Humans Choose More Diverse Goals Than LLMs in Self-Directed Learning Study of 5 Models
ValerioCapraro · x · 2026-11-06
Researchers compared 175 humans with five LLMs (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Qwen3 32B, Centaur) in a self-directed learning game where players picked which potions to make and discovered recipes. Humans explored broadly and gradually diversified goals, while most models fixated on a single goal or recipe, with distinct patterns per model and little variability across instances. Chain-of-thought and persona steering barely helped. The UC Berkeley / Amazon AGI Lab authors warn that as AI increasingly decides which problems and hypotheses are worth pursuing, it may steer humanity toward a narrow set of questions. Paper linked in replies.
Related event: LLMs Pick Far Narrower Goals Than Humans, Study Finds(2 posts)→
More from AGI Musings
- US added 165,000 jobs in October as Gary Marcus mocks Dario's AI job-loss predictions — GaryMarcus · 2026-11-06
- Hassabis: AI now boosts engineers 3-5x, but closed-loop self-improvement is the risky next phase — haider1 · 2026-11-06
- 5 months of university report work in 2021 is now one deep research query away — gaganghotra_ · 2026-11-06
- Schmidhuber: Mind Uploads Won't Stay Human—They'll Evolve Beyond Recognition — SchmidhuberAI · 2026-11-06
- Dan Shipper's closing note: what humans should keep doing as AI takes over work — kieranklaassen · 2026-11-06
- AI Book Club picks "The New Dark Ages": AI accelerates the post-literate society — LuizaJarovsky · 2026-11-06