The “AI is like a junior associate” analogy is making the rounds in legal right now. It’s one of the better ones. It gives non-lawyers a mental model for calibration that actually maps onto something real, and it gives lawyers a frame for explaining AI limitations without dismissing the technology entirely.
But it breaks in a place nobody’s talking about.
The analogy works on the surface because the failure modes look similar. A junior associate out of law school is eager, technically trained, and frequently wrong in ways that aren’t obvious from the output. They’ll give you a confident answer that misses the business context, cite the right rule and draw the wrong conclusion, draft a clause that’s technically defensible and practically unworkable. AI does all of these things too. So the instinct to treat it like a smart but green lawyer, useful under supervision and dangerous without it, is reasonable as far as it goes.
Where it stops going far enough is on the question of what actually produces judgment.
A junior associate has skin in the game. License, reputation, livelihood. When they get something wrong, there are consequences. Not the long hours or the anxiety that biglaw culture likes to treat as a rite of passage, but something more structural: the fact that being wrong costs them something real. That feedback loop, playing out over years of practice, is what eventually produces the thing we call legal judgment. The ability to read a situation, weigh competing risks, and give advice that accounts for what the client actually needs rather than just what the question technically asked.
AI has none of that. It cannot be sanctioned. It cannot lose a client. It has no professional reputation to protect or career to derail. And no amount of better prompting, more structured playbooks, or carefully developed institutional knowledge changes that structural reality. You can make AI more accurate. You cannot give it stakes.
This matters because the confidence doesn’t track the correctness. A junior associate who’s out of their depth usually knows it. They hedge, they check, they ask. The internal discomfort of not knowing is itself a signal that produces better behavior. AI produces the same confident tone regardless of whether it’s right or wrong, well within its training distribution or operating at its edges. The output looks the same either way.
So the more honest version of the analogy: AI is like a junior associate who can never be made partner, and who will never develop judgment, because judgment requires consequence and consequence requires skin in the game.
Useful at the right tasks, in the right structure. Genuinely dangerous if you mistake the confidence for something it hasn’t earned.
Derek Francisco is a licensed attorney and fractional in-house counsel. He writes on law, legal technology, and the economics of institutional change. Views expressed are solely his own.