A Realistic Expectation | AI Learning Guide

The conclusion here is uncomfortable but it follows directly from the path we drew. AI made the earliest stage of drug development faster and cheaper, and that is a real achievement. But that stage was never what limited how many medicines reach patients, so improving it changes the number of candidates entering the path without changing the number leaving it. For the picture to change, something has to move in the later filters: trials that run faster or answer questions more cleanly, evidence standards that can handle a design process nobody can fully explain, and funders willing to absorb the failure rate. Those move slowly because they involve institutions and people, not software. So expect change to arrive gradually and unevenly, first where the biology is well understood and the evidence bar is already clear. And when you see the next headline, ask the one question that cuts through it: which filter does this move, and by how much?

0:00 / 0:00

Auto-play

AI improved the stage of drug development that was never the bottleneck. Throughput is set by the slowest remaining filter, so visible change depends on movement in trials, evidence standards, and funding — not on further gains in molecule design.

Why change would arrive unevenly

The filters are not equally tight everywhere. In disease areas with well-understood biology, identifiable patient groups, and established evidence standards, candidates move through faster. Where the biology is unclear or trials are hard to run, the same candidate faces a much tighter filter. The same technology therefore produces different visible results in different areas, which is why a single overall verdict on AI drugs is misleading.

A question to carry forward

When you encounter a claim about AI-designed drugs, ask which of the four filters it actually moves and by how much. A claim that moves none of them is activity, not progress.