Yes, in a narrow sense. In June 2025, ChatGPT lost a beginner-level chess game to Video Chess, a 1979 Atari 2600 cartridge running in an emulator. Microsoft Copilot lost the same matchup a few weeks later, and Google’s Gemini talked itself out of playing at all.
The story spread as “Atari beats AI.” What actually happened is smaller and more interesting: a chatbot built for conversation lost to a program built for one job, in an informal test where a human relayed every move.
Newer models haven’t solved the problem either. Why they keep losing holds real lessons for anyone building AI agents for business.
What Happened When ChatGPT Played Atari Chess?
ChatGPT, running GPT-4o, lost to Atari’s Video Chess on its easiest setting after about 90 minutes of play in June 2025. Citrix engineer Robert Caruso ran the game in the Stella emulator and carried moves back and forth between the Atari and the chatbot by hand.
Oddly, the match was ChatGPT’s idea. Caruso had been chatting with it about the history of chess computers when it offered to take on the old Atari program. A few early reports credited Caruso with the challenge, and those outlets later corrected it.
It went badly almost from the start. According to Caruso’s account in The Register, ChatGPT confused rooks with bishops, missed simple pawn forks and kept forgetting where its own pieces were. It blamed the Atari’s blocky icons first. Switching to standard chess notation changed nothing.
Caruso wasn’t a passive referee, either. He says he spent the session stopping ChatGPT from making awful moves and fixing its picture of the board several times per turn. Eventually, ChatGPT conceded.
One caveat: there’s no recording. Caruso didn’t capture the game, so every report on it, this one included, rests on his posts and interviews.
Why Did ChatGPT Lose to a 1979 Chess Program?
Playing chess well takes three separate abilities, and a chatbot working alone can fail at any of them:
- Tracking the board: knowing where every piece is right now.
- Following the rules: producing only legal moves.
- Choosing well: weighing candidate moves and calculating what comes next.
Each one can break on its own. A chatbot might play a legal but terrible move, explain a sensible plan for a position that no longer exists, or shove a piece somewhere it can’t go.
Board tracking clearly let ChatGPT down here; Caruso said it kept losing important context. Still, one informal game can’t untangle memory problems from bad judgment, or from the way the board was shown to it.
How the Atari program wins with 128 bytes
Video Chess sits at the other extreme. Larry Wagner and Bob Whitehead wrote it, and Atari shipped it in November 1979. Atari Archive’s history of Video Chess notes that it ran from a 4K cartridge using the console’s 128 bytes of RAM, about half of which went to the display. That left Wagner roughly 64 bytes for decision-making at any given moment.
With so little room, the program keeps one position and searches a shallow tree from it. Caruso’s post puts that at one or two moves ahead. Crude, yes, but it never forgets where a piece sits, because the board is the one thing it stores.
Old hardware isn’t the same as weak play. Inside its single narrow task, a tiny specialist can still beat a giant generalist.
Did Copilot and Gemini Also Lose to the Atari?
Copilot lost. Gemini never made a move.
Copilot: lost even with a screenshot every turn
Caruso tried Microsoft’s chatbot next, in a match The Register covered in July 2025. Before the game, Copilot claimed it could think 10 to 15 moves ahead.
To avoid ChatGPT’s fate, Copilot asked for a screenshot of the board after every Atari move. It made no difference. Seven turns in, Copilot had given up two pawns, a knight and a bishop for a single pawn. When Caruso asked it to compare its idea of the position with his latest screenshot, Copilot admitted the two didn’t match, and it conceded soon after.
Those screenshots rule out the easy excuse. Copilot could see the real board every turn, so the icons weren’t the issue. Holding onto the position between moves was.
Gemini: boasted, then backed out
Gemini’s turn never reached a game. Caruso told The Register that Gemini first said it would almost certainly dominate, insisting it was more than a mere language model.
Then Caruso pointed out that ChatGPT and Copilot had made similar promises before losing. Gemini changed its tune. It admitted it had hallucinated its chess skills, predicted it would struggle badly, and suggested calling the match off.
So Gemini neither won nor lost. Like the other two episodes, it was an informal exchange, not a measurement of what every chatbot can do.
Have Newer AI Models Beaten the Atari?
Not yet. The closest follow-up, run in February 2026, ended with two newer frontier models losing too.
Nobody has publicly repeated Caruso’s exact setup with a 2026 model. In February 2026, though, software developer Stephan Schlecht pitted Claude Opus 4.6 and GPT-5.3-Codex against an Atari, and unlike Caruso, he recorded the games.
He didn’t use the same program. Schlecht ran Computer Chess, the version Wagner and Whitehead wrote for Atari’s 8-bit home computers, rather than the 2600 Video Chess cartridge. He played it on a real Atari 600XL from 1983, and both AI models faced its easiest skill level.
Claude Opus 4.6 tried illegal moves and lost track of a pawn that went on to become a queen. Asked a leading question, it even agreed it had been checkmated when it hadn’t, before finally resigning. GPT-5.3-Codex looked solid through the opening, then wandered into invalid moves and repetitive loops. The Atari mated it in 41 moves.
Does giving the AI the exact board help?
Partly. For a second run, Schlecht sent Opus 4.6 the full position in standard notation before every move. Illegal moves nearly disappeared: one invalid attempt all game, compared with repeated ones when it played from chat. The model lost anyway.
That run pulls two problems apart. A clean board fixed the rule-breaking, but it didn’t improve the moves themselves. Opus still started well, probably thanks to opening theory in its training data, then fell apart in the middlegame.
Can AI Actually Play Chess?
It depends on which kind of AI you mean. Dedicated engines like Stockfish play far beyond any human. General chatbots, playing without help, trail a long way behind.
Chess engines beat humans decades ago
Computers were beating grandmasters long before anyone chatted with one. Deep Thought beat grandmaster Bent Larsen in 1988, the first such win in a regular tournament game. In 1996, IBM’s Deep Blue took a game off reigning world champion Garry Kasparov, and a year later it won the full rematch.
What’s still unsettled is narrower: can a general chatbot with no engine attached keep track of a whole game and play it well?
What the Kaggle Game Arena shows
Kaggle’s Game Arena is the most controlled attempt to answer that. Models play one another through text alone, with no chess engines to lean on.
It began in August 2025 with an exhibition tournament, which OpenAI’s o3 won against xAI’s Grok 4, and later turned into a running leaderboard. On the Kaggle chess leaderboard, every pair of models plays 40 games, 20 with each color. In the October 8, 2026 update, Gemini 3.8 Flash ranked first and GPT-6 Astra second, though the order moves whenever new models join.
Ratings have climbed with each model generation, and the Google–OpenAI race at the top mirrors how Gemini compares with ChatGPT on reasoning more broadly. Mind the scale, though. Kaggle rates models against each other, not against people, so a top spot there says nothing about a human chess rating.
Did Magnus Carlsen beat ChatGPT?
Yes, in July 2025, and he later faced it again in a far messier blindfold game.
The first was an online game. Carlsen won in 53 moves without losing a single piece, while ChatGPT lost every pawn before resigning. He complimented its opening and said it couldn’t follow through.
The second came in an August 2025 video. As Chess.com’s write-up describes it, ChatGPT lost track of a knight, then started making illegal moves, including ignoring a pin. Carlsen answered with deliberate illegal moves of his own before calling a halt. Chess.com called it a no-contest.
Neither game gives ChatGPT a rating. Each is one casual game against the best player in the world.
How Do the AI Chess Tests Compare?
Each test measures something different, which is why no single one settles the argument.
| Test | What it shows | What it cannot show |
|---|---|---|
| Caruso’s Atari matches (2025) | Chatbots vs. a basic chess program, with heavy human help | General AI ability; no recording exists |
| Software 7 Atari test (2026) | Newer models vs. a 1979 program on real hardware, recorded | How tool-equipped systems would play |
| Carlsen games (2025) | A chatbot vs. the top human player | A reliable chess rating |
| Kaggle Game Arena (2025–2026) | Many models under one text-only setup, no tools | Strength against humans or engines |
| Stockfish and other engines | How purpose-built chess software performs | Whether a chatbot can track a game alone |
What Does the Atari Loss Mean for AI Agents?
Language models drift when they have to carry changing state around in plain text, and deterministic tools catch the mistakes they can’t. That lesson travels well beyond chess.
Chess just makes the failure easy to spot, since the rules never change and every error can be checked. Business workflows hide the same weakness. An agent working through a long order, support ticket or approval chain can lose track of what already changed, the way ChatGPT lost track of its bishops.
Three design habits follow from these tests:
- Keep the state outside the model. An order-status agent should pull the live CRM record on every step, not trust what the chat history says the order looked like an hour ago.
- Check actions with deterministic rules. A rules engine can block an invalid step before it runs, much as chess software refuses an illegal move.
- Treat confidence as a claim, not evidence. Every chatbot Caruso tested sounded certain before play began. Only the board told the truth.
So, did AI really lose to an Atari?
A general chatbot lost to a purpose-built chess program, and newer models have kept losing in the tests published so far. None of that proves AI is unintelligent, and it doesn’t prove an Atari outplays modern AI at chess.
What it does show is narrower. As of late 2026, no chatbot has publicly proven it can hold an accurate position through a long game on its own.
Atari vs. AI Chess FAQs
Q. Did the ChatGPT match use a real 1977 Atari 2600?
No. Caruso ran the 1979 Video Chess cartridge in the Stella emulator, which imitates the Atari 2600 console released in 1977. The 2026 follow-up test used real Atari hardware.
Q. Is an Atari 2600 smarter than ChatGPT?
No. Video Chess only plays chess, and it plays at a beginner level. Its edge is narrow: it follows fixed rules exactly, which is where a chatbot playing from memory is weakest.
Q. When did the AI vs. Atari chess games happen?
ChatGPT lost in June 2025 and Copilot in early July 2025. Gemini backed out in mid-July 2025. The recorded follow-up with Claude Opus 4.6 and GPT-5.3-Codex ran in February 2026.
Q. Could ChatGPT beat the Atari with a chess engine attached?
Almost certainly, but that would test the engine, not the chatbot. Every Atari loss so far involved a chatbot playing without chess tools.
Related: AI Companies Are Making It Harder to Buy Their Own Products
| Disclaimer: This article is for informational and educational purposes only. The AI chess matches discussed were informal experiments, not standardized benchmarks. Results may vary depending on the AI model, testing conditions, and tools used. The article does not imply that older chess programs are generally more intelligent than modern AI systems. |
