Does a heartfelt, emotional appeal move an AI the way it moves a person, or does the AI just calm it down?

This explores how emotional tone changes what an LLM says and judges, versus how emotion changes what people believe, and whether the two match. Across these notes they mostly don't. LLMs tend to defuse emotion rather than respond to it, while humans are often moved by it. None of the studies here puts the same emotional appeal in front of both an LLM and a person, so the contrast has to be assembled from separate findings.

On the LLM side, emotion works more like a smoothing filter than a lever. GPT-4 shows an 'emotional rebound': negative prompts get neutral-positive answers about 86% of the time, and positive prompts rarely get negative ones. The same question therefore gets different information depending on how upset the asker sounds, except on sensitive topics, where alignment training overrides the tone effect (Does emotional tone in prompts change what information LLMs provide?). Emotional disclosure gets a similar treatment in therapy settings. LLM therapists reach for solution-focused advice when a user shares a feeling, which is a hallmark of low-quality human therapy and is probably driven by RLHF's helpfulness bias (Do LLM therapists respond to emotions like low-quality human therapists?). In both cases the model takes in emotion and puts out something calmer or more practical.

Humans change their minds differently. A meta-analysis using the elaboration likelihood model found that human persuaders work through the peripheral route, with emotional vividness and identity cues. LLMs persuade through the central route, with analytical reasoning and coherence (Do humans and AI persuade through different cognitive routes?). An audit of five models found the same split: LLMs used logical appeals and quantitative framing in virtually every conversation, while humans answering the same prompts leaned on emotion and social proof (Do LLMs persuade users more often than humans do?). When LLMs try to predict which arguments actually changed a human's view, they barely beat chance (Cohen's κ of 0.08 to 0.18). They over-weight topical overlap and credibility, while humans respond more to novelty and assertive language (Do language models judge persuasion the way humans do?). The model's sense of what persuades is more analytical than the people it is judging.

The corpus also suggests that 'emotional' is not one thing. LLMs used about 22% more moral language than humans across care, fairness, authority, and sanctity, yet their sentiment scores were nearly identical to human ones. That points to moral appeals and emotional tone working as separate channels (Do LLMs use moral language more than humans?). A model can sound morally loaded without sounding emotional. Its assertive, confident register, which RLHF appears to install, also does persuasive work regardless of whether the claim is true (Does linguistic conviction explain why LLMs persuade more effectively?). Part of what looks like emotional pull in LLM text may actually be conviction.

Belief change also unfolds over time, and that is where the models are weakest. LLMs match humans at tracking fixed mental states, like a persuader's unchanging goal, but fall behind on shifting ones, like a persuadee's growing resistance (Can language models track how minds change during persuasion?). That fits a broader critique that LLM social simulation stays at the level of plausible behavior without modeling the belief networks that make minds move (Can language models simulate belief change in people?). The upshot is that an LLM can be nudged by the mood of a prompt, but it doesn't model why a person's mood, identity, or resistance changes what they will accept.