Clinical competency in the age of AI safe, equitable and effective practice in AI-augmented healthcare

Summary

A recent white paper, Clinical Competency in the Age of AI,  presents findings from a systematic narrative synthesis of 445 studies examining clinical competency requirements in AI-augmented healthcare. It addresses a structural gap in how current competency frameworks prepare clinicians for AI-assisted practice.

In addition to examining the breadth of research into clinical risks associated with use of AI in clinical care, the research analysed 23 existing AI competency and capability frameworks, including the NHS Health Education England AI and Digital Healthcare Technologies Capability Framework and the DECODE international consensus framework.

It found that across all reviewed frameworks, the competencies most critical for frontline patient safety—critical appraisal of AI recommendations, detection of biased outputs, governance escalation, and protection of professional moral accountability—are largely limited to awareness statements for frontline users. Clinicians are expected to understand what AI is. They are not equipped to practise safely with it.

The white paper proposes a five-domain competency framework, specified across three career stages, that translates intersecting AI risks into assessable clinical capabilities for practising clinicians.

Content

Key findings

• AI erodes clinical reasoning without competency safeguards. The Budzyń et al. (2025) multicentre colonoscopy study provides the first real-world evidence: adenoma detection rates fell from 28% to 22% among endoscopists after three months of AI assistance. The skill had not been assessed. It had not been exercised. It had atrophied.
• Cognitive overload drives uncritical AI acceptance. Alert override rates of 90–96% have been documented in deployed clinical AI environments—a workforce adapting to unsustainable demand by reducing evaluative effort. AI tools assessed as safe under controlled conditions carry significantly higher risk in busy, overstretched environments where they are most needed.
• Governance infrastructure is inadequate. Over 70% of NHS trusts lack documented clinical safety assurance for deployed AI tools (Oskrochi et al., 2025). Clinicians in these settings carry full personal professional accountability for AI-assisted decisions without the institutional infrastructure that should underpin them.
• Risks compound, but are treated as parallel separate risks. Time pressure increases automation bias severity. Automation bias accelerates deskilling. Deskilling undermines safety governance capacity. Equity failures concentrate where burnout is highest and training resources most limited. Current frameworks miss these feedback loops.
• Healthcare-specific competency frameworks are insufficient. Over 75% of medical students receive no formal AI education. Where training exists, assessment tools lack specificity for healthcare contexts. This research defines what AI clinical competency requires: technical understanding, critical appraisal, equity awareness, safety governance knowledge, and professional identity maintenance, integrated rather than treated as separate modules.
• Implementation guidance remains fragmented. Governance frameworks address safety. Education frameworks address training. Workforce research addresses burnout. Each treats its domain rigorously while missing the system dynamics. This research consolidates evidence into practical principles for curriculum development, organisational deployment and regulatory strengthening.
• Harm concentrates in those least able to detect it. The populations most at risk from biased AI outputs are served by clinicians least equipped to recognise that bias, in settings least able to monitor it. This convergence is structural and will not be resolved by improving AI performance alone.