Define the task before judging the tool
An AI system that helps draft text, a model that classifies selected slide images and software used in a diagnostic workflow perform different tasks. Success in one does not establish suitability for another.
WHO identifies possible uses of generative AI in clinical care, education and administrative work, while warning about inaccurate, biased or incomplete outputs and automation bias. [1]
Ask what evidence supports the claimed use
My suggested evaluation starts with the intended task and the material tested. Was performance measured on selected textbook-like images or on cases resembling the service’s actual workload? Were rare lesions and difficult mimics represented? Was the evaluation independent of the training data?
Ask how errors were counted and whether the system can communicate uncertainty. An impressive overall accuracy figure may conceal a weakness in a clinically important category.
Keep the complete case in view
A selected image contains less information than a full slide set, and slides contain less clinical context than the whole patient record. A useful assessment should make clear what information the tool received and what conclusion that information can support.
Check outputs against the underlying evidence. Do not accept a confident explanation or a plausible reference without verification.
Build oversight into the workflow
WHO’s ethics guidance emphasises meaningful human oversight and attention to patient rights and data governance. [2] Before using a tool with clinical material, clarify permission, privacy, access and retention arrangements, as well as the intended clinical role.
Assign responsibility for reviewing outputs and recording how problems are handled. A person approving a report needs enough information and time to assess it meaningfully.
Use curiosity with discipline
AI deserves careful study. A constructive approach is to identify a specific problem, evaluate a proposed tool against that problem and review whether it improves the real workflow.
The important question is what evidence supports using this particular system for this particular task, with these safeguards and these limitations.
References
[1] World Health Organization. Guidance for large multi-modal models, release and risk summary. Read source
[2] World Health Organization. Ethics and Governance of Artificial Intelligence for Health. Read source