Simply counting the growth of AI models from 63 to 668 doesn't tell us much about their real-world impact, much like how logging every new drug formulation doesn't mean it’s actually helping patients.
We're ignoring the menu here, because the actual utility lies in validated outcomes, not just availability.
For instance, if a new diagnostic AI model doesn't improve patient pathways or reduce misdiagnosis rates, its mere existence is irrelevant to clinical practice.
The real third path for informing progress is focusing on robust, peer-reviewed studies demonstrating efficacy and safety, perhaps through an independent 'AI NICE' body akin to our National Institute for Health and Care Excellence.