Coined by Dell’Acqua and colleagues from a pre-registered study of Boston Consulting Group consultants, the jagged frontier describes something counterintuitive about AI capability: it doesn’t map onto task difficulty the way human intuition expects. Some genuinely hard tasks sit inside the frontier, where the model performs at or above expert level. Some deceptively simple-looking tasks sit just outside it, where the model confidently produces wrong answers. From the outside, both kinds of task can look identical.
The practical danger is on the outside edge. Workers who have seen a model succeed repeatedly on hard, inside-the-frontier work develop justified trust, then extend that trust to a task that merely resembles the ones the model has been good at. In the BCG study, consultants using AI on a task positioned just outside the frontier were measurably less accurate than consultants working without it. Confidence and correctness didn’t move together.
The frontier also isn’t fixed. It shifts with every model release, so a task ruled out six months ago may now be inside it, and one that used to work may not survive a model swap unchanged. The only reliable way to know which side of the line a specific task sits on is to test that task directly, not to reason from a general impression of what AI can do.