BYLINE: Ala N. Tak, PhD Candidate, Computer Science, Viterbi School of Engineering; Research Assistant, Affective Computing Group
As a PhD candidate in Computer Science at USC’s Viterbi School of Engineering and a researcher in the Affective Computing Group at the Institute for Creative Technologies, I have been investigating one of the most fascinating and concerning aspects of our increasingly AI-integrated world: how large language models perceive and process human emotions.
I am honored that IEEE Transactions on Affective Computing has accepted our paper “Aware Yet Biased: Investigating Emotional Reasoning and Appraisal Bias in Large Language Models” (Ala N. Tak, Jonathan Gratch, and Klaus R. Scherer), which validates the importance of investigating emotional reasoning and bias in AI systems.
The question that drives my research is deceptively simple yet profoundly complex: When advanced language models encounter emotionally charged situations, do they reason about emotions the way humans do? And perhaps more critically, what biases do they bring to these assessments that could impact the millions of people now interacting with AI systems in contexts ranging from therapy to legal proceedings?
The Foundation of Emotional Understanding
My investigation into the emotional reasoning capabilities of large language models began with appraisal theory—a cornerstone framework in affective science that suggests our emotions arise not from situations themselves, but from how we evaluate or “appraise” those situations. When two sports fans witness their team losing, their emotional responses may differ dramatically based on their individual appraisals: one might focus on remaining game time (hope), while another fixates on the score deficit (despair).
This theoretical foundation provided the perfect lens through which to examine LLMs. Working alongside my advisor, Jonathan Gratch, and Klaus Scherer—a leading figure in emotion theory, I designed studies to test whether models like GPT or Gemini could accurately predict human emotional responses and, crucially, whether they exhibited the same systematic biases that characterize human emotional reasoning.
Discovering Remarkable Capabilities and Troubling Limitations
In our first study, I analyzed LLM performance against a substantial multilingual corpus containing 5,636 autobiographical descriptions of emotional experiences collected across English, French, and German speakers. The results were simultaneously impressive and concerning.
Both GPT-4 and Gemini-2.0 demonstrated remarkable accuracy in predicting human emotions and appraisals, showing what can only be described as sophisticated emotional intelligence. The model consistently identified emotions like joy and anger with striking precision, and its appraisal predictions aligned closely with human self-reports across multiple languages. This cross-linguistic consistency was particularly noteworthy—the model performed equally well regardless of whether I prompted it in English while analyzing French or German emotional narratives.
However, beneath this impressive surface lay systematic failures that revealed the model’s limitations. Both models consistently struggled with specific emotions—shame, fear, and irritation—suggesting gaps in their understanding of these more complex emotional states. More troubling was their near-random performance on appraisal dimensions related to control and power, fundamental aspects of how humans evaluate their ability to cope with challenging situations.
Uncovering Systematic Bias
These failures led me to Study 2, where I investigated whether LLMs exhibit systematic appraisal biases—the tendency to evaluate situations in consistently skewed ways. Using a specialized corpus designed to identify individual differences in emotional appraisal, I discovered that both GPT-4 and Gemini-2.0 exhibit remarkably similar personality-like biases.
The pattern that emerged was striking: both models appraise situations as if they possessed the personality profile of someone with high agreeableness and emotional competence, but low power, low life satisfaction, and low self-efficacy. In essence, these models seem to embody the perspective of individuals who are emotionally aware and cooperative, yet feel powerless and dissatisfied with their circumstances.
This finding carries profound implications. When an AI system consistently underestimates human agency or overemphasizes helplessness in emotionally charged situations, it could reinforce negative thought patterns in vulnerable users. In therapeutic contexts, such biased appraisals might invalidate a client’s genuine concerns or provide overly pessimistic assessments of their coping abilities.
The Promise and Peril of Debiasing
Perhaps most intriguingly, my research suggests that these biases may be addressable through careful prompt engineering. By incorporating personality information into prompts, I found evidence that LLMs can be “debiased” to some extent, producing appraisals more aligned with specific personality profiles. However, this debiasing proved inconsistent and model-specific, with GPT-4 and Gemini-2 responding differently to the same personality-informed prompts.
This variability highlights a critical challenge in deploying LLMs in sensitive contexts: even when we identify biases, our ability to control them remains limited and unpredictable. The implications extend far beyond academic curiosity—as these models increasingly influence healthcare diagnostics, legal decisions, and personal counseling, their embedded biases could shape human emotional experiences in ways we are only beginning to understand.
Implications for Our AI-Integrated Future
My research reveals that LLMs are not neutral observers of human emotion but active interpreters that bring their own systematic biases to emotional reasoning. These models exhibit what can only be described as machine personalities—consistent patterns of appraisal that influence their emotional assessments.
The clinical implications are particularly concerning. If individuals regularly interact with AI systems that consistently appraise situations through the lens of low power and life satisfaction, these interactions could potentially reinforce or exacerbate similar dispositions in humans. The very technology designed to support human wellbeing might inadvertently undermine it through persistent exposure to biased emotional reasoning.
Charting the Path Forward
This research represents just the beginning of understanding emotional bias in AI systems. As LLMs become more sophisticated and ubiquitous, we must develop more precise methods for identifying, measuring, and controlling their emotional biases. Future work should explore whether these biases originate in pretraining data or emerge during post-training processes, and whether intervention in the model’s internal representations might offer more reliable debiasing than prompt-based approaches.
The stakes could not be higher. As we stand at the threshold of an era where AI systems will increasingly mediate human emotional experiences, understanding and addressing their biases becomes not just an academic exercise, but a moral imperative. The emotional minds of machines, as I have discovered, are neither neutral nor perfectly aligned with human reasoning—they are complex, biased, and potentially influential in ways that demand our urgent attention and careful stewardship.
Through continued research into these mechanisms, we can work toward AI systems that not only understand human emotions more accurately but do so in ways that genuinely support human flourishing rather than inadvertently undermining it.
//