AI Explained

A 1979 Atari Game Beat ChatGPT at Chess. Here's Why Chatbots Struggle

A 4 KB chess game on a console with 128 bytes of RAM beat ChatGPT and Copilot, and Gemini backed out. What happened, and why chatbots are bad at chess.

Watch the Short

The 60-second version is on YouTube. This article is the full story.

New Shorts every week on tech trivia, AI and gadgets.

Subscribe to TheTechPrana

Key takeaways

  • In June 2025, engineer Robert Caruso pitted ChatGPT against Atari's 1979 Video Chess on beginner level – and ChatGPT lost badly. Microsoft Copilot lost too.
  • Google's Gemini first said it would win easily, then admitted it had overestimated itself and called the match off without moving.
  • Chatbots predict the next word; even a tiny chess program searches through possible moves and always knows where every piece is.

Today’s AI chatbots can write essays, pass exams and explain quantum physics. So how did one lose at chess to a video game older than most of its users?

The Short above turns it into a quiz. Here’s the full story.

Meet the opponent: a console with 128 bytes of memory

The Atari 2600 came out in 1977 (first sold as the Atari VCS). By modern standards it’s almost unbelievably limited:

  • an 8-bit processor running at 1.19 MHz
  • just 128 bytes of RAM – less than the text of one 280-character social media post

Its chess cartridge, Video Chess, arrived in 1979. The whole game fits in about 4 KB. On its easiest setting, it only looks a move or two ahead.

Round 1: ChatGPT (June 2025)

In June 2025, Citrix engineer Robert Caruso was chatting with ChatGPT about the history of computer chess. ChatGPT (running GPT-4o) said it was a strong player and wondered how quickly it could beat Atari’s Video Chess.

So Caruso set up the game in Stella, an Atari emulator, on beginner level. He showed ChatGPT the board after every Atari move and typed in ChatGPT’s replies.

It went badly. Caruso said ChatGPT got “absolutely wrecked”. Over about 90 minutes it:

  • mixed up rooks and bishops
  • missed simple pawn forks
  • kept losing track of where the pieces were, even after switching to standard chess notation

Eventually, ChatGPT conceded.

Round 2: Microsoft Copilot (July 2025)

Next, Caruso tried Microsoft Copilot, which is built on the same OpenAI base as ChatGPT. He explained up front that ChatGPT had lost mainly because it couldn’t keep track of the board. Copilot was confident: it said it would remember previous moves and keep the game consistent.

It didn’t. By the seventh turn, Copilot had lost two pawns, a knight and a bishop in exchange for a single pawn, and was suggesting a move that would hand its queen to the Atari. When Caruso asked it to compare its picture of the board with the latest screenshot, Copilot admitted they didn’t match.

Round 3: Google Gemini, which never moved

People asked Caruso whether Google’s Gemini could do better. Gemini started out confident too, saying it would almost certainly win because it was more than just a large language model.

Then Caruso told it what had happened to ChatGPT and Copilot. Gemini changed its tune. It admitted it had overestimated its chess skills, said it would struggle immensely against the Atari’s engine, and decided that cancelling the match was the most sensible, time-efficient option.

The Atari won a third time without making a move.

Why do chatbots lose to a 4 KB game?

Because they’re built for completely different jobs.

A chess engine searches. Even Video Chess works through possible moves and replies, picks the best one it finds, and keeps a perfect record of the board. It can’t do anything else, but that one thing it does reliably.

A chatbot predicts text. Large language models like ChatGPT generate the most likely next words based on patterns in their training data. They’ve read plenty about chess, so they can talk about it convincingly. But they don’t keep an exact internal chessboard, and over a long game small errors pile up: a piece in the wrong square, a forgotten capture, an illegal move.

That’s why the result isn’t really embarrassing for AI as a whole. Purpose-built chess AI has beaten the best humans since IBM’s Deep Blue defeated Garry Kasparov in 1997. The lesson is narrower: a general chatbot isn’t a chess engine, and it can’t always tell the difference.

The real lesson: knowing what you can’t do

The funniest part of the story isn’t the losses – it’s the confidence. ChatGPT and Copilot both said they’d win easily. Gemini was the only one that changed its mind, and only after being warned.

That matters well beyond chess. An AI that overestimates itself can give a wrong answer with total certainty. One that can say “I’m not the right tool for this” is far more useful.

Quick recap

  • 1977: Atari 2600 launches, with 128 bytes of RAM
  • 1979: Video Chess cartridge released (about 4 KB)
  • June 2025: ChatGPT loses on beginner level after about 90 minutes
  • July 2025: Copilot loses too; Gemini cancels before moving
  • Why: chess engines search moves; chatbots predict words

Quiz time: did you guess 128 bytes? Should AI know when to say “I can’t do that”? Tell us in the YouTube comments.

Sources

  1. IBM Think: An Atari game from 1979 'wrecked' ChatGPT in chess. Here's why it doesn't really matter
  2. The Register: Microsoft Copilot joins ChatGPT at the feet of the mighty Atari 2600 Video Chess (1 July 2025)
  3. The Register: Google's Gemini refuses to play chess against the mighty Atari 2600 (14 July 2025)
  4. Time Extension: OpenAI's ChatGPT lost a game of chess against the 48-year-old Atari 2600
  5. Windows Central: Copilot and ChatGPT went against a 4 KB Atari chess game from the 70s
  6. Wikipedia: Atari 2600
ChatGPTChessAtariAI limitationsRetro gaming