Why AI's promise to fix medicine faces a regulatory, not technical, ceiling

The author's core argument rests on a distinction often overlooked in Silicon Valley: the difference between intelligence and real-world change. The piece draws on two domains—medicine and general technology diffusion—to show that capability alone does not translate to impact. In biomedicine, the evidence is stark. Eroom's Law (the inverse of Moore's Law) documents that despite exponential gains in scientific tools, drug approval rates have declined for decades. The author traces this not to lack of computational power but to the grinding, unsexy work of clinical governance: trials lasting seven years and costing over $1 billion per drug; data access delays; and regulatory validation timelines measured in years even when underlying evidence exists. China's ascent in biotech—now capturing more than half of major Western pharma licensing deals—reinforces this: the shift came from regulatory reform enabling faster trials, not superior basic science.

The patent system further illustrates how incentives, not intelligence, shape outcomes. Because patents protect specific molecules rather than biological insights, companies avoid the target risk inherent in discovering new biology. Once a target is validated (the hardest part), fast followers design competing molecules and capture patent value, incentivizing herd behavior around de-risked targets. Even AI-for-biology startups—despite access to unprecedented computational tools and billion-dollar valuations—respond to these incentives, focusing on molecule optimization rather than the riskier, higher-impact work of validating novel targets. The author notes that surrogate endpoints and AI-driven biomarkers could theoretically make some trials 10 times faster and cheaper, but even this gain requires governance changes (faster FDA validation, easier data access) that lie outside the AI lab.

Beyond medicine, the author observes a broader pattern: real-world adoption of AI lags far behind headline capabilities. LLMs can generate complex code and spawn agents, yet customer service chatbots remain frustrating; GDP growth has not spiked despite recent advances; entry-level employment has held steady or grown despite repeated automation prophecies. Economist Tyler Cowen made exactly this prediction in 2023—that capabilities would arrive faster than real-world changes, GDP growth would disappoint, and job displacement would not materialize—and data have borne him out. The core insight is that diffusion is hard: a technology existing is not the same as being absorbed into institutions and translated into outcomes people care about. The author attributes Silicon Valley's blind spot on this to community insularity and the success bias of those who have bet on ideas that once looked insane, lending them outsized credibility on matters where lived experience in governance, policy, and institutional friction would serve better.