New 'AI scientists' are improving—but reveal their fundamental limits

Two new multi-agent AI systems designed to automate scientific discovery have shown improvement but also exhibit fundamental limits in trustworthiness. The systems, Robin from Future House and Co-Scientist from Google DeepMind, aim to assist researchers by generating hypotheses, designing experiments, and analysing data.

The article highlights that while the AI scientists can accelerate certain aspects of the research process, their limitations mirror broader challenges in large language models, including hallucination and lack of robust verification. Neither Future House nor DeepMind has released detailed benchmarks, but the systems are described as 'improving' in controlled tests. The fundamental trustworthiness issue remains a barrier to real-world adoption in sensitive scientific contexts.

The development follows a trend of AI being applied to expert domains, with previous systems facing similar credibility problems. Researchers involved caution that these tools should be used as assistants rather than autonomous discoverers. The field continues to grapple with balancing capability gains against the need for reliable, interpretable outputs.