How Poker Became a Laboratory for Artificial Intelligence

From Deep Blue to Polaris, Cepheus, AlphaGo, DeepStack, Libratus, and Pluribus, games became laboratories for increasingly sophisticated forms of machine intelligence. But poker posed a distinctive challenge: how can a machine make rational decisions when crucial information is deliberately hidden? Today, the question is no longer whether AI can beat humans at poker, but whether humans can still be certain they are playing against humans.

13 – 19 min

Editorial Note — September 16, 2026

This article is a complete rewrite of an earlier AI-Talks.org piece on poker AI and the evolving relationship between artificial intelligence and poker. The original article was written at a very different stage in the development of AI and focused mainly on Polaris and the possibility of using artificial intelligence in online poker. Since then, the field has changed dramatically: Cepheus essentially solved heads-up limit Texas Hold’em; DeepStack and Libratus achieved superhuman performance in heads-up no-limit poker; and Pluribus extended those advances to multiplayer poker. AlphaGo is included as a broader landmark because its victory over Lee Sedol helped bring game-playing AI into global public consciousness and turn artificial intelligence into a major technological and cultural phenomenon. Revisiting the original text made clear that a conventional update would not be enough: its historical framing was incomplete, some factual claims required correction, and its discussion of ethics and legality no longer reflected the far more complex reality of modern poker AI, solvers, bots, and real-time assistance. I therefore chose not to revise the old article paragraph by paragraph, but to rewrite it entirely from scratch. The result is a fundamentally new article that treats poker not simply as a game machines learned to beat, but as a laboratory for artificial intelligence—a domain in which machines must reason under uncertainty, hidden information, strategic deception, and adversarial behavior. It also explores what these systems reveal about the distinction between intelligent performance and genuine understanding. What follows is a complete 2026 reconsideration of the subject.

---

When Machines Learned to Bluff: How Poker Became a Laboratory for Artificial Intelligence

In July 2007, two professional poker players sat down in Vancouver to face an unusual opponent.

Phil Laak and Ali Eslami were playing heads-up limit Texas Hold’em against Polaris, a computer program developed by researchers at the University of Alberta. The match came ten years after IBM’s Deep Blue had defeated world chess champion Garry Kasparov, but the problem confronting Polaris was fundamentally different.

The humans officially won. Across four duplicate matches, they won two, lost one, and drew one. The duplicate format was designed to reduce the influence of luck by exposing the human and machine teams to the same card sequences from opposite positions. Polaris lost the contest, but narrowly enough to make the real lesson difficult to miss: elite poker was no longer safely beyond the reach of machines.

One year later, Polaris returned in Las Vegas.

This time, the humans lost.

Polaris 2.0 defeated a team of professional players in the second Man-Machine Poker Competition, converting an intriguing experiment into a much more consequential demonstration.

The event attracted only a fraction of the attention generated by Deep Blue. Yet poker posed a problem that, in an important sense, looked more like the world outside games.

Chess asks a machine to choose well when the relevant state of the game is visible.

Poker asks:

What should you do when part of reality is hidden from you?

That distinction would turn poker into one of the most revealing laboratories in artificial intelligence.

Why Poker Is Not Chess

Chess, checkers, and Go are extraordinarily complex, but they belong to the class of perfect-information games. Both players can observe the current state of the board.

The difficulty lies in what comes next.

Poker introduces another kind of uncertainty: you do not even know the complete state of the game now.

Your opponents possess private cards. Chance determines future cards. Betting actions convey information, but that information is ambiguous because strategically competent players sometimes have an incentive to conceal, distort, or deliberately confuse what their actions reveal.

A large bet may indicate a powerful hand.

It may also indicate a weak hand masquerading as one.

Consequently, the poker-playing machine cannot merely ask which move maximizes its expected reward from a known position. It must maintain beliefs over possible hidden states of the game and continually update those beliefs.

What hands might the opponent hold?

What does the opponent’s previous behavior imply?

What does my own behavior reveal about my cards?

How predictable have I become?

And how should I act if my opponent is simultaneously trying to answer the same questions about me?

The challenge is not simply computation.

It is decision-making under imperfect information.

This is precisely what gives poker significance beyond poker. The real world rarely provides agents with a complete and reliable description of the state they inhabit. Negotiators do not know each other’s reservation prices. Firms do not know their competitors’ future strategies. Cybersecurity defenders cannot directly observe an attacker’s intentions. Investors possess different information. Political and military actors may deliberately transmit misleading signals.

Reality keeps some of its cards face down.

Can a Machine Bluff?

Poker adds another intriguing complication: bluffing.

A human bluff seems inherently psychological. I know my hand is weak. I want you to believe it is strong. I deliberately act in a way that encourages you to form an incorrect belief.

This sounds like behavior requiring intention, deception, perhaps even a model of another mind.

But a poker-playing computer does not need any of those things in their ordinary human sense.

Bluffing can emerge from the mathematics of optimal strategy.

Imagine a player who makes a very large bet only when holding an exceptionally strong hand. The strategy initially looks sensible, but it contains a fatal weakness: the player’s actions reveal too much information.

Once opponents discover the pattern, a large bet becomes almost equivalent to showing them the cards.

To prevent this, the player must sometimes make similar bets with weaker hands. The action then ceases to identify the underlying state so reliably.

At the opposite extreme, bluff too frequently and opponents can exploit that pattern by calling more often.

A difficult-to-exploit strategy therefore requires a controlled mixture of actions. Some strong hands are played aggressively. Some weaker hands must occasionally be played in similar ways. The precise mixture depends on the game state and strategic equilibrium.

In game-theoretic terms, bluffing emerges as part of a mixed strategy.

The important point is conceptual.

The machine does not have to feel deceptive to produce behavior that is strategically indistinguishable from deception.

Its bluff can simply be a probability.

But poker contains another layer of hidden information. The cards may be concealed, yet the player is still visible.

If a machine can learn to bluff, can it also learn to detect the subtle signals humans reveal while trying to hide what they know?

Can AI Read a Poker Face?

Can AI Read a Poker Face?

Poker hides the cards, but not the player.

Humans have always searched for tells—changes in gaze, posture, movement, breathing, or voice that may reveal what an opponent is trying to conceal. AI raises an obvious question: could a machine detect these signals more systematically than a human?

The technical ingredients already exist. Computer vision can track facial movement, gaze, blinking, head position, and posture. Audio models can analyze pitch, intensity, speaking rate, and prosody. Multimodal systems can combine these signals.

Poker players do leak information.

A 2013 study found that observers could infer aspects of professional players’ hand quality from arm movements, even when facial information was less useful. The way chips were moved appeared to reveal information the players were trying to hide.

Voice may also matter. In real-money no-limit Hold’em experiments, changes in voice pitch, together with physiological signals, were associated with high-stress situations and attempts to detect bluffing.

Computer vision has since been applied directly to poker footage. A 2022 dataset containing more than 31,000 facial images from professional games explored whether neural networks could identify patterns associated with bluffing.

But detecting a pattern is not the same as reading a mind.

Research on emotion recognition shows that facial expressions do not map neatly onto specific emotions. Context, culture, body movement, and individual habits all matter. A machine that detects a change in gaze has detected a signal—not necessarily fear, confidence, or deception.

Poker makes this harder because skilled players deliberately suppress or manipulate tells.

A 2026 preregistered computer-vision study found that facial and pose features had some predictive value within individual sessions, but did not generalize reliably across sessions. The study is still a preprint, but its implication is important: there may be no universal face of bluffing.

A more realistic AI would therefore learn the player, not the expression.

It could model how a particular opponent blinks, moves chips, shifts posture, or changes vocal patterns—and correlate those signals with hands eventually revealed.

That is less like emotion reading than discovering another layer of imperfect information.

A future system combining strategic reasoning with reliable visual and acoustic modeling could therefore confront an even richer version of poker:

not only the cards a player hides, but the information the player fails to hide.

Yet the strongest poker AIs did not need to read humans at all. Their breakthrough came from learning how not to be exploited.

Teaching Machines to Regret

The researchers developing poker AI were not primarily trying to construct better psychological models of professional players.

A more powerful objective was available: develop strategies that opponents would have difficulty exploiting at all.

One of the central ideas that emerged from this research tradition was counterfactual regret minimization, or CFR.

The word regret sounds psychological, but here it has a mathematical meaning.

Suppose an algorithm repeatedly encounters a class of decision situations and chooses action A. It can retrospectively estimate what would have happened if it had chosen B instead whenever those situations occurred.

If B would systematically have produced better results, the algorithm accumulates counterfactual regret for not choosing it.

Future strategy is adjusted accordingly.

Repeated over immense numbers of simulated games, this process gradually redistributes the probabilities assigned to different actions. In appropriate two-player zero-sum games, average regret can be driven downward and the strategy can approach a Nash equilibrium.

At such an equilibrium, neither player can substantially improve simply by unilaterally changing strategy.

This is not the same as memorizing a collection of poker tips.

The program is learning something much closer to the strategic structure of the game itself.

Polaris was an early public indication of how powerful that idea could become.

It was only the beginning.

2015: A Form of Poker Is Essentially Solved

In 2015, Michael Bowling, Neil Burch, Michael Johanson, and Oskari Tammelin published a remarkable result in Science.

Their program, Cepheus, had essentially weakly solved heads-up limit Texas Hold’em. The breakthrough relied on CFR+, a highly efficient development of counterfactual regret minimization.

The word solved requires care.

Cepheus did not know which cards would appear next. Chance still existed. Even a game-theoretically optimal player can lose an individual hand, session, or long sequence of hands.

Nor had every version of poker been solved.

The result concerned two-player limit Texas Hold’em, where the range of allowable bet sizes is constrained.

What had been achieved was subtler: a strategy sufficiently close to game-theoretic optimality that no opponent could exploit it enough to obtain a practically meaningful long-term advantage under the criterion used in the research.

This was a major milestone.

Computers had already become dominant in several perfect-information games. Cepheus showed that an important competitive game containing private information could also be pushed astonishingly close to its equilibrium solution.

Outside game theory, computer science, and poker, however, relatively few people noticed.

One year later, another board game would make artificial intelligence impossible to ignore.

2016: AlphaGo Makes AI a Global Spectacle

In March 2016, Google DeepMind’s AlphaGo faced Lee Sedol, one of the strongest Go players of his generation, in Seoul.

AlphaGo won four games to one.

More than 200 million people watched the match worldwide, according to DeepMind.

Go had long been considered a formidable challenge for artificial intelligence. Although its rules are simple, its space of possible positions is enormous, making traditional exhaustive search impractical. Elite Go had also become strongly associated with intuition, pattern recognition, positional judgment, and the accumulated strategic knowledge of centuries.

AlphaGo combined deep neural networks with tree search, reinforcement learning, and self-play. Its victory therefore felt different from earlier machine achievements.

The system was not merely evaluating more positions per second.

It appeared to be finding strategies.

Game Two provided the most famous example. AlphaGo’s Move 37 initially appeared highly unconventional to many expert observers. It later proved strategically powerful and became a symbol of a machine discovering play outside established human convention.

AlphaGo turned advances in artificial intelligence into a global cultural event.

Its effect was particularly striking in China.

Go originated in China more than 2,500 years ago and has deep historical and intellectual associations there. Contemporary analyses described hundreds of millions of predominantly Chinese viewers following the AlphaGo–Lee Sedol confrontation.

Brookings later cited Kai-Fu Lee’s description of the reaction as an “AI fever” that swept China’s technology community. The episode has also frequently been described as a Chinese technological Sputnik moment.

This should not be turned into a simplistic causal story. China already had substantial capabilities and ambitions in computing, machine learning, internet technology, and automation. AlphaGo did not create China’s AI strategy.

But it made the implications of rapid AI progress extraordinarily visible.

In July 2017, China’s State Council published the New Generation Artificial Intelligence Development Plan, explicitly identifying AI as a strategic opportunity and setting national development objectives extending to 2030.

AlphaGo had helped move artificial intelligence from a specialist research subject into public consciousness and geopolitical strategy.

Yet Go still differed from poker in one fundamental respect.

Everyone could see the board.

Poker kept its cards hidden.

And in 2017, researchers made another major attack on that harder information problem.

2017: DeepStack and Libratus Take on No-Limit Poker

Limit Texas Hold’em constrains the size of bets.

No-limit poker removes that restriction.

A player may potentially wager an entire stack, generating a vastly larger strategic space because bet sizing itself becomes part of the problem.

Two major systems demonstrated in 2017 that this frontier could also be crossed.

The first was DeepStack, developed by researchers associated with the University of Alberta, Charles University, and the Czech Technical University.

DeepStack combined game-theoretic reasoning, game decomposition, and a learned value function. Rather than attempting to solve the entire poker game tree at once, the system continually re-solved the strategically relevant subproblem as new cards and bets appeared, using a neural network to estimate the value of future situations that were not searched all the way to the end.

Against professional poker players, DeepStack achieved statistically significant winning performance over 44,000 hands of heads-up no-limit Texas Hold’em.

At approximately the same time, Carnegie Mellon researchers Noam Brown and Tuomas Sandholm developed Libratus.

Libratus used a large precomputed strategic blueprint, refined particular subgames during play, and included mechanisms for identifying and repairing potential weaknesses in its strategy.

It then faced four specialist professional players over 120,000 hands of heads-up no-limit Texas Hold’em.

The humans lost.

By this stage, the old question—whether computers could compete with elite poker professionals—was becoming much less interesting.

They could.

But an important limitation remained.

All these headline results concerned two-player poker.

Real poker tables often contain several opponents.

And that changes the mathematics significantly.

2019: Pluribus Enters the Table

Two-player zero-sum games possess an unusually elegant structure.

One player’s gain is the other’s loss, and Nash-equilibrium reasoning provides strong theoretical guarantees.

Add more players and this simplicity disappears.

A decision can help one opponent while harming another. A strategically useful response to one player can alter the incentives of everyone else. The equilibrium guarantees that make the two-player case so attractive do not transfer cleanly.

Multiplayer poker was therefore a substantial challenge.

Then came Pluribus.

Again developed by Noam Brown and Tuomas Sandholm through Carnegie Mellon University and Facebook AI, Pluribus tackled six-player no-limit Texas Hold’em.

The system relied heavily on self-play. Rather than simply ingesting a canonical encyclopedia of human poker strategy, it improved through repeated competition against versions of itself.

When tested against elite human professionals, Pluribus achieved superhuman performance under the experimental conditions.

Poker AI had reached the multiplayer table.

Something else had also become unmistakable.

The machines were not necessarily reproducing the strategic habits that humans had developed.

They could discover different ones.

Unconventional bet sizes and unfamiliar patterns could emerge because the optimization process had no obligation to respect tradition, elegance, fashion, intimidation, or human intuition.

It had only the objective.

Humans could subsequently study what the machine discovered.

This has become one of the recurring features of modern AI. A system does not always outperform humans by learning to imitate us more faithfully.

Sometimes it succeeds precisely because it does not inherit all our habits.

Poker Was Never Really About Poker

It would be easy to tell this history as a familiar succession of human defeats:

Kasparov loses at chess. Poker professionals lose to algorithms. Lee Sedol loses at Go. Another domain of human expertise falls.

That framing is dramatic, but scientifically shallow.

The important question is not whether a machine can humiliate another champion.

The important question is what kind of intelligence the problem requires.

Poker is particularly valuable because it formalizes a situation that appears everywhere outside games: several decision-makers interact while possessing different information and different incentives.

Consider an auction. Bidders have private valuations.

Consider negotiation. Neither side normally discloses the minimum agreement it would accept.

Consider cybersecurity. Defender and attacker continually infer intentions from incomplete evidence while concealing their own capabilities.

Consider markets. Participants act with asymmetric information and expectations about how other participants will react.

In each case, the agent must act without knowing the complete state of the world.

Poker is useful because it turns that problem into a precisely specified experimental environment.

The rules are known. Outcomes are measurable. Information asymmetry is explicit. Strategic deception is permitted.

The cards are merely laboratory equipment.

The real subject is uncertainty.

There Is No Straight Line From Deep Blue to ChatGPT

From the perspective of today’s generative-AI boom, it is tempting to draw a simple history:

Deep Blue → poker AI → AlphaGo → ChatGPT.

As cultural history, this sequence captures something real. Each system changed public expectations about what machines could do.

As technical history, however, it is misleading.

These systems emerged from different research traditions and solved different problems.

Deep Blue relied heavily on search, specialized evaluation functions, enormous computing power, and expert-designed chess knowledge.

Cepheus and Libratus emerged primarily from computational game theory, abstraction, optimization, search, self-play, and regret minimization.

AlphaGo combined deep neural networks, tree search, reinforcement learning, and self-play.

Modern large language models emerged principally from neural sequence modeling, transformer architectures, large-scale pretraining, enormous datasets, and increasingly large computational resources.

The histories intersect. Techniques migrate. Reinforcement learning, search, self-play, representation learning, and optimization increasingly appear in hybrid forms.

But Cepheus was not an immature chatbot.

AlphaGo was not an early language model.

Pluribus was solving a fundamentally different problem from ChatGPT.

There is no single ladder called “intelligence” up which every AI system simply climbs.

Different tasks expose different components of what we casually call intelligent behavior.

And poker exposes a particularly strange one:

behavior that resembles deception without requiring us to establish that the machine understands deception at all.

That takes us from game theory into philosophy.

The Chinese Room at the Poker Table

In 1980, philosopher John Searle introduced one of the most famous thought experiments in the philosophy of artificial intelligence: the Chinese Room.

Imagine a person who does not understand Chinese locked inside a room.

Messages written in Chinese enter the room. The person possesses detailed instructions, written in a language they understand, explaining how to manipulate the unfamiliar Chinese symbols and which symbols should be returned in response.

By following those rules with sufficient precision, the person may produce answers so convincing that Chinese speakers outside the room conclude that whoever is inside understands Chinese.

But, Searle argues, the person does not understand Chinese at all.

They manipulate syntax without possessing the corresponding semantics.

Searle used the thought experiment to challenge what he called Strong AI: the claim that executing the appropriate computer program would, by itself, constitute genuine understanding. The argument has generated extensive criticism and counterarguments and remains highly contested in philosophy of mind and cognitive science. It should not be treated as a proof that machines cannot understand.

Its relevance to poker is therefore not that it settles anything.

It is that it exposes a distinction that poker makes unusually vivid:

behavior and understanding are not the same question.

Consider the bluffing machine.

It may bet at exactly the right frequency.

It may infer an opponent’s hidden range.

It may exploit systematic weaknesses in that opponent’s strategy.

It may respond to a raise exactly as an elite player would.

Its behavior may be indistinguishable, at the table, from sophisticated strategic deception.

But does the machine know that it is bluffing?

Does it understand deception?

Does it possess anything comparable to the human concepts of risk, uncertainty, intimidation, confidence, or fear?

Game theory does not need to answer those questions.

To determine whether the strategy works, it needs only the behavior.

This produces a poker version of the Chinese Room problem.

From outside the room—or from across the poker table—we observe actions and infer intelligence, intention, perhaps even understanding.

Yet successful behavior alone may not tell us what exists behind it.

The distinction has become even more provocative in the age of generative AI, where machines routinely produce linguistic behavior that invites humans to attribute intention, belief, personality, and understanding.

Poker encountered a narrower version of that puzzle earlier.

A machine did not need to understand bluffing like a human in order to bluff better than many humans.

Perhaps intelligence and understanding overlap less neatly than our intuitions suggest.

Or perhaps our tests are simply better at measuring competent behavior than subjective understanding.

Poker cannot decide between those possibilities.

It can only make the question difficult to ignore.

From Building Poker AI to Keeping It Out of Poker

Then the story acquired an irony.

For decades, researchers asked:

Can we build a machine good enough to sit at a poker table with expert humans?

Modern online poker increasingly faces the opposite problem:

How can we know that the player at the table is actually human?

Poker solvers now analyze strategically complex situations at levels that would have been unimaginable to ordinary players when Polaris appeared.

Used away from the table, such tools can be powerful training systems.

Used secretly during live play, they create a fundamentally different problem.

A bot can partially or completely automate play.

Real-time assistance, or RTA, can provide a human player with computer-generated recommendations while the hand is still in progress.

Major poker platforms now explicitly prohibit such assistance.

PokerStars’ current terms prohibit tools including AI and bots that eliminate or reduce genuine human decision-making, as well as systems offering real-time advice and certain advanced calculations during play.

GGPoker’s current Security & Ecology Policy, dated March 13, 2026, explicitly prohibits RTA, bots, solvers, charts, and other external assistance used during gameplay. Its stated potential penalties for RTA violations include permanent bans and confiscation of funds.

This should not be converted into the sweeping claim that “using AI in poker is illegal.”

That depends on jurisdiction, gambling law, contractual obligations, and the specific conduct involved.

The more precise problem is competitive integrity.

If the game is presented as a contest among humans making their own decisions, covert machine assistance changes the nature of the contest.

The historical circle is almost complete.

Researchers first developed AI to see whether a machine could defeat poker players.

Poker operators must now develop technological and statistical methods to detect players who may secretly be relying on machines.

The opponent has become invisible.

The Game After the Game

The history of poker AI is therefore much more than another chapter in the long competition between humans and machines.

Deep Blue showed that enormous computational search could defeat the world’s strongest human in a perfect-information game.

Cepheus demonstrated how close machines could come to game-theoretic optimality when crucial information was hidden.

AlphaGo combined learning, search, and self-play to defeat one of the greatest Go players in history, turning artificial intelligence into a global spectacle and contributing to an extraordinary surge of interest in AI in China.

DeepStack and Libratus pushed machine strategy into heads-up no-limit poker.

Pluribus moved it to the multiplayer table.

Together, these systems illustrate different routes toward behavior we recognize as intelligent.

Poker’s contribution is distinctive because an agent must operate while reality remains partially concealed. It must reason over possibilities, infer hidden states, remain difficult to exploit, anticipate other strategic agents, and sometimes take actions whose effectiveness depends precisely on concealing what it knows.

That is much closer to many real decisions than a fully visible chessboard.

Not because life is a casino.

Because life rarely reveals all its cards.

The original question has therefore largely been answered:

Can a machine bluff?

Yes.

More precisely, a machine can discover that protecting private information requires actions operationally indistinguishable from bluffing.

It does not need a poker face because it has no face.

It does not need courage because it does not fear losing the pot.

It does not need to feel deceptive for deception-like behavior to emerge from optimization.

And the Chinese Room reminds us to be careful about the next inference.

From successful behavior, we can conclude that the behavior is successful.

Whether the machine understands what it is doing is a different and much harder question.

For decades, that philosophical puzzle could be expressed as:

If a machine behaves intelligently, is there genuine understanding inside?

Online poker has produced an almost comic inversion.

The immediate problem is no longer philosophical at all.

Is there even a human inside?

For decades, researchers struggled to build a machine intelligent enough to sit at a poker table.

Today, the question is no longer whether AI can beat humans at poker, but whether humans can still be certain they are playing against humans.

---

References and Further Reading

Related Content on AI-Talks.org

• Game Theory
• Does the Intelligent Machine Really Knows What is doing?
• The Ultimatum Game: Where Rational Choice Meets Fairness, Emotion, and Artificial Intelligence
• Why Every Intelligent Decision Begins Before Certainty
• Classical Multi-Armed Bandit: Learning Under Uncertainty
• The Gittins Index: One of the Most Elegant Algorithms Ever Created
• When the World Refuses to Stay Still: From Restless Bandits to Adaptive Intelligence

Barrett, Lisa Feldman, Ralph Adolphs, Stacy Marsella, Aleix M. Martinez, and Seth D. Pollak. “Emotional Expressions Reconsidered: Challenges to Inferring Emotion from Human Facial Movements.” Psychological Science in the Public Interest 20, no. 1 (2019): 1–68.

Bowling, Michael, Neil Burch, Michael Johanson, and Oskari Tammelin. “Heads-Up Limit Hold’em Poker Is Solved.” Science 347, no. 6218 (2015): 145–149.

Brown, Noam, and Tuomas Sandholm. “Superhuman AI for Heads-Up No-Limit Poker: Libratus Beats Top Professionals.” Science 359, no. 6374 (2018): 418–424.

Brown, Noam, and Tuomas Sandholm. “Superhuman AI for Multiplayer Poker.” Science 365, no. 6456 (2019): 885–890.

Choi, Minwoo. MediaPipe and DINOv2 Representations for Hand-Strength Prediction in Broadcast Poker. Preprint, 2026.

Feinland, Joshua, Jacob Barkovitch, David Lee, Andrew Kaforey, and Umur A. Ciftci. “Poker Bluff Detection Dataset Based on Facial Analysis.” In Image Analysis and Processing – ICIAP 2022, 400–410. Lecture Notes in Computer Science 13233. Cham: Springer, 2022.

George, S. M., and P. M. Ilyas. “A Review on Speech Emotion Recognition: A Survey, Recent Advances, Challenges, and the Influence of Noise.” Neurocomputing 568 (2024): 127015.

Moravčík, Matej, Martin Schmid, Neil Burch, Viliam Lisý, Dustin Morrill, Nolan Bard, Trevor Davis, Kevin Waugh, Michael Johanson, and Michael Bowling. “DeepStack: Expert-Level Artificial Intelligence in Heads-Up No-Limit Poker.” Science 356, no. 6337 (2017): 508–513.

Searle, John R. “Minds, Brains, and Programs.” Behavioral and Brain Sciences 3, no. 3 (1980): 417–424.

Silver, David, Aja Huang, Chris J. Maddison, et al. “Mastering the Game of Go with Deep Neural Networks and Tree Search.” Nature 529 (2016): 484–489.

Slepian, Michael L., Steven G. Young, Abraham M. Rutchick, and Nalini Ambady. “Quality of Professional Players’ Poker Hands Is Perceived Accurately from Arm Motions.” Psychological Science 24, no. 11 (2013): 2335–2338.

State Council of the People’s Republic of China. Plan for Development of the New Generation of Artificial Intelligence. State Council Gazette, no. 22, 2017.

Sung, Michael, and Alex Pentland. “Stress and Lie Detection through Non-Invasive Physiological Sensing.” International Journal of Biomedical Soft Computing and Human Sciences 14, no. 2 (2009): 111–118.

University of Alberta Computer Poker Research Group. The First Man-Machine Poker Championship. 2007.

University of Alberta Computer Poker Research Group. The Second Man-Machine Poker Competition. 2008.

Zhang, S., Y. Yang, C. Chen, X. Zhang, Q. Leng, and X. Zhao. “Deep Learning-Based Multimodal Emotion Recognition from Audio, Visual, and Text Modalities: A Systematic Review of Recent Advancements and Future Prospects.” Expert Systems with Applications 237 (2024): 121692.

---

Maurício Veloso Brant Pinheiro, PhD
Professor of Physics, Federal University of Minas Gerais (UFMG)
Founder, Author and Editor, AI-Talks.org
About Us

Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.

---

Copyright 2026 AI-Talks.org