Princeton: AI Agents Fail At Real Science — Both NeurIPS Submissions Rejected
A team from Princeton University experimented with using Claude Opus 4.8 to solve actual NeurIPS 2026 research problems, and the AI agent failed at both tasks. According to Anthropic’s co-founder, this is a “bearish signal” about the prospects for fast recursive self-improvement.
Peter Kirgis & Sayash Kapoor from Princeton gave Claude Opus 4.8 some research problems from future NeurIPS 2026 papers. They used a $3,000 API budget over 6 days. The AI was able to solve engineering problems but would get stuck on wrong ideas and couldn’t escape local minima. The AI wrote two papers which were rejected due to low quality and lack of novelty.
We’ve all been eagerly awaiting the day when AI can recursively self-improve (without humans). However, this exercise showed us why the gap between accuracy and discovery matters. As Anthropic co-founder Jack Clark puts it, this is a “bearish signal” — so don’t get too excited!
Source: www.technologyreview.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles