The Boundary Is Not Where You Think

The pattern that confuses everyone

Someone uses AI to draft a genuinely difficult document. Excellent result.

Emboldened, they hand it something simpler. It fails, and fails oddly, in a way no competent person would.

They conclude the tool is unreliable, or that they got lucky the first time.

Neither. They ran into the shape of the capability boundary, which does not look like anything they expected.

Jagged, not a line

The intuitive model is a line. Easy tasks on one side, hard on the other, and as models improve the line moves.

The real boundary is jagged. It cuts back and forth in ways that do not track human difficulty at all.

Tasks that feel similar to us sit on opposite sides, because what determines the side is not difficulty. It is how well the task matches patterns present in training text, and that has almost nothing to do with how hard we find it.

This framing came out of field research where consultants using AI improved substantially on tasks inside the frontier and did measurably worse on tasks just outside it, without being able to tell which was which.

Roughly where the edges are

Generalising cautiously, and expecting exceptions:

Usually inside. Transforming text you supplied, summarising, restructuring, drafting from material, explaining well-documented things, translating, generating options.

Usually outside. Precise arithmetic, counting, current events, your private information, anything needing real-world verification, and knowing what it does not know.

Genuinely unpredictable. Domain-specific reasoning, multi-step analysis, niche technical work. These need testing rather than assumption.

That third category is the honest one. Much of what matters at work lands there.

Can it, versus can I rely on it

Here is the distinction people miss, and it is more important than knowing where the frontier is.

A model that succeeds eight times in ten is clearly capable. It is also unusable without checking, because you cannot tell which run is one of the two.

Demos show capability. Work needs reliability. The gap between them is where most disappointment lives, and it is why "I saw it do this" is such a poor basis for depending on it.