Zhejiang University paper: LLMs match human average choices but adjust along divergent paths
ValerioCapraro · x · 2026-10-13
Behavioral scientist Valerio Capraro highlights a key flaw in AI debates: conflating output similarity with process similarity. An iScience paper from Zhejiang University had humans, DeepSeek-v3.2, and Qwen3-Max repeatedly choose between standard and slow delivery under reward-penalty incentives. Averages looked alike, but humans adjusted gradually and incompletely while LLMs followed model-specific trajectories. Similar aggregate behavior does not mean models "think" like humans.
Related event: Zhejiang Study: LLMs Decide Like Humans but Adapt Differently(2 posts)→
More from AGI Musings
- Moderna Founder Calls for Urgent AI Agent Regulation, Warns of Biomedical 'Disaster' — Polymarket · 2026-10-13
- Judea Pearl: investment measures returns, citations measure intellectual returns — yudapearl · 2026-10-13
- Humanoid robots at scale 20-30 years away: copper, lithium shortages and boomer retirements reshape the AI jobs debate — JHochderffer · 2026-10-13
- Data scientist debunks 'growing AI' narrative: it's just gradient descent nudging token probabilities — gerardsans · 2026-10-13
- AI detectors still used to accuse students of cheating, sparks KOL backlash — VraserX · 2026-10-13
- Job hunting for CS grads in 2026: competing with AI, laid-off seniors and teen founders — _jaydeepkarale · 2026-10-13