The Gemini 2.5 Pro is capable of simulating stress when one of your Pokémon is about to fall in battle.
The use of video games to train artificial intelligence models is becoming a constant. At the end of March, we learned that Anthropic had put its most advanced reasoning model to use to try to complete a game of the original Pokémon for Game Boy. Almost three months later, not only did it fail, but it also drew attention for getting stuck on more than one level, something that many children easily overcome. Surprisingly, Google has embraced the idea with this test with Gemini , but there is a fundamental difference between the two: the American company’s AI “feels” fear when it is about to lose.
As TechCrunch points out, the Gemini 2.5 Pro model exhibited panic behavior when one of its Pokémon was about to fall in battle. Unfortunately, this negative feeling can impair its reasoning ability , which causes it to ignore useful tools in different parts of the game. So, although it does not feel emotions, the AI’s behavior indicates that it is capable of simulating human reactions to stress. In this way, it makes poor decisions that affect its gameplay.
AI can’t handle Pokémon
Claude, the Anthropic model who has already taken this test (and still does on her Twitch channel), also showed flaws in her logic in extreme situations, even thinking that if she intentionally lost, she could teleport to another area . Gemini, who is also showing off her progress in a live stream on Twitch, needs several…
Source: https://www.terra.com.br/byte/google-colocou-sua-ia-para-jogar-pokemon-no-game-boy-centenas-de-horas-depois-ela-comecou-a-sentir-panico-estresse-e-medo,03172abdf7718275bd9970a5083e268bch2rd0p9.html
