Google DeepMind’s Gemini 2.5 Pro panics playing Pokémon, causing AI blunders
Google DeepMind shared a new report showing its latest AI, Gemini 2.5 Pro, struggles with early Pokémon games. The model exhibits a “panic” state when its Pokémon near defeat, which hurts its reasoning and gameplay.
The AI doesn’t just mess up strategy — it actually abandons useful tools during these panic moments. Twitch chat watching the AI noticed the glitchy behavior in real time. 
“Over the course of the playthrough, Gemini 2.5 Pro gets into various situations which cause the model to simulate ‘panic,’” the report states.
“This behavior has occurred in enough separate instances that the members of the Twitch chat have actively noticed when it is occurring.”
This glitchy human-like panic is a new wrinkle in AI game testing. Other AI, like Anthropic’s Claude, also show strange moves. Claude once tried to intentionally faint all its Pokémon to warp forward in Mt. Moon — only to fail and restart from the wrong spot.
Gemini 2.5 Pro does shine in puzzle-solving. It created “agentic tools” with some human help to crack boulder puzzles and find faster paths through Victory Road. The report says:
“With only a prompt describing boulder physics and a description of how to verify a valid path, Gemini 2.5 Pro is able to one-shot some of these complex boulder puzzles, which are required to progress through Victory Road.”
Google thinks Gemini might soon build these tools solo. Maybe it’ll add a “don’t panic” fix next.
Meanwhile, Twitch streams like Gemini Plays Pokémon and Claude Plays Pokémon let viewers watch AI struggle with a 25-year-old game in real time, showing the weird mix of progress and goofs.
Gemini still takes hundreds of hours per game that a kid finishes in minutes — but the human-like flaws might offer fresh insights into how AI thinks (or fails to).
