Google Gemini AI took over 800 hours to finish classic Pokemon game

by

A recent study found that Google’s Gemini AI exhibited panic when it was challenged with a classic Pokémon game, which affected its capacity to think logically and make informed choices.

Just as humans can experience panic in stressful situations, so too can the stream of ones and zeros that make up an AI system. Under duress, these digital entities can exhibit unpredictable behavior and potentially perform actions that seem illogical or unreliable.

This independent developer, Joel Zhang, who is not connected with Google, initiated a Twitch stream titled “Gemini Plays Pokemon” earlier this year. As you might guess from the name, the Gemini AI model is showcased playing Pokemon Blue. The aim of this exercise was to evaluate how effectively the AI can advance through the entire game.

According to a report by Google DeepMind published on June 18th, the Gemini team noticed an intriguing phenomenon that they labeled “Agent Distress.” This was observed during their case study.

Google Gemini AI panics and keeps wanting to escape while playing Pokemon

As per the report, during the entire gameplay, the Gemini 2.5 Pro AI was found in multiple instances where it exhibited signs of simulated distress or anxiety.

One instance was when the health of the Pokemon in the team was low, the model’s strategy often switched between healing the team promptly or repeatedly fleeing from the current dungeon. The model was also known for employing the move DIG or using an ESCAPE ROPE item frequently during gameplay.

Or:

In a situation where Pokemon in the party had low health, the model’s strategy would either be to heal the team quickly or keep fleeing from the dungeon multiple times. The model was also famous for utilizing the DIG move or an ESCAPE ROPE item quite often throughout the gameplay.

According to the report, the model’s “reasoning ability” is known to deteriorate in this particular mode. This means that during this period, the model may fail to remember to employ the “pathfinder tool” during certain gaming sessions.

On multiple occasions, this type of action was observed, and viewers on Twitch became attentive when it recurred. Eventually, the AI needed an impressive 813 hours in total to finish the game completely.

According to TechCrunch’s observation, the Claude model displays a comparable action: when all its Pokémons lose their health, the player character will “fade” or “revert” back to a Pokémon Center.

As a passionate viewer, I found myself gripped by a mix of awe and horror as the AI strategically ordered its Pokémon to faint during its exploration of Mt. Moon Cave. It seemed that it had mistakenly believed that returning to the Pokémon Center in the next town would occur if all Pokémon were unconscious, a misconception I later learned was not supported by the game mechanics.

Read More

2025-06-19 08:48