A new study has found that cancer diagnosis can be less accurate when AI is used to help, with people over relying on AI rather than their own opinion.

The study, from University of Warwick psychologists, asked trained people to diagnose cancer from signs of it in mammograms (breast X-rays).

It found that when people were told the accuracy of the AI with they were worse at correctly identifying cancer compared to when they were not told the accuracy.

Participants showed worse performance with increased false alarms, decreased sensitivity to symptoms of cancer, an increase in women being unnecessarily recalled in a clinical setting and a decrease in women being recalled who actually had cancer.

Overall, the study found that participants who were told how accurate their AI assistance was were much less likely to correctly identify cancer than those who had AI help but did not know how accurate the AI was.

The data also showed that improvement with AI help overall was limited, even when participants did not know how accurate it was.

Dr Melina Kunar, Turing Fellow at the Psychology Department at Warwick, led the study. She said:

“By using AI with different levels of accuracy, we have seen that people can easily veer into overreliance on it, rather than trusting their own opinion. People actually perform better when they aren’t aware of how accurate the help from AI is.

“Whilst AI continues to be used in more and more medical settings all over the world, often with amazing results, this study clearly shows that we must put safeguards in place to monitor accuracy.

“If we’re not careful, this incredible tool could result in over-reliance on AI against the expertise of our own medical professionals”.

The study on 645 participants used six types of AI with varying levels of accuracy, as well as asking some participants to try to diagnose the cancer with no AI assistance at all. The study used four different types of cancer to simulate a real clinical setting.

Psychologists believe that transparency about the accuracy of the AI assistant led to worse cancer detection rates as study participants became over reliant on AI. Participants were willing to accept the recommendation of the AI, even if it disagreed with their own judgement.

The new data has been released to mark the end of October's Breast Cancer Awareness Month.

The full paper can be read here.