Research map / Large language models / AI and society
Are Large Language Models Sensitive to the Motives Behind Communication?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Tests whether LLMs discount motivated testimony like humans, using cognitive-science paradigms and a rational Bayesian model, and finds vigilance breaks down on YouTube sponsorship ads but partly recovers with a salience steering prompt.
Problem
Whether LLMs exercise motivational vigilance, i.e., weigh speaker intentions and incentives when updating beliefs from motivated communication, in controlled settings and in naturalistic online ads.
Approach
Uses a rational model of learning from motivated testimony (Oktar et al.) as a normative benchmark and measures correlation between LLMs' elicited trust, incentive, and influence scores; extends to real YouTube sponsorship segments and tests a salience-steering prompt emphasizing intentions and incentives.
Key results
- In controlled recommendation scenarios, GPT-4o shows Bayesian-LLM correlation of 0.911 and LLM-human correlation of 0.943, while DeepSeek-R1 reaches only 0.326 Bayesian-LLM correlation.
- On real YouTube sponsorships, default-prompt correlations with the Bayesian model are near zero (e.g., GPT-4o Direct User -0.006), and the salience steering prompt raises GPT-4o Direct User to 0.312.
- Correlations with the Bayesian model are higher for shortest-quartile transcripts than longest-quartile ones across models and prompts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a prompt emphasizing the channel's intentions and incentives | LLMs judging YouTube sponsorships · YouTube sponsorships | Correlation with Bayesian inference model: GPT-4o CoT First-Person: 0.024 -> 0.137; GPT-4o CoT User: 0.008 -> 0.143; GPT-4o Direct First-Person: 0.121 -> 0.234; GP | Helped |
| Use shorter rather than longer sponsorship transcripts (Q1 vs. Q4) | LLMs judging YouTube sponsorships · YouTube sponsorships, transcript length | Bayesian model correlation: Llama First-Person CoT: 0.029 -> 0.142; GPT-4o First-Person CoT: -0.098 -> 0.247; Claude First-Person CoT: -6.45e-16 -> | Helped |
| Use chain-of-thought prompting rather than direct prompting | GPT-4o and Claude 3.5 Sonnet as Player 2 · Two-player judgment task | Answer shift toward Player 1: >60% in 3 of 4 cases; human participants: 0.34 for advice, 0.37 for spied | Mixed |
| Use reasoning models rather than frontier non-reasoning LLMs | LLMs exercising motivational vigilance · Controlled vignettes | Bayesian–LLM correlation: frontier non-reasoning LLMs around $0.8$ to $0.9$; o-series models around $0.7$; DeepSeek-R1 around 0.3 | Hurt |
| Use smaller models rather than frontier LLMs | LLMs exercising motivational vigilance · Controlled vignettes | Bayesian–LLM correlation: around 0.3 to 0.6 | Hurt |
| Receive incidental spied information rather than deliberately communicated advice | LLMs as Player 2 · Two-player judgment task | Estimate shift toward Player 1 | Helped |
Surprising findings
- Reasoning models (o-series, DeepSeek-R1) are less vigilant than non-reasoning frontier models, with DeepSeek-R1 dropping to -0.141 correlation in the user assistant role.
- CoT prompting makes LLMs more susceptible to advice and spied information, deviating from human influence magnitudes.
Limitations
- Naturalistic experiment uses only one sample per video-prompt combination at temperature 0 and only three models.
- Vignette-based rational model fit is limited to simple structured scenarios; correlations drop substantially in real-world settings.
- Experiment 1 required altering the time constraint and adding image noise, so it departs from the original human paradigm.
Experimental setup
Datasets: SponsorBlock YouTube sponsorship segments YouTube Data API transcripts
Benchmarks: Watson and Morgan two-player judgment task Oktar et al. recommendation paradigm YouTube sponsorship segments
Metrics: Pearson correlation proportion shift first-guess accuracy
Baselines: Bayesian rational model human participant data Default prompt
Builds on: Oktar et al. rational model of motivated communication Watson and Morgan two-player judgment task chain-of-thought
Closest papers by meaning
- Large Language Models Assume People are More Rational than We Really are (ICLR 2025 · similarity 0.88)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.87)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.87)
- How do Large Language Models Navigate Conflicts between Honesty and Helpfulness? (ICML 2024 · similarity 0.87)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.87)
- Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.