Review Jagged Capabilities

Jagged capabilities are the uneven pattern of AI performance in which a system can excel at one task yet fail at another that appears similar in subject or difficulty.

An AI capability boundary is not a smooth line from easy to hard. A model may perform expert-looking work on one assignment, miss a simpler-looking neighboring assignment, and give equally confident responses to both. The edge of what it can do has cliffs and pockets.

A strong demonstration does not establish broad competence. Teams must test the actual tasks, variations, and failure conditions they plan to use because intuitive judgments about which nearby tasks should also work are often unreliable.

A model produces a strong market overview and persuasive interview plan, then gives a factually wrong recommendation on an adjacent analytical task that uses the same case materials. The polished earlier work did not prove capability across the whole workflow.

Jagged capabilities are not the same as random output or Model Drift. The uneven boundary can be repeatable at a given time. The lesson is that capability in one area does not justify assuming capability in a seemingly similar area.

Capability has cliffs, not a smooth edge.

* [[capability-provenance|Capability Provenance]] * [[emergent-capabilities|Emergent Capabilities]] * [[calibration|Calibration]] * [[model-card|Model Card]] * [[evaluation-gap|Evaluation Gap]]

* [Dell'Acqua et al., “Navigating the Jagged Technological Frontier”](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) — reports field-experimental evidence that AI assistance improved some knowledge-work tasks while worsening another apparently similar task. * [INFORMS, *Organization Science* record](https://ideas.repec.org/a/inm/ororsc/v37y2026i2p403-423.html) — summarizes the paper's concept of an uneven frontier of AI capabilities across tasks.

* Preserve task-specific evidence through [[capability-provenance|Capability Provenance]]. * Use [[evaluation-gap|Evaluation Gap]] to identify where a successful test has been stretched into an unsupported capability claim.