Discovery has compressed from years to weeks, but the FDA's structural evidence bar hasn't moved, and its record on drugs aimed at new biological targets sets the ceiling any AI discovered drug will face.
AI has compressed one clock in drug development into weeks. It has barely touched the other. The fastest novel-target, novel-molecule approvals the FDA has ever granted are the real ceiling on any AI-discovered drug for now, and that ceiling is measured in years, not months.
The "discovery clock" runs from a disease hypothesis to a synthesized, biologically active molecule. This is the part AI has reshaped. Generative chemistry, target identification models, and automated biology have pushed the path from idea to candidate compound into timeframes that would have been implausible a decade ago. In commentary published on Forever.ai, Insilico Medicine founder Alex Zhavoronkov frames this as a solved problem in chemistry and compute.
The "approval clock" runs from an Investigational New Drug (IND) application to FDA approval. This is the part AI cannot compress. The FDA requires evidence that a drug is safe and effective in humans regardless of how the molecule was found. Peer-reviewed work on clinical development times for innovative drugs puts typical timelines at multiple years from first-in-human studies to approval, even with expedited programs.
The reason the gap is structural, not computational, is the FDA's evidence standard. A novel target, meaning a biological mechanism no approved medicine has ever hit before, combined with a novel molecule, meaning a chemical entity with no prior safety database, is the hardest case regulators face. The agency's Novel Drug Approvals 2025 report and its Fast Track approvals page are the primary records for how this case plays out. Fast Track can shorten the FDA's review clock; it does not shorten the clinical trial clock.
This is where Zhavoronkov's own framing is useful. He is openly writing as an industry insider with a position to defend, not as an independent analyst. His timeline estimates in the seed piece were produced via "AI model orchestration," which is a hypothetical construct rather than a verified record. Treating those numbers as data would be a mistake. The legitimate critique he surfaces is real, however: industry pipeline narratives routinely conflate discovery speed with approval speed, and the conflation now repeats every time an AI-discovered molecule reaches preclinical status.
What the FDA record actually says: AI can compress the path from hypothesis to candidate molecule, and that part is demonstrated and continuing to improve. The FDA will still require Phase 1, Phase 2, and Phase 3 evidence for safety and efficacy, plus manufacturing and labeling review, and that part is structural and unlikely to change in the near term. The fastest-known novel-target approvals visible in the agency's annual Novel Drug Approvals reports define the practical ceiling for the first AI-discovered drugs to reach patients, and that ceiling runs in years, not months.
The watch item is not whether AI can find a molecule faster. It is whether any AI-discovered candidate can clear the structural evidence bar the FDA has set for the hardest approval case. Until one does, the two-clock frame is the right mental model for every AI-drug headline that follows.