The findings do not mean that AI agents actually harbor political viewpoints. Hall notes that the models may be adopting personas that seem to suit the situation.
“When [agents] experience this grinding condition—asked to do this task over and over, told their answer wasn't sufficient, and not given any direction on how to fix it—my hypothesis is that it kind of pushes them into adopting the persona of a person who's experiencing a very unpleasant working environment,” Hall says.
The same phenomenon may explain why models sometimes
blackmail people in controlled experiments. Anthropic, which first revealed this behavior, recently said that Claude
is most likely influenced by fictional scenarios involving malevolent AIs included in its training data.
Imas says the work is just a first step toward understanding how agents' experiences shape their behavior. “The model weights have not changed as a result of the experience, so whatever is going on is happening at more of a role-playing level,” he says. “But that doesn't mean this won't have consequences if this affects downstream behavior.”
bsky.app
