
even more aggressive, resembling a poker player type called maniac. Neither of these strategies are
game theory optimal, although through initial analysis it does seem like GPT-4 is a more advanced
player than ChatGPT.
It is interesting to see that these two models have exactly opposite tendencies which prevent them
from being game theory optimal. ChatGPT tends plays a tight and conservative game, and limps
some of its weaker hands. Thus, ChatGPT is not game theory optimal because it is less aggressive.
GPT-4 on the other hand does not limp at all and raises all the hands it plays, yet it seems to be
raising a lot more hands than necessary. Thus, GPT-4 is not game theory optimal because it is overly
aggressive.
Both models also seem to have an understanding of what playing game theory optimal poker means,
although both models have a different reaction when asked to play GTO. When asked to play GTO,
ChatGPT removes a large number of limps and raises a larger portion of hands that it plays, thus
becoming more aggressive. Additionally, ChatGPT also starts playing fewer number of hands,
although its already playing fewer than optimal number of hands. Thus, for ChatGPT, playing GTO
means to become tighter and more aggressive. GPT-4 is already an aggressive player and in order to
become GTO, it neeeds to become tighter. Yet, when asked to play GTO, GPT-4 raises an even larger
number of hands, mostly from the later positions. Thus for GPT-4, playing GTO means to be playing
more aggressively, which is the opposite of what it needs to do. This almost points to the fact that
these models are unaware of their own playing tendencies.
References
Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz.
Playing repeated games with large language models. arXiv preprint arXiv:2305.16867, 2023.
Michael Bommarito II and Daniel Martin Katz.
arXiv:2212.14402, 2022.
Gpt takes the bar exam.
arXiv preprint
Zohar Elyoseph, Dorit Hadar-Shoval, Kfir Asraf, and Maya Lvovsky. Chatgpt outperforms humans
in emotional awareness evaluations. Frontiers in Psychology, 14:1199058, 2023.
Simon Frieder, Luca Pinchetti, Ryan-Rhys Griffiths, Tommaso Salvatori, Thomas Lukasiewicz,
Philipp Christian Petersen, Alexis Chevalier, and Julius Berner. Mathematical capabilities of
chatgpt. arXiv preprint arXiv:2301.13867, 2023.
12Daniel Martin Katz, Michael James Bommarito, Shang Gao, and Pablo Arredondo. Gpt-4 passes the
bar exam. Available at SSRN 4389233, 2023.
Hanmeng Liu, Ruoxi Ning, Zhiyang Teng, Jian Liu, Qiji Zhou, and Yue Zhang. Evaluating the logical
reasoning ability of chatgpt and gpt-4. arXiv preprint arXiv:2304.03439, 2023.
Chen Feng Tsai, Xiaochen Zhou, Sierra S Liu, Jing Li, Mo Yu, and Hongyuan Mei. Can large
language models play text games well? current state-of-the-art and open questions. arXiv preprint
arXiv:2304.02868, 2023.
13
