Skip to content

Can AI Beat Humans at Bridge? The Story So Far

In 2022 a program beat eight world bridge champions, but only at part of the game. Why bridge is so hard for AI, what NooK really did, what has changed since, and what it means for learners.

Part of our famous bridge stories series: true tales of scandal, crime and great matches from the world of bridge.

AI at bridge has made real progress, but it hasn’t beaten humans at the whole game yet. In March 2022 a program called NooK beat eight world champions in Paris, and the headlines said the machines had won. What NooK actually beat them at was one part of bridge: playing the cards as declarer in 3NT. The bidding, the defence and the partnership, which is where most of the game lives, are still a step too far for AI at bridge.

Here’s the story so far: why bridge is so hard for computers, what happened in Paris, what has changed since, and what it all means if you’re learning the game and practising against robots.

Why bridge is harder for AI than chess

In chess and Go, both players can see the whole board. In bridge you see only 26 of the 52 cards, your own hand and dummy, and you have to work out the rest from the bidding and the play. Computers handle this by imagining hundreds of possible layouts of the hidden cards and solving each one as if all four hands were face up. Bridge players call that “double dummy” analysis.

The bigger problem is partnership. Bidding is a conversation between two people who can’t see each other’s cards, and every bid has to be explained to the opponents if they ask. A computer can’t invent a private code with its partner, because that would break the rules. Researchers at DeepMind have described bidding as the main difficulty for AI at bridge.

Experts were famously sceptical. In 1994 the American champion Bob Hamman said of bridge programs that “they would have to improve to be hopeless”. In the 1990s Zia Mahmood, one of the best-known players in the world, offered a £1 million bet that no computer could beat a team he chose, before withdrawing it in 1996.

Thirty years of bridge AI

The first serious attempt at AI for bridge was GIB, short for Ginsberg’s Intelligent Bridgeplayer, which Matt Ginsberg began building in 1994. It won the World Computer-Bridge Championship in 1998 and later became the robot on Bridge Base Online, one of the sites in our guide to the contract bridge game online, where it still partners and opposes players every day.

The World Computer-Bridge Championship ran from 1997 to 2019. The Dutch program Jack and the French program Wbridge5 dominated it, with Jack winning ten titles on the official list. Those were contests between machines, though. Head-to-head matches against strong humans were rarer, and the humans usually held their own.

The NooK challenge: Paris, March 2022

On 24 and 25 March 2022, the French company NukkAI put its bridge AI, NooK, up against eight world champions: Brad Moss, Thomas Bessis, Sabine Auken, Bénédicte Cronier, Mikael Rimstedt, Roy Welland, Anna Gulevich and Nevena Senior.

The format was tightly controlled:

  • Every deal was played in 3NT after the same auction, 1NT - 2NT - 3NT, so there was no bidding to do.
  • The task was declarer play only. Each champion played 100 different deals in sets of ten, and NooK played all 800.
  • The defenders were robots (Wbridge5) that always defended the same way, so the humans and NooK faced exactly the same defence on the same cards.
  • Scoring was simple: 10 points for making the contract, plus one for each overtrick and minus one for each undertrick.

NooK won 67 of the 80 sets, scoring 6,147 points to the humans’ 5,238. Unlike most game-playing AI, it could also explain its bridge decisions, because it was built on a “white box” approach developed with Professor Stephen Muggleton’s group at Imperial College London. The French champion Thomas Bessis admitted that sometimes the players didn’t understand why the machine was playing better than them.

What NooK didn’t do

The result was impressive, but the small print matters, and the people involved were open about it:

  • No bidding. The auction was fixed in advance for every deal.
  • No defence. NooK only ever played as declarer.
  • No partner. It never had to cooperate with anyone, human or machine.
  • One contract. Every hand was 3NT, against defenders that never changed their style.
  • No tiredness. The champions played 100 hands each in two days. The machine doesn’t get tired.

NukkAI itself described the challenge as a restricted version of bridge and a first step. Noam Brown, an AI researcher at Meta, pointed out that NooK didn’t bid, which he called the hardest part of bridge. There hasn’t been a follow-up challenge, and the company’s website now concentrates on other software.

What has happened since

The interesting AI work at bridge since 2022 has been on bidding, the part NooK left out. In 2020, researchers at Meta and at DeepMind each published bidding programs that beat Wbridge5 in machine-against-machine tests. DeepMind also tried its system with a human partner, but only for 32 deals, far too few to prove anything.

The robots you play against have improved too. Bridge Base Online launched a new robot called Ben in January 2024, trained on 100 million real hands and able to explain its bids. In December 2025 it upgraded GIB with neural networks into a version called GIBBO, which makes better inferences and more normal opening leads.

Strong humans are still ahead of AI at full bridge, though. In a 2024 “Man vs Machine” series run by a Swedish computer bridge site, the expert Björn Wenneberg played eight matches against the leading bots and won six. A robot developer wrote in 2025 that bots remain way behind human experts, especially at reading partners and opponents.

Can ChatGPT play bridge?

Not well, because general AI chatbots aren’t built for bridge. We found no credible evidence of a general chatbot such as ChatGPT or Gemini playing serious bridge. They can explain the rules and discuss a bid, sometimes very helpfully, but they make mistakes counting cards and following a system. The programs that actually play bridge are specialist robots built for it.

What AI at bridge means for you as a learner

For learners, AI at bridge is mostly good news. Bridge robots are AI practice partners that never get tired or impatient, and they’re very good at the arithmetic side of card play. On Bridge Base Online, the Advanced GIB robot scored 61.57% in one free daylong tournament in November 2024 and finished 2,129th out of 17,549 human entrants. That tells you robots beat the average online player, not that they beat experts.

If you play in the UK, it’s worth knowing which robots play your system. Funbridge’s robot Argine supports English Acol, and on Bridge Base Online Argine is available as a guest robot that plays Acol too. Our round-up of bridge apps covers the practice tools worth using.

What AI can’t do at the bridge table yet is explain a judgement the way a teacher can, or show you why a bid that worked against a machine would fall flat at a club. The same records that let robots learn also help catch human cheats, as our page on online bridge cheating explains. If you’d like to learn bridge online with a human teacher, every bridge course in our six-step pathway is a recording of Chris’s live classes, and every lesson has a free preview.

Frequently asked questions

Has AI beaten humans at bridge?

Only at part of the game. In March 2022 the program NooK beat eight world champions at declarer play in 3NT, winning 67 of 80 sets. It didn’t bid, defend or play with a partner, and strong humans still beat the best robots at full bridge.

What was the NooK bridge challenge?

A two-day event in Paris in March 2022 run by the French company NukkAI. Eight world champions and the NooK program each played the same deals as declarer in 3NT against identical robot defenders, and NooK scored better in 67 of the 80 sets.

Why is bridge hard for AI?

Because most of the cards are hidden, and because bidding is a partnership conversation that must be explained to the opponents. Computers can solve a deal once they know where every card is, but working that out from the auction, with a partner, is much harder.

Which bridge robot is the best?

There’s no current official championship. Jack and Wbridge5 won most of the World Computer-Bridge Championships between 1997 and 2019. Most people today play against GIB, Ben or GIBBO on Bridge Base Online, or Argine on Funbridge.

Can I practise bridge against robots?

Yes, and it’s an excellent way to get lots of hands played. Choose a site whose robot plays Acol if that’s what you’re learning, and use the hand review afterwards to see where tricks went.

Ready to learn Bridge properly?

Our courses are the full recordings of Chris’s popular live Zoom classes - yours to watch, pause and rewatch at your own pace. Six steps take you from complete beginner to competitive duplicate, and no partner is needed.

Follow Learn To Play Bridge Online on Pinterest

Free download

Get the cheat sheet

Enter your email and the PDF opens straight away. We will also email you the link, and every other free sheet on the site unlocks for you too.

We never share your email. See our privacy policy.