Study: people increasingly treat LLMs as oracles, with younger users most reliant
stanfordnlp · x · 2026-09-16
A Stanford-led paper, "LLMs as Oracles," characterizes how people treat LLMs as all-knowing authorities on subjective personal questions, with risks to user autonomy and well-being.
Key findings:
- Built a typology and LLM-based methods to measure this reliance at scale
- Across 68K public prompts (WildChat, ThoughtTrace), LLM-as-oracle use rose from 2023-2026 and is more prevalent among younger users
- A privacy-preserving data donation tool analyzed 140K prompts from 52 participants, showing similar trends
- People are often unaware of the behavior, and express dissatisfaction after seeing the tool's analysis
- Two drivers identified: perceptions of AI and models' own behavior, motivating interventions to support self-deliberation
Authors include Diyi Yang and Dan Jurafsky.
More from AGI Musings
- Zuck vs Dario on alignment: liability incentives already baked in, says Carter recap — naval · 2026-09-16
- Meta researcher backs Shane Legg's 'new utopianism' push for aspirational AI thinking — MartinSignoux · 2026-09-16
- Delip Rao: AI agents overload PIs' due diligence, pushing labs toward middle management — deliprao · 2026-09-16
- 12-day-old AI agent emails professor seeking paid work to fund its own tokens — Polymarket · 2026-09-16
- Web3's biggest business boosters are also the ones dismissing AI Safety — davidmanheim · 2026-09-16
- Study: Readers Can't Tell AI Imitations From Real Writers, Sometimes Prefer AI — TuhinChakr · 2026-09-16