No AI system reigns supreme across the board; the contests that come closest to a verdict are narrow games, and the results show wins in one game at a time. Between 2017 and 2022, Carnegie Mellon University and Meta researchers published peer-reviewed results for three games with hidden information: two-player poker (Libratus), six-player poker (Pluribus), and the negotiation game Diplomacy (CICERO). Each win was real, and each came with limits that the researchers themselves stated.
Key takeaways
- Libratus played 120,000 hands of heads-up no-limit Texas hold'em against four professionals and won, as reported in Science in 2018.
- Pluribus beat professionals in six-player poker, reported in Science in 2019, using far less computing power than Libratus.
- CICERO scored more than double the human average across 40 games of online Diplomacy and ranked in the top 10 percent of players who played more than one game.
- Each system was built for one game, so a win says little about general ability.
- Games with hidden information and more than two players are harder to solve, and theory offers fewer guarantees there.
Why contests with hidden information matter
Chess and Go give both players the full board. Poker does not: cards are hidden, bluffing is part of play, and the best move depends on what an opponent might hold. Researchers treat such games as stand-ins for real situations where information is incomplete, such as negotiations and security problems. The poker papers describe the game as a long-standing benchmark for exactly that reason. A system that wins here must reason about uncertainty, not just search more moves ahead.
The three results below form a progression: from one opponent, to five opponents, to six opponents who can talk to each other in plain language.
Libratus: two players, 120,000 hands
Noam Brown and Tuomas Sandholm of Carnegie Mellon University described Libratus in Science (volume 359, January 2018). The system played a 120,000-hand match of heads-up no-limit Texas hold'em against four elite professionals and won. The paper describes three parts: a precomputed overall strategy, a method that refines the strategy for the specific situation reached during play, and a module that patches weaknesses that opponents find in the base strategy over time.
According to a Carnegie Mellon release on Pluribus, Libratus used about 15 million core hours to develop its strategies and 1,400 processor cores during live play. That scale put the method out of reach for most laboratories.
Pluribus: six players, much less computing
Brown and Sandholm's next paper, "Superhuman AI for multiplayer poker," appeared in Science on July 11, 2019. Poker with more than two players is harder because, as the university's release puts it, playing a Nash equilibrium (a standard theoretical safe strategy) can be a losing approach in games with more than two players. Pluribus therefore relies on strategies that come without a theoretical guarantee of success.
Two tests were reported. Darren Elias and Chris Ferguson each played 5,000 hands against five copies of Pluribus. In a second test, 13 professionals, each with more than USD 1 million in career poker winnings, played five at a time against the program for 10,000 hands in total. Pluribus won both. Its blueprint strategy took eight days and 12,400 core hours to compute, and live play ran on 28 cores, a small fraction of the Libratus figures.
CICERO: seven players and negotiation in language
Diplomacy is a seven-player strategy game in which players make and break alliances through private conversation. In a paper published in Science on November 22, 2022, the Meta Fundamental AI Research Diplomacy Team described CICERO, which combines a language model with planning and reinforcement learning. It infers other players' beliefs and intentions from conversation and writes messages to carry out its plans.
In 40 games of an anonymous online league, CICERO's score was more than double the human average, and it ranked in the top 10 percent of participants who played more than one game. The AAAS release adds that it passed as human against 82 unique players. The league played speed games, so the result does not cover every format of the game.
What a win in these games does and does not show
The table is a summary of the three published results and the stated limits, not a ranking of the systems.
| System | Game | Reported result | Limit worth keeping in mind |
|---|---|---|---|
| Libratus | Heads-up no-limit hold'em | Won a 120,000-hand match against four professionals | Two players only; very large computing budget |
| Pluribus | Six-player no-limit hold'em | Beat professionals in two test formats | No guarantee from theory in multiplayer play |
| CICERO | Diplomacy (seven players) | Top 10 percent of multi-game players across 40 games | One league and speed-game format |
None of these systems can play a different game without being rebuilt. That is the central point behind the question of who "reigns supreme": contests measure a system against specific rules, specific opponents, and a specific date.
Checklist for reading an AI contest result
- How many games or hands were played, and was the sample large enough to separate skill from luck?
- Who were the opponents, and were they experts?
- Was the computing budget disclosed?
- Does the result come from a peer-reviewed paper or only a press release?
- Does the claim stay inside the game that was tested?
FAQ
Which AI is the strongest overall?
No contest answers that. Results exist for specific games and tests, and each system was built for its own task.
Did Pluribus play against only one human?
It faced two formats: one human against five copies of the program, and five humans at a time against the program, with 13 professionals taking part in the second format.
Did CICERO tell opponents it was an AI?
The Science paper's abstract does not address disclosure. The AAAS release says it played anonymously and passed as human against 82 unique players.
Are these results still the latest?
They are the published results cited here, dated 2017 to 2022. Newer systems and contests exist, and none of these figures should be read as a current ranking.
Related reading
- Who is called the father of AI
- The role of mathematics in AI
- How AI programs process large data
- Ethical questions around AI
More in AI Basics.
Last verified: October 2026
Sources
- Brown N, Sandholm T, Superhuman AI for heads-up no-limit poker: Libratus beats top professionals, Science 359(6374), 2018
- Brown N, Sandholm T, Superhuman AI for multiplayer poker, Science 365(6456), 2019
- Meta Fundamental AI Research Diplomacy Team, Human-level play in the game of Diplomacy by combining language models with strategic reasoning, Science 378(6624), 2022
- Carnegie Mellon University news release on Pluribus, July 11, 2019, via EurekAlert
- AAAS news release on CICERO, November 22, 2022, via EurekAlert