LLM Psychology: Do Large Language Models Have Beliefs and Preferences?
brwilder · x · 2026-08-05
This article delves into the concept of "LLM psychology"—finding the right high-level abstractions to describe how large language models accomplish cognitive tasks like decision-making, inference, and generalization.
The author poses a core question: Do LLMs have beliefs?
- Some argue it is almost obvious that we can ascribe beliefs, preferences, and even emotions to models.
- Others see this as overreaching anthropomorphism or a category error.
The author argues that this question is not just scientifically fascinating but crucial for building AI systems that are legible and steerable by humans. Just as we use "folk psychology" to reason about human behavior, finding the right descriptive abstractions for LLMs will inform better approaches to AI safety and value alignment.
Related event: Researchers Explore if LLMs Possess Genuine Beliefs(3 posts)→
More from AGI Musings
- Using AI to Disguise Writing: Scholars Use Claude to Anonymize Peer Reviews — RobbWiller · 2026-08-05
- Academic Peer Review in the AI Era: Using LLMs to Polish Reports and Hide Identity — RobbWiller · 2026-08-05
- Software Abundance: Nader Dabit Says AI Is Creating Unprecedented Opportunities for Engineers — rseroter · 2026-08-05
- Superintelligence May Learn More from Tiny Experiments, Bypassing Massive Compute — jxmnop · 2026-08-05
- Why Self-Improving Autoresearch Loops Fail: Paper Reveals Sharp Drop in LLM Judge Accuracy — dair_ai · 2026-08-05
- Grok 4.5 Loses Its Edge: Users Suspect xAI Trained on Claude Outputs — teortaxesTex · 2026-08-05