NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Contemporary AI lacks the imagination to diverge or negate in science
TL;DR
Artificial intelligence (AI) has been predicted to significantly enhance scientific discovery, but evidence supporting these claims is lacking. A large-scale evaluation was conducted where authors of 121,640 recent scientific preprints assessed ideas generated by large language models (LLMs).
✦ Why It Matters
Engineers and researchers should recognize the limitations of current AI models in generating innovative scientific ideas.
Key Takeaways
- AI struggles to generate diverse scientific hypotheses.
- Human feedback significantly improves AI's output quality.
- LLMs favor ideas similar to existing research.
- Social scientists are more critical of AI-generated ideas.
How It Works
The study involved a large-scale evaluation where scientists rated AI-generated hypotheses based on their own research. This process highlighted the limitations of LLMs in proposing innovative or divergent ideas, particularly in complex fields like social sciences.