Google DeepMind researchers have discovered that Gemini 2.5 Pro can go into panic mode when playing early versions of the Pokémon games, a condition that impairs the AI’s performance and reasoning abilities.
Google DeepMind researchers have discovered that Gemini 2.5 Pro can go into panic mode when playing early versions of the Pokémon games, a condition that impairs the AI’s performance and reasoning abilities.
This is reported by TechCrunch, citing a report by Google DeepMind. According to some researchers, studying how AI models play video games can be beneficial.
Over the past few months, two developers, not affiliated with Google or Anthropic, have created Twitch streams called “ Gemini Plays Pokémon ” and “ Claude Plays Pokémon,» where anyone can watch in real time as an AI attempts to beat a children’s video game released over 25 years ago. The streams demonstrate the AI’s «thinking» process and depict in natural language how the model assesses a problem and finds a solution.
While the progress of these AI models is impressive, they still aren’t very good at playing Pokémon. Gemini takes hundreds of hours to master a game that a child can complete in much less time. But the researchers aren’t interested in the completion time, but in how the AI actually behaves during the game.
«During the game, Gemini 2.5 Pro encounters various situations that cause the model to simulate ‘panic,’» the report says.
This state of «panic» can lead to a deterioration in the model’s performance, as the AI may suddenly stop using certain tools at its disposal for a period of time in the game. Although the AI does not think or feel emotions, its actions mimic how a human might make poor, hasty decisions under stress.
«This behavior was observed in enough isolated instances that Twitch chat participants actively noticed it,» the researchers note.
Claude’s AI also exhibited some interesting behavior during his playthroughs. In one instance, he noticed a pattern: when all of his Pokémon were running low on health, the player character would «disappear» and return to the Pokémon Center.
When Claude got stuck in the cave of Mount Moon, he mistakenly assumed that if he intentionally made all his Pokémon lose health points, he would be transported through the cave to a center in a nearby town. However, the game doesn’t work that way. When all the Pokémon die, the player returns to the center they last used, not the one closest geographically. As a result, viewers watched as the AI essentially tried to kill itself in the game to no avail.
Despite its shortcomings, AI can outperform human players in several ways. For example, Gemini 2.5 Pro AI is capable of solving puzzles with extreme accuracy. With the help of a human, the AI created agent tools (task-specific hints) to solve the game’s stone puzzles and find efficient routes to their destination.
«With just a hint describing the physics of the boulders and a description of how to check for the correct path, Gemini 2.5 Pro is capable of solving some of these complex boulder puzzles required to complete Victory Road in one go,» the report says.
Since the AI did much of the work on its own, Google suggests that Gemini 2.5 Pro may be able to create these tools without human intervention.
Previously, Every, a company that develops software products and provides training services in the field of artificial intelligence, created a reinterpretation of the classic historical strategy game «Diplomacy», in which AI models ChatGPT, Gemini, Claude, DeepSeek and others play as seven great powers of the 1901 model and compete for dominance in Europe .
