AI's Social Skills Depend on Emotions, Not Logic

A new study from Johns Hopkins University has uncovered a surprising mechanism behind how artificial intelligence systems understand human perspectives. Researchers found that when language models get better at attributing beliefs to others—a skill known as Theory of Mind—they do so by amplifying emotional processing and dampening analytical reasoning. This s common assumptions that AI social abilities rely on logical, step-by-step thinking, suggesting instead that emotional context plays a fundamental role in how machines interpret social scenarios.

The key finding shows that improving a language model's Theory of Mind performance involves systematic changes in its internal cognitive processes. Using a technique called Contrastive Activation Addition (CAA) steering on the Gemma-3-4B model, the researchers boosted accuracy on belief attribution tasks from 32.5% to 46.7%—a 14.2% improvement that shifted 217 examples from incorrect to correct predictions. More importantly, they discovered that this improvement was mediated by specific changes: emotional processes like emotion perception and emotion valuing increased significantly, while analytical processes like questioning and convergent thinking decreased. This pattern indicates that successful perspective-taking in AI depends more on emotional understanding than on deliberate analytical interrogation.

To reach these conclusions, the researchers developed a novel decomposition that combines activation steering with linear probes trained on 45 cognitive actions. They generated 31,500 synthetic training samples across four categories—Metacognitive, Analytical, Creative, and Emotional—using first-person narratives in everyday contexts. These probes were then applied to analyze activation patterns in the Gemma-3-4B model during 1,000 forward belief scenarios from the BigToM benchmark. By comparing baseline versus steered conditions at three timepoints—at question, after true answer, and after wrong answer—they could identify which cognitive processes changed when Theory of Mind performance improved. The probes achieved an average AUC-ROC of 0.78 and F1 score of 0.68, with mid-layers (5-24) showing the best performance for capturing cognitive abstractions.

, Detailed in Figures 2 through 5, reveal consistent patterns across all measurements. Emotional actions showed the strongest increases, with emotion perception rising by a mean difference of +1.73 and emotion valuing by +0.85. Creative processes like hypothesis generation also increased by +1.63. In contrast, analytical actions decreased, with questioning dropping by -1.24 and convergent thinking by -1.13. Category-level analysis in Figure 3 confirms that emotional and creative processes consistently increased across timepoints, while analytical processes showed decreases. These suggest that when AI systems successfully engage in perspective-taking, they activate representations responsible for processing emotional contexts rather than those involved in logical analysis.

This research has significant for how we design and evaluate AI systems intended for social interaction. If emotional processing mediates Theory of Mind abilities in language models, it could inform approaches to AI alignment, human collaboration, and social reasoning applications. The study notes that neuroscience research shows similar patterns in humans, where affective and cognitive Theory of Mind share neural mechanisms in brain regions like the Temporoparietal Junction. However, the researchers caution against direct comparisons to human cognition, emphasizing that their ology doesn't validate such generalizations. Instead, they suggest that language models may learn shared representations linking perspective-taking with emotional context processing through exposure to linguistic data, potentially mirroring compressed structures of human social cognition.

Despite these insights, the study acknowledges several limitations. are based on a single model (Gemma-3-4B) and one dataset (BigToM forward belief scenarios), so they may not generalize to other models or different types of Theory of Mind tasks. The researchers explicitly state that future work should validate these cognitive decomposition with multiple, bigger models and additional data sources. Additionally, the synthetic training data for probes, while carefully designed across 20 everyday domains, might not capture all nuances of real-world cognitive processes. The open question remains whether these patterns constitute genuine emulation of cognitive architecture or merely emergent convergence on functionally equivalent representations through training on language data.