LLM-generated forum replies are more neutral and less varied in sentiment than human replies¶

Claim · All claims
Evidence · 1 study · causal r1 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q2 i? Gemma and LLaMA generated 80% or more neutral replies versus 62% for human replies, and all three LLMs produced fewer positive replies than humans. → Liu 2025

Evidence¶

Liu 2025¶

Liu, Z., Xing, W., Jiao, X., & Li, C. (2025). Exploring Fairness and Explainability in LLM-Generated Support for Online Learning Discussion Forums. Journal of Learning Analytics, 12(3), 8–33. https://doi.org/10.18608/jla.2025.8885

q2 · i? · causal · r1

Sentiment analysis of 8,322 post-reply pairs using BERT and SOWB classifiers on replies from three fine-tuned LLMs compared with original student replies; the study reports "80% or more of replies from the Gemma and LLaMA models were neutral" versus 62% human, and positive-reply proportions of 12%, 10%, and 6% versus 27%.

"In contrast, 80% or more of replies from the Gemma and LLaMA models were neutral—much higher than the 62% observed in human replies. For positive reply proportions, all three LLMs produced fewer positive replies (12%, 10%, and 6%) than humans did (27%)."

Discussion¶

Related Claims¶

• Counterfactual fine-tuning reduces sentiment bias in LLM-generated forum replies — related
• Sentiment scores of LLM-generated replies differ significantly from human replies under both classifiers — related
• GPT-generated tutor feedback is predominantly positive or neutral in sentiment — related