News · 14 August 2025
When AI helps and may also change the clinician: the colonoscopy deskilling study
In August 2025 The Lancet Gastroenterology & Hepatology published an observational study of endoscopists at several centres in Poland. After AI-assisted polyp detection was introduced, the adenoma detection rate in non-AI-assisted colonoscopies appeared to fall, from roughly 28% before AI exposure to roughly 22% afterwards. The authors suggested that regular reliance on AI might reduce clinicians' unaided detection skills.
How to read it
Caution is needed. The study was observational, not a randomised trial, so it cannot prove that AI caused the change. Other factors, such as changes in case mix, workload or behaviour, may contribute. The result needs replication.
But it points to a real design concern that human-factors researchers have discussed for years: automation complacency and skill decay. When a tool does part of a task reliably, humans pay less attention to that part.
Implications for design
If we want clinician and AI together to outperform either alone, the system must keep the human engaged and accountable. Some design ideas:
- Show evidence, not just flags. A highlighted region with a rationale invites assessment. A silent beep invites compliance.
- Make disagreement easy and informative. Capture overrides, learn from them and share outcomes with clinicians.
- Vary the workflow deliberately. Periodic unaided reads or second-reader checks help maintain skills and provide a baseline.
- Measure the human-AI team. Evaluate performance with and without AI, over time, in real settings, not only the model's accuracy on a test set.
- Train for failure modes. Teach users where a particular tool tends to be wrong.
What it means for explainable AI
Explainability is sometimes sold as a trust-builder. This study suggests a more demanding purpose: to keep experts thinking. A good explanation does not just justify the output. It provides the information a skilled person needs to confirm or contradict it.
The aim is augmented clinicians, not dependent ones. Studies like this are valuable because they ask what happens after deployment, which is where the evidence gap in medical AI is widest.
Written in August 2025; details reflect the situation at that time.