A customer service chatbot tells an upset client, "I understand how frustrating this must be." It sounds genuinely empathetic.
Yet an empirical study co-authored by Dr. Jiaoping Chen, an assistant professor of information systems at the University of Baltimore’s Merrick
School of Business, suggests that what happens inside the AI model looks very different
from how a person understands someone.
Chen helped lead a new study with first author Xiaoyan Li and a team of computational and healthcare researchers, comparing how AI models process
information with how the human brain does, using brain-imaging (fMRI) data.
The paper, titled "Attention Mechanisms for Theory of Mind: Weak Brain-Transformer
Alignment Despite Behavioral Success," evaluated four language models against brain
scans from 58 people who listened to spoken stories about social interactions and
physical movements. It has been accepted for the 2026 IEEE International Conference on Bioinformatics and Biomedicine.
The findings reveal a clear split: while larger AI models answered social reasoning
test questions with up to 83 percent accuracy, their internal calculations showed
very weak alignment with the brain regions humans use to understand one another.
In cognitive psychology, the ability to track what other people think, feel, or want
is called Theory of Mind. Humans rely on a network of brain regions, including the
medial prefrontal cortex and temporoparietal junction, to pick up unspoken cues, detect
awkward social missteps, and anticipate behavior.
To test whether AI works the same way, the researchers fed identical story transcripts into four transformer-based language models, ranging in size from 355 million to 7 billion parameters.
The 7-billion-parameter models solved classic social reasoning puzzles with 75% to 83% accuracy. They also separated social stories from physical ones with up to 98% accuracy.
The models' internal activity showed little correlation with human brain data. The strongest correlation reached 𝑟 =0.047 in the medial prefrontal cortex, which did not remain statistically significant after correcting for multiple comparisons.
Think of it like a student who passes a reading test by spotting patterns in the questions
rather than understanding the story. The AI produces the right answer by predicting
the most likely next word, but it does so without building an internal model of human
thoughts or intentions.
The models often arrive at the right answer, but the way they get there looks quite different from what we see in the human brain. A high benchmark score does not mean a system understands people the way we do.
The study represents measured, incremental progress in cognitive science rather than
an overhaul of the field. For corporate executives, human resource directors, and
technology officers, the findings provide practical guardrails when evaluating automated
tools:
As artificial intelligence takes on more workplace tasks, Chen's co-authored study offers a useful reminder: AI remains an effective text-processing tool, but understanding people remains uniquely human.
