Jagged Frontier

The jagged frontier is the idea that AI capability has an uneven, unpredictable boundary: tasks that look equally hard to a human fall on opposite sides of it. A model may draft a nuanced strategy memo flawlessly, then fail a simple counting or ordering task a child could do. The line between "AI is great at this" and "AI quietly fails at this" is jagged, not a neat difficulty gradient, and it doesn't match human intuitions about what's easy or hard. The term comes from a 2023 study of knowledge workers using AI. For builders, the practical implication is that you cannot reason about whether a model will handle a given task by analogy to a similar-seeming one — apparent difficulty is a poor predictor. Practical note: test each capability you plan to ship on real examples rather than assuming that strength on one task transfers to a neighboring one. Map where your specific use case sits on the frontier empirically, and re-check when you change models.