AI still struggles to act like a real scientist

Researchers from Princeton and the UK tested top AI models on the task of writing scientific papers. They gave the tools six days and three thousand dollars in computing credits to do the work. Despite these resources the models could not produce quality results.

Experts reviewed the work and rejected every paper the AI created. While the technology can handle basic coding and technical tasks it lacks the judgment needed for real discovery. The models struggled to solve problems creatively or change direction when a plan was not working.

These findings challenge the idea that autonomous AI research is ready for prime time. Companies often claim their software is close to doing science alone. This study proves that human insight is still essential for meaningful progress.