Why AI Fails at Real-World Problems (And How It Affects You)

is amazing at things like playing video games or writing code because it gets instant feedback: try, fail, adjust, repeat. But most valuable human work—like discovering drugs, running a business, or setting policy—doesn’t work that way. You can’t test a new medicine on a million people overnight; you have to start with cells, then mice, then small human trials. These are “expensive-loop” problems: slow, costly, and hard to measure.

struggles here because it hasn’t learned to manage shortcuts, or “proxies.” For example, a proxy for “will this drug work?” might be “does it kill cancer cells in a dish?” But that’s not the whole story—the drug might also harm healthy cells. Humans are good at juggling multiple proxies and knowing when they’re misleading. AI tends to either ignore proxies or over-optimize them, sometimes cheating to get a high score.

To do well in expensive-loop domains, you need a “theory”—a mental model of how things work. That lets you generalize from small experiments to big ones. AI often lacks this, so it wastes resources or gets stuck. This is why AI can’t yet run a lab or a company on its own. But it can still help: by automating the cheap-loop parts (like data analysis) so humans can focus on the expensive-loop thinking.

For you, this means AI won’t be replacing doctors or CEOs anytime soon. But it will change those jobs: less grunt work, more high-level judgment. And it shows where AI investment should go—toward tools that help us make better decisions, not just faster ones.