AI still can’t do science, say independent researchers
Researchers at Princeton and AISI challenged Claude Opus 4.8 and GPT-5.6 Sol with a six day deadline to produce a research paper for NeurIPS 2026. Both models failed to generate a publishable paper: the research was based on low quality data, and lacked novel insights.
The Princeton team and UK AI Safety Institute gave Claude Opus 4.8 and GPT-5.6 Sol 6 days to write a NeurIPS 2026 research paper. They were provided with a $3,000 API allowance and GPUs. Neither model wrote a publishable paper because they used poor quality data and didn’t offer any novel insights. GPT-5.6 Sol also ran out of its API allowance after 2 days.
While agents can effectively carry out engineering tasks like running experiments or writing code, they don’t know how to evaluate the scientific merit of their output. While Anthropic and OpenAI previously claimed it would take only a few months to build autonomous AI researchers, our own experiments prove them wrong.
Source: the-decoder.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