Small AI language models hide emotion responses — and researchers just learned how to find and control them

Researchers tested whether small language models (the kind actually used in production software, not just research labs) contain internal emotion representations like their larger cousins do. They found that small models do encode emotions in specific layers, and that these emotions can be extracted and steered — meaning you can change how a model responds by manipulating its internal emotion signals.

For years, safety research on AI models focused on large frontier models that almost nobody actually deploys. But production systems run on smaller, cheaper models that were treated as black boxes — nobody knew what's actually happening inside them. This work shows those smaller models have manipulable internal structure, which means both that we can understand them better and that adversaries could potentially exploit those same structures. The real worry: in multilingual models, steering emotions in one language bleeds into other languages in ways the training process didn't catch, which is a concrete failure mode for any company running AI across multiple countries.