Four questions still without settled answers
- Do AI-designed molecules succeed more often in human testing than conventionally designed ones, or do they mainly reach the same outcome faster and more cheaply?
- How can we check whether a model has learned a causal mechanism rather than a correlation that happens to hold in the training data?
- How should the field generate the missing data in underrepresented tissues and patient groups, given that better models cannot create measurements that were never taken?
- What evidence from a model will regulators and clinicians accept when the prediction cannot be inspected or explained?
Why these are open rather than merely unanswered
Each question is open because the evidence needed to close it does not yet exist, not because no one has looked. The first requires years of clinical programs to mature. The second requires methods that are still being developed. The third requires data collection that is slow and costly by nature. The fourth depends on regulatory practice that is evolving alongside the technology. Recognizing this prevents the mistake of treating a current limitation as a permanent one, or a current claim as a settled result.