AI Can Learn What You Did, But Not Why

The running coach put it like this: the day he heard Garmin had acquired his training platform, his first thought was "someone just bought my office." He'd used TrainingPeaks for nearly a decade. Every interval session trimmed, every recovery week added, every adjustment to every athlete's load. Ten years of records. What the platform doesn't have is why. When he cut two sets from an athlete's session, it was because that person had been working overtime for two weeks, their kid had just started school, and with eleven weeks to the race, the risk wasn't worth it. None of that is in the database. Garmin bought ten years of decisions. The reasoning behind them was never stored anywhere. Robert C. Martin, known as Uncle Bob in the software world, spent thirty years teaching engineers how to write readable code. Last month, he announced he no longer reads AI-written code. The announcement drew hundreds of thousands of shares, many of them critical. People said he was abandoning standards. But the next sentence was the point. He doesn't read it because he wrapped it in something else: layers of tests that the code has to pass before he accepts it. When the tests pass, he trusts the result. Who designed those tests? He did. He built them from nearly sixty years of writing code, reading bad code, and fixing broken tests. He poured that judgment into the verification layer. Then he let himself look away. The coach is spending his days now writing things down: what conditions call for lower intensity, what to say when an athlete is struggling, what to leave unsaid. Some decisions he couldn't fully explain until he stopped and worked through them. That's how he knows they weren't in the data. The coach is still writing it down. Uncle Bob wrote his into code sixty years ago.