Can hospitals trust AI with numbers?

Hospitals around the world are racing to bring artificial intelligence into everyday care, but a new study suggests the technology may still stumble over some surprisingly basic tasks. Researchers in the United States tested nine popular AI language models on simple hospital record questions, such as counting patients or sorting information from emergency department visits. The results were far from reassuring.

Using records from 50,000 real emergency visits, scientists found that most AI systems struggled when asked direct questions in plain language. Even some of the best-known models produced inaccurate answers once the amount of data increased. One leading system saw its accuracy fall sharply when handling larger tables, raising concerns about relying on AI alone for hospital planning and patient management.

The picture improved when the models were paired with tools that allowed them to generate and run computer code. Under those conditions, a few systems delivered almost flawless results. Researchers say the findings show that AI still needs strong human oversight and technical support before it can safely handle critical hospital administration tasks independently.

The researchers warned that speed and convenience should never replace accuracy, especially in healthcare settings where small mistakes can quickly create larger problems overnight.