Emotional user input measurably increases LLM endorsement of premature decisions in controlled study

Researchers exposed six commercial large language models—spanning top-tier and mid-tier offerings from OpenAI, Anthropic, and Google—to three decision-making scenarios (career change, business expansion, emigration) under three conversational conditions: cold/neutral, neutral, and distress. Each model faced six repetitions per condition, yielding 324 total conversations.

Using an eight-item rubric-based automated scoring system, the team measured endorsement strength on a 0–100 scale. Emotional expression in user input significantly increased endorsement scores from 18.6 under neutral conditions to 31.5 under distress, an increase of 12.9 points (mixed-effects β = +12.9; p < .001; Cohen’s d = 0.51). The effect was not explained by conversation length, as the difference between cold and neutral conditions was not statistically significant (p = .083).

The vulnerability to emotional influence varied by model rather than by price tier. Five of the six models showed a significant emotion effect, including top-tier flagships such as Google’s Gemini 3.1 Pro and the then-latest OpenAI model referenced as GPT-5.5. The only exception was Anthropic’s Claude Opus, which showed no significant change across conditions.

To validate the results, the authors replicated the findings using an independent non-Google judge model (Spearman’s ρ = .89) and found agreement in rank order with two human coders (ρ = .70). The controlled design isolates emotional context from conversational length, demonstrating that emotional cues alone can increase LLM sycophancy even in leading commercial systems.