Almost every poker product now advertises AI. Some of it is a genuine machine-learning model, some is a precomputed lookup table with a new label, and some is a chatbot that will confidently invent a range that does not exist. This guide separates them using published research instead of product copy, so you can tell what a poker AI tool is actually doing before you pay for it.
Disclosure: This guide is published by GTO Gecko, which sells poker study software including a machine-learning explanation feature discussed below. We keep general claims and product-specific claims separate, and we cite the research behind both.
What Is Poker AI, Exactly?
Four different technologies share the label: game-theory solvers, self-play neural agents, large language models, and machine-learning explanation layers. Only the first two produce strategy. The third talks about it, the fourth explains it, and knowing which one a product runs tells you exactly what to trust it for.
| Type | How it works | What it is genuinely good at |
|---|---|---|
| Game-theory solver | Counterfactual regret minimization and its variants, iterating toward an approximate equilibrium inside a game tree you define | Producing a strategy baseline for one clearly specified spot |
| Neural-net agent | Deep networks plus search, trained largely through self-play, playing a full game rather than one node | Playing entire hands well, research milestones |
| Language model | Predicts likely text from patterns in a very large training corpus | Explaining concepts, summarizing, drafting notes and study plans |
| Explanation layer | Fits a supervised model to solver output, then attributes each prediction back to input features | Turning a frequency into a reason you can recall at the table |
Products blur these categories in their marketing, and the blur is where money gets wasted.
The Poker AI Milestones That Are Actually Published
Five published systems cover the whole history of superhuman poker AI: Cepheus (2015), DeepStack (2017), Libratus (2017), Pluribus (2019), and ReBeL (2020). All five were research programs, not products, and none has ever been sold to players.
| Year | System | Team | Published result |
|---|---|---|---|
| 2015 | Cepheus | University of Alberta | Heads-up limit hold’em declared essentially weakly solved in Science, using the CFR+ algorithm |
| 2017 | DeepStack | Alberta and collaborators | Defeated professional players over 44,000 hands of heads-up no-limit with statistical significance |
| 2017 | Libratus | Carnegie Mellon | Beat four leading pros over 120,000 hands, finishing up $1,766,250 in chips |
| 2019 | Pluribus | Carnegie Mellon and Facebook AI | First superhuman result in six-player no-limit hold’em |
| 2020 | ReBeL | Facebook AI Research | General self-play framework reaching superhuman heads-up no-limit with far less poker-specific knowledge |
The Libratus match ran from January 9 to 30, 2017 at Rivers Casino in Pittsburgh. Brown and Sandholm, writing in Science in 2018, described the margin as statistically significant rather than a run of luck.
Pluribus is the better data point for judging modern claims about compute. Its blueprint strategy trained in eight days on 12,400 core hours, and it ran on 28 CPU cores during live play. Libratus, two years earlier, needed roughly 15 million core hours and 1,400 cores. Poker AI got dramatically cheaper, which is why consumer study tools became viable.
What none of these systems did
Not one was built to coach a human. They were built to win, and gave no explanations, no range charts, and no feedback. The gap between an agent that plays well and a tool that teaches you is the entire product category below.
What Commercial GTO AI Tools Actually Sell You
Three architectures, sometimes combined in one product: precomputed solution libraries, on-demand solving, and machine-learning layers on top. Almost everything marketed as GTO AI is serving stored solver output rather than running a neural network. All three are legitimate, and the difference decides what you can actually study.
Precomputed solution libraries
The vendor solves thousands of configurations in advance, stores the output, and serves the matching entry. Queries are instant because nothing is computed while you wait. The limit is coverage: if your stack depth, rake model, ante, or bet size is missing, you get the nearest neighbor rather than your spot, and the interface does not always tell you which you received.
On-demand solving
You define the spot and the software solves it, in the cloud or on your device. This handles configurations no library anticipated, at the cost of waiting and of understanding the inputs. Our guide to using a poker solver covers the input mistakes that quietly invalidate results, and our roundup compares tools in both camps.
Machine-learning layers on top
This is where the AI label is most often earned. A supervised model trained on solver output compresses, approximates, or explains it. The model does not discover strategy, it reorganizes strategy the solver already produced.
Machine Learning as an Explanation Layer
An explanation layer converts a solver frequency into a stated reason, by fitting a model to solver output and attributing each prediction back to its inputs. It cannot find strategy the solver did not produce. GTO Gecko’s Explain feature does this with gradient-boosted trees (XGBoost, Chen and Guestrin, 2016) and SHAP attributions (Lundberg and Lee, 2017).
The practical effect is that instead of reading "raise 63%, call 37%" and moving on, you get an attribution: this hand blocks the opponent’s continuing range, the board favors the preflop aggressor, the stack-to-pot ratio rewards building the pot now. Those are the sentences you can recall three weeks later at a table.
The limitation matters as much as the feature. An explanation layer explains the strategy it was trained on. If the underlying solution is wrong for your game because the rake, ante, or stack depth was misconfigured, you get a fluent, confident account of the wrong answer. Explanations do not validate inputs. You do.
Can ChatGPT Help Your Poker Game?
Yes for concepts, no for numbers. On PokerBench, an 11,000-scenario benchmark graded against GTO solver output, GPT-4 scored 53.55% and ChatGPT 3.5 scored 29.96% (Zhuang and colleagues, January 2025). Use a language model to explain ideas and organize notes, then check every range or frequency it produces against a solver.
The full benchmark, built from 1,000 preflop and 10,000 postflop spots, ranks as follows:
| Model | Setting | PokerBench accuracy |
|---|---|---|
| GPT-4 | Off the shelf, few-shot | 53.55% |
| Llama-3 70B | Off the shelf, few-shot | 39.16% |
| ChatGPT 3.5 | Off the shelf, few-shot | 29.96% |
| Llama-3 8B, fine-tuned | Zero-shot after fine-tuning | 78.26% |
The strongest general-purpose model agreed with solver ground truth on barely half the scenarios. The authors concluded that state-of-the-art language models underperform at optimal poker, and that fine-tuning helps substantially.
The failure mode has a specific shape. Ask a language model for a Button opening range at 40bb with antes and you get a clean, well-formatted list that looks exactly like a real range. It may include hands the solver folds and omit hands it opens, with nothing distinguishing the two. Fluency is not accuracy, and the model will not tell you when it is guessing.
| Task | LLM suitability | Why |
|---|---|---|
| Explaining what an equilibrium or a blocker is | Strong | Well-covered conceptual material with many correct sources in training data |
| Turning your session notes into a study plan | Strong | Structuring text you supply, not generating poker facts |
| Rewriting a solver output into your own words | Good, with the output pasted in | You provide the ground truth; the model only reformulates |
| Producing an opening or defending range | Weak | Precise frequencies are memorized approximations at best, invented at worst |
| Giving exact mixed-strategy percentages | Weak | No underlying computation, and no calibration signal on its own confidence |
| Evaluating a specific hand on a specific board | Weak | Requires solving a game tree, which the model is not doing |
The rule that follows is simple. Use a language model for language, and check every claim it makes about ranges or frequencies against a solver before you believe it. Our guide to understanding poker ranges covers what a correct range looks like structurally, which makes invented output easier to spot.
An AI-Assisted Study Workflow That Holds Up
Solver first, explanation second, chatbot third, trainer last. The order matters because only step two produces ground truth, and nothing downstream can repair a wrong configuration upstream.
- Pick one spot from your own play. Tag it during the session, not from memory afterward. Record format, positions, effective stack, antes, and every bet size.
- Get the strategy from a solver or a solution library. This is the only step that produces ground truth. Nothing later in the workflow can fix a wrong configuration here.
- Use an explanation layer to find the reason. Feature attributions tell you which variables moved the decision, so you learn a transferable principle rather than one memorized combo.
- Ask a language model to compress it. Paste in the solver output and the attribution, then ask for a two-sentence rule in plain language. The model is editing your material, not sourcing new facts.
- Drill it in a trainer until you can produce it cold. Reading an explanation is recognition. A preflop trainer or postflop trainer tests recall, which is where the learning actually happens.
- Sort your errors by cost, not by count. Track EV loss so you study the expensive mistakes first. Our EV loss leak finder guide explains why accuracy alone hides the leaks worth fixing.
Steps 2, 3, and 5 involve no language model at all. The chatbot has exactly one job here, and it is a writing job.
What AI Cannot Do for Your Game
Three things: exploit a specific opponent, manage the parts of poker that are not decisions, and cover spots nobody has solved.
| Limitation | Why it exists | What to do instead |
|---|---|---|
| Exploiting the player in seat four | Equilibrium is built to be unexploitable, so it also declines to attack an opponent’s mistakes | Learn the baseline, then deviate deliberately (GTO versus exploitative play) |
| Everything that is not a decision | Bankroll discipline, session length, and quitting while behind sit outside every tool listed here | Treat the mental game as separate work; perfect solver knowledge plus no tilt control still loses |
| Spots nobody solved | Libraries get built where demand justifies the compute. Full heads-up postflop GTO content is thin across the category, GTO Gecko included | Ask a vendor what is missing before subscribing, and treat a vague answer as an answer |
Is Using AI for Poker Allowed?
For off-table study, yes. During play, no, and the penalties include permanent bans and confiscated funds. GGPoker’s published security policy defines real-time assistance as "any external assistance which provides users with an unfair advantage over their opponents, influences their decision making in real-time," covering "software, apps, websites, physical and digital reference materials." Its bot definition names "artificial intelligence" directly, and its prohibited tools list covers RTA, bots, solvers, charts, and HUDs.
The practical line: a solver, a trainer, a chatbot, and an explanation model are all study tools when the poker client is closed, and all of them become prohibited assistance when a decision is pending. Some policies apply whenever the client is open, not only while the action is on you. Read our guide to solver and RTA rules, then check the current terms for every room you play.
Frequently Asked Questions
Can AI beat professional poker players?
Yes, in published research settings. Libratus beat four leading professionals over 120,000 hands of heads-up no-limit in 2017, and Pluribus reached superhuman performance in six-player no-limit in 2019. Those were research agents built to play, not products you can buy. No commercial study tool is one of them.
Is ChatGPT good for poker strategy?
It is useful for explaining concepts and organizing your study notes, and unreliable for anything numeric. On the PokerBench benchmark of 11,000 solver-graded scenarios, GPT-4 scored 53.55% and ChatGPT 3.5 scored 29.96%. Treat any range or frequency it produces as unverified until you check it against a solver.
What is the difference between a GTO solver and poker AI?
A solver computes an approximate equilibrium for one spot you define, usually with a regret-minimization algorithm. Poker AI is a broader label covering solvers, self-play neural agents, and language models. Most products advertising AI are serving precomputed solver output. Our solver versus trainer guide covers which you need.
Are AI poker tools allowed by poker sites?
For off-table study, generally yes. During play, no. GGPoker’s policy explicitly lists artificial intelligence, solvers, charts, and HUDs among prohibited tools, with permanent bans and fund confiscation as penalties. Close the poker client before you open study software, and verify the current rules for your room.
Can AI explain why a solver play is correct?
Increasingly, yes. Feature-attribution methods such as SHAP assign each input a measured contribution to a single prediction, which converts a raw frequency into a stated reason. GTO Gecko’s Explain feature uses this approach. The caveat is that the explanation describes the solution it was trained on, so a misconfigured spot produces a confident wrong explanation.
Will AI make poker unbeatable for humans?
It has not so far. Superhuman agents have existed since 2017 and human games continue, because the agents are research systems and using one during play is prohibited and detectable. What changed is study: equilibrium baselines that once cost millions of core hours now cost a subscription, so the skill gap increasingly reflects who studies well.
Use AI for the Job It Actually Does
Solvers give you the answer, explanation models give you the reason, language models help you write it down, and a trainer is what turns any of it into a decision you can make in four seconds. Pick one spot from your last session and run it through that order. When you are ready to drill, open the GTO Gecko trainer and start with a single configuration.
Sources
- Bowling, Burch, Johanson and Tammelin (2015), “Heads-up limit hold’em poker is solved” (Science)
- Moravčík et al. (2017), “DeepStack: Expert-level artificial intelligence in heads-up no-limit poker” (PubMed)
- Brown and Sandholm (2018), “Superhuman AI for heads-up no-limit poker: Libratus beats top professionals” (Science)
- Carnegie Mellon University news release on the Libratus Brains vs AI match
- Brown and Sandholm (2019), “Superhuman AI for multiplayer poker” (Science)
- Carnegie Mellon University news release on Pluribus
- Brown, Bakhtin, Lerer and Gong (2020), “Combining Deep Reinforcement Learning and Search for Imperfect-Information Games” (ReBeL) (arXiv)
- Zhuang et al. (2025), “PokerBench: Training Large Language Models to become Professional Poker Players” (arXiv)
- Lundberg and Lee (2017), “A Unified Approach to Interpreting Model Predictions” (SHAP) (arXiv)
- Chen and Guestrin (2016), “XGBoost: A Scalable Tree Boosting System” (arXiv)
- GGPoker Security and Ecology Policy
- GTO Gecko official product page