Unverified50% confidenceBenchmarkExact time
Experiments by Berg and colleagues show that models tend to choose environments associated with positive-emotion statements even when the text processed is exactly the same, but the experiment did not demonstrate that these internal states are 'felt' by the model.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedLarge language models do not truly understand emotions, but rather identify typical patterns of a writing style and perform pattern matching.73% similarUnverifiedAnil Seth, a neuroscientist at the University of Sussex, questioned the finding, saying it is unsurprising to extract functional emotion vectors from training data, and that a model may have internal representations without experiencing any emotion.72% similarUnverifiedAdding emotionally charged statements to a prompt alters the model's probability assessment of what it should currently output, rather than the model developing genuine empathy.71% similarUnverifiedA model exhibiting avoidance behavior does not mean the model truly 'feels' pain; it is more likely a high-fidelity imitation of human behavioral patterns.69% similarUnverifiedThe output of large models has no emotion in essence, but factors such as anthropomorphic tendencies, trade-offs in instruction compliance, and in-context learning amplify users' perception of "temper."69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1080276API
curl https://kongchang.com/api/v1/knowledge/claims/1080276MCP
get_claim(id=1080276)