A Verification Hardness Framework for AI in Drug Development

Artificial intelligence (AI) is accelerating the generation of hypotheses, predictions, and candidate solutions throughout the drug-development process. However, AI-generated outputs must still be verified through scientific review, experimentation, clinical evaluation, and regulatory assessment. This creates an AI verification gap between the speed of generation and the speed of verification. To characterize this gap, we introduce a Verification Hardness Framework that evaluates verification across five dimensions: ground-truth accessibility, verification cost, verification latency, human-judgment dependence, and consequence of error. In this framework, we developed a scoring system and applied it to 23 published studies representing five stages of drug development: target identification, drug discovery, preclinical study, clinical study, and regulatory review. Our analysis found that target identification has the lowest Verification Hardness Index (VHI), followed by patient screening and molecule discovery, indicating relatively lower verification barriers to AI adoption. This framework provides a methodical approach for evaluating the AI verification gap and may help predict where AI can achieve faster adoption in the drug-development workflow. We argue that verification hardness should be a key metric for evaluating AI in drug development alongside the conventional measures of model performance.