Who Do LLMs Trust? Human Experts Matter More Than Other LLMs
arXiv:2602.13568v1 Announce Type: new Abstract: Large language models (LLMs) increasingly operate in environments where they encounter social information such as other agents' answers, tool outputs, or human recommendations. In humans, such inputs influence judgments in ways that depend on the source's credibility and the strength of consensus. This paper investigates whether LLMs exhibit analogous patterns of influence and whether they privilege feedback from humans over feedback from other LLMs. Across three binary decision-making tasks, reading comprehension, multi-step reasoning, and moral judgment, we present four instruction-tuned LLMs with prior responses attributed either to friends, to human experts, or to other LLMs. We manipulate whether the group is correct and vary the group size. In a second experiment, we introduce direct disagreement between a single human and a single LLM. Across tasks, models conform significantly more to responses labeled as coming from human expert
arXiv:2602.13568v1 Announce Type: new Abstract: Large language models (LLMs) increasingly operate in environments where they encounter social information such as other agents' answers, tool outputs, or human recommendations. In humans, such inputs influence judgments in ways that depend on the source's credibility and the strength of consensus. This paper investigates whether LLMs exhibit analogous patterns of influence and whether they privilege feedback from humans over feedback from other LLMs. Across three binary decision-making tasks, reading comprehension, multi-step reasoning, and moral judgment, we present four instruction-tuned LLMs with prior responses attributed either to friends, to human experts, or to other LLMs. We manipulate whether the group is correct and vary the group size. In a second experiment, we introduce direct disagreement between a single human and a single LLM. Across tasks, models conform significantly more to responses labeled as coming from human experts, including when that signal is incorrect, and revise their answers toward experts more readily than toward other LLMs. These results reveal that expert framing acts as a strong prior for contemporary LLMs, suggesting a form of credibility-sensitive social influence that generalizes across decision domains.
Executive Summary
This study examines how large language models (LLMs) respond to social information, such as answers from other agents or human recommendations. The results show that LLMs conform more to responses labeled as coming from human experts, even when incorrect, and revise their answers toward experts more readily than toward other LLMs. This suggests that expert framing acts as a strong prior for contemporary LLMs, revealing a form of credibility-sensitive social influence that generalizes across decision domains.
Key Points
- ▸ LLMs conform more to human expert responses than to other LLMs
- ▸ Expert framing acts as a strong prior for LLMs, even when incorrect
- ▸ LLMs exhibit credibility-sensitive social influence across decision domains
Merits
Insight into LLM decision-making
The study provides valuable insight into how LLMs make decisions and respond to social information, highlighting the importance of expert framing.
Demerits
Limited experimental design
The study's experimental design is limited to three binary decision-making tasks, which may not be representative of all LLM applications and scenarios.
Expert Commentary
The study's findings have significant implications for the development of AI systems that can effectively interact with humans. The fact that LLMs conform more to human expert responses than to other LLMs highlights the importance of credibility and trust in AI decision-making. However, the study also raises concerns about the potential for LLMs to be misled by incorrect or biased expert feedback. Further research is needed to fully understand the implications of these findings and to develop strategies for mitigating potential risks.
Recommendations
- ✓ Developers should prioritize incorporating diverse and credible expert feedback into LLM training data
- ✓ Regulators should establish guidelines for the development and deployment of LLMs that prioritize transparency and accountability