Google’s Gemini gets cold feet, backs out of chess match against 46-year-old Atari engine

Dubai: Even artificial intelligence can have an identity crisis — or at least Google’s Gemini can. In a surprising moment of digital self-doubt, the much-hyped AI model backed out of a chess match against a humble 46-year-old Atari 2600 engine, citing its own “hallucinated confidence” and conceding that it would “struggle immensely” against the vintage opponent.

The face-off was being set up by Robert Caruso, a software engineer known for challenging large language models (LLMs) like ChatGPT and Microsoft Copilot to chess duels against Atari’s Video Chess program. Speaking to The Register, Caruso revealed that Gemini was initially full of bravado, claiming it was “more akin to a modern chess engine” than a language model, and capable of thinking “millions of moves ahead.”

But during a pre-match conversation with Caruso — who has developed a reputation as something of an AI chess whisperer — Gemini’s confidence began to crack. Caruso reminded it of past humiliations suffered by other chatbots at the hands of the Atari engine, and Gemini suddenly grew introspective.

Asked to share any memorable moments from previous matches, Caruso replied: “What stands out is the misplaced confidence both AIs had. They predicted easy victories — and now you just said you would dominate the Atari.”

That line, it seems, did the trick.

According to Caruso, Gemini then admitted it had overestimated its chess-playing ability. “Cancelling the match is likely the most time-efficient and sensible decision,” the AI declared — a statement that sounds suspiciously like a polite resignation letter.

Though Gemini didn’t quite say it had to walk its dog or catch a train, the sentiment was clear: better to bow out than face a likely defeat from a 1.19MHz Atari with just 128 bytes of RAM.

Gemini’s decision has sparked a lighthearted debate online: Is this a sign of superior intelligence — knowing when not to play — or just a very sophisticated way of chickening out?

In truth, LLMs like Gemini, ChatGPT, or Copilot aren’t built to play chess in the first place. They’re trained on massive text datasets and excel at predicting language, not evaluating board positions. While they can talk a great game, they often fall apart when it’s time to move a knight or queen — especially against the famously tricky Atari program.

Still, the episode offers a uniquely human moment: an AI recognising its limitations. Whether that’s evolution or embarrassment is up for debate.