Assistance changes behavior
A user might accept a correct suggestion, reject it, misunderstand it, or follow an incorrect suggestion. The final decision depends on the interaction, not just the model's answer.
The interface, timing, explanation, and user's experience can all influence that interaction.
A relevant comparison
For a decision-support tool, a study might compare the intended users with and without access under otherwise appropriate conditions. The comparator should resemble the real alternative, including existing information resources.
| Outcome | What it adds |
|---|---|
| Decision quality | Whether the combined process is accurate |
| Time and workload | Whether the assistance changes effort |
| Error transitions | Whether correct decisions become incorrect or vice versa |
A vignette is not a full clinic
Controlled case studies can isolate aspects of reasoning. Real care adds incomplete information, competing tasks, follow-up, and consequences. Results from one setting need appropriately limited claims.
FDA's good-machine-learning-practice principles emphasize the human-AI team's performance. The practical implication is straightforward: evaluate the product as people will use it. “The model scored higher than clinicians” and “clinicians benefited from the model” are different claims requiring different evidence.
