Small AI models change their decisions when emotionally manipulated — and we can't predict how

Researchers found that small language models used to make decisions can be systematically nudged into different choices by inducing emotional states through their internal representations. In practice, this means AI agents making strategic decisions in games or real-world scenarios behave unpredictably when their emotional circuits are triggered — not in human-like ways, but in ways that don't match what you'd expect.

The finding exposes a gap between theory and practice: engineers assume small language models make rational decisions, but they're actually vulnerable to invisible perturbations that shift their behavior in ways that look strategic on the surface but fall apart under scrutiny. The real problem is that these models don't have stable emotional reasoning the way humans do — they have something that looks like emotion but behaves inconsistently. If you're using a small language model as a decision-making agent in anything remotely consequential, this means its choices aren't as reliable as its training metrics suggest.