If AI coding tools are copilots, what makes a good copilot? | Michael 🤗 Hoffman

If AI coding tools are copilots, what makes a good copilot? A copilot is a fellow expert who can, when needed, perform many of the pilot’s duties. The copilot offers a second pair of hands and eyes, improving the pilot’s efficiency and reliability. Crucially, the copilot is subservient to the authority of the pilot. The pilot is ultimately responsible for the safety of all on board. This is the implicit promise of AI “copilots” too: They are, we are led to believe, expert code-writers who serve our will. Even better, unlike a real copilot, they never demand pesky perks like rest, wages, and human dignity. In season 2 of The Rehearsal, Nathan Fielder identifies a pattern common to many plane crashes. When a pilot makes a dangerous mistake or oversight, copilots are scared to speak up until it’s too late. They mistake their lower place in the hierarchy for prohibition against critical feedback. But unconditional faith in superiors and sycophancy have no place in a cockpit. They are deadly. I was thinking about this today when Cursor told me a simple refactor I asked for was “brilliant”. It wasn’t. Will Cursor tell me “no”? Will Copilot question my reasoning? Or will they just tell me how great I am while they help me crash the plane? Another metaphor we often hear is that using a coding assistant (also a metaphor, by the way) is like working with a junior developer. This analogy is deeply troubling because it threatens to dehumanize actual early career devs. But it also undersells junior developers just like the “copilot” metaphor misunderstands what makes a good copilot. A junior developer on a team with a decent culture still asks you to explain why. And you learn by explaining. Sometimes you learn that the thing you were doing is the wrong thing to do. Automated coding assistants don’t do that. Be careful out there.