The emergence of AI tools that support healthcare professionals, such as radiologists and surgeons, in performing their work is generally praised. These AI tools contribute to more accurate and better diagnoses in radiology and often ensure that surgical procedures can be performed with greater precision and less risk of side effects or damage to healthy tissue. An AI model for independently training surgeons is even being tested. However, recent research shows that there is also a danger in using AI tools. Researchers looked specifically at the impact of AI on the skills of healthcare professionals who perform colonoscopies.
New research published in The Lancet Gastroenterology & Hepatology warns that routine use of AI in colonoscopies may undermine the ability of experienced endoscopists to detect adenomas without AI support. Colonoscopy is crucial for detecting and removing precancerous growths, an essential mechanism in the prevention of colon cancer.
Effect of long-term AI use still unclear
Although previous research has shown that AI increases adenoma detection rates, the effect of long-term use on the clinical skills of physicians remained unclear. To explore this, a total of 1,443 colonoscopies were analysed between September 2021 and March 2022 in four centres in Poland.
Of these procedures, 795 were performed before the introduction of AI and 648 after, in each case without AI support. All of the endoscopists had performed more than 2,000 procedures. Results showed a significant decrease in adenoma detection in non-AI procedures: from 28.4 to 22.4 per cent, representing a relative decrease of 20 per cent and an absolute reduction of 6 percentage points. AI-assisted procedures showed an adenoma detection rate of 25.3 per cent.
Evidence of a negative effect of AI?
Prof. Marcin Romańczyk emphasises that this may be the first evidence of a negative effect of regular AI use on the clinical skills of healthcare professionals. Prof. Yuichi Mori adds that previous studies may have produced biased results because endoscopists became accustomed to relying on AI, which could have caused their normal performance to deteriorate.
The authors acknowledge that there are limitations: it was an observational study, and the cohort consisted only of experienced doctors, which limits its generalisability. Further research is therefore needed, both with less experienced healthcare providers and in other medical fields where AI is playing a growing role.
Dr Omer Ahmad of UCL emphasises that this clinical evidence should temper enthusiasm for rapid AI adoption. While AI remains promising for improving healthcare outcomes, he warns of a potentially insidious erosion of fundamental clinical skills.
