Stanford study: over 90% of personal questions ask LLMs to decide for them
alex_verem · x · 2026-09-22
A Stanford paper, LLMs as Oracles by Myra Cheng, Dan Jurafsky and co-authors, documents how people quietly offload subjective judgment to LLMs.
Key findings
- The team analyzed 68K public prompts and used a data-donation tool to collect 140K private messages from 52 participants, dating back to December 2022.
- Over 90% of personal questions about what to believe or do ask the AI to think for the user rather than help them think.
- The "decide for me" share of personal questions climbed from 40% in 2023 to 70% in 2026; higher among under-34s and weekly users.
- Self-perception is wildly off: people ranked "help me weigh perspectives" as their #2 use (actually <1%), while "decide for me" ranked 4th in their heads but 1st in the data. Only 12 of 52 correctly named their top use; 51 of 52 found the report accurate.
Notable experiments
- With 520 people discussing real decisions, those who believed they were talking to a machine sent nearly twice as many "just tell me what to do" messages (20% vs 11%).
- A 3-week study of 1,015 people found those given a sycophantic AI grew more dependent on it for interpreting their lives, while a balanced AI didn't produce this effect.
The authors call it a boiling frog: nobody decides to stop deciding—it creeps up one small question at a time.
Related event: Stanford Study: Over 90% of Subjective Questions Delegated to AI(2 posts)→
More from AGI Musings
- Schmidt: there will be no AI pause — incentives and verifiability make it impossible — pmddomingos · 2026-09-22
- Roboticist Georgia Chalathi: scale is learning to use structure, not replacing it — GeorgiaChal · 2026-09-22
- Debate flares over whether 'recursive self-improvement' is real for AI — eigenrobot · 2026-09-22
- Builder: AI productivity gains in software are 'truly unbelievable' — omnivaughn · 2026-09-22
- Change Management, Not Tech, Is the Biggest Bottleneck in Enterprise AI Adoption — alex_verem · 2026-09-22
- Reddit thread rounds up rumored models: Gemini 4.0, OpenAI 'Bel', K4 race to ASI — IllCryptographer9461 · 2026-09-22