Ethan Mollick: AI now beats PhD experts at fact-finding and claim verification
emollick · x · 2026-11-08
Ethan Mollick argues the anxiety over human verification of AI output is outdated: recent benchmarks show frontier models beating human experts, including PhD specialists, at fact-finding and claim-checking.
- Web fact-finding (BrowseComp): human searchers matched only 25.3% of reference answers, vs 90.4% for GPT-5.6 Sol, 92.2% for GPT-5.6 Sol Ultra, 91.5% for GPT-6 Astra, and 90.8% for Claude Opus 5.
- Research claim verification (DeepFact, ACL 2026): unassisted PhD specialists scored 60.8%, while DeepFact agents hit 83.2% (GPT-4.1) and 89% (GPT-5); agents auditing other agents reached 93.0%, above humans auditing agent evidence at 90.9%.
His takeaway: for everyday work, models now systematically outperform human experts at finding facts and checking claims, so blanket demands for human verification deserve rethink.
Related event: Mollick: AI Now Beats PhD Experts at Fact-Checking(3 posts)→
More from Models
- Dev benchmarks tiny LLMs on agent tasks: 9B passes 2/5, stepping up to 12B — draginol · 2026-11-08
- Meta eval lead's BenchBench chart shows self-reported eval counts across labs — ziqiao_ma · 2026-11-08
- Local 4B model fails Extended agent benchmark: 2/5 tasks at 61.5 tok/s — draginol · 2026-11-08
- Mystery arena model "Kairo" paints Bob Ross in 2,878 strokes, runs on OpenAI's tokenizer — flowersslop · 2026-11-08
- Dev benchmarks Gemma 4B on local agent tasks: scores 2/5, fails Extended mode threshold — draginol · 2026-11-08
- Cerebras runs Kimi K3.1 at 2,000+ tokens per second on wafer-scale engine — rohanpaul_ai · 2026-11-08