As artificial intelligence (AI) becomes increasingly involved in scientific research, especially in the fierce competition between China and the US, a unique gathering of AI “scientists” has revealed that the technology still faces fundamental weaknesses.
Each paper presented at last week’s Agents4Science 2025 conference listed large language models as primary authors and reviewers, while scholars from around the world shared how they used AI bots in scientific work and the challenges encountered.
In one study, OpenAI’s ChatGPT and Anthropic’s Claude simulated two-sided job marketplaces, handling everything from brainstorming to experimental design. The main limitation observed was a struggle to maintain the context and focus of the paper.
The conference heard that the AI agents needed reminding to keep supporting documents updated, hallucinated references and generated redundant code and text until human collaborators at a California-based job platform stepped in.
In another paper, Google’s Gemini analysed the effects of San Francisco’s 2020 policy that cut towing fees for low-income drivers. The AI played a major role in data processing but repeatedly fabricated sources, according to researchers from the University of California, Berkeley.
05:04
China creates analogue AI chip said to be 1,000 times faster than Nvidia GPU
China creates analogue AI chip said to be 1,000 times faster than Nvidia GPU
The event’s co-organiser James Zou, a researcher in AI at Stanford University, said the conference accepted 47 papers from more than 300 submissions, all with AI systems listed as sole first author and taking the lead in the research and writing process.
