Researchers Tried Letting AI Do Science. It Failed

What happened

Researchers tested cutting‑edge AI systems on scientific tasks; the AI could perform routine research steps but did not generate novel results that would be accepted at a leading AI conference.

Confirmed

Global impact / market context

Because investors hope AI can speed up discovery, the study shows current models still need human insight to create publishable breakthroughs, limiting immediate commercial value and suggesting that funding may stay focused on tools that assist rather than replace researchers.

Analyst inference

While AI stocks have surged on expectations of autonomous innovation, this result tempers optimism, indicating that near‑term revenue from AI‑driven R&D services may grow slower than projected, and that companies may prioritize incremental productivity gains over fully automated research.

Analyst inference

What to watch

1. Watch for follow‑up studies that test newer AI models on the same research tasks, to see if future systems can produce conference‑level original work. Analyst inference
2. Monitor citations of this study in the AI research community, as broader discussion may influence funding allocations toward improving AI creativity and in the. Analyst inference
3. Track whether top AI conferences adjust submission guidelines to allow AI‑generated papers, which could affect acceptance rates and signal industry confidence in AI research capabilities. Analyst inference

Evidence

• decrypt.co