The capability-reliability gap is the distance between what an AI system can do once and what it can do dependably. Narayanan and Kapoor use the label for a pattern they say recurs, and they connect it to why AI agents have struggled to automate real-world tasks.
The answer, in one line
The distance between what an AI system can do once and what it can do dependably. Narayanan and Kapoor use the label in AI as Normal Technology (Knight First Amendment Institute, 2025) and say it has been a major barrier to AI agents automating real-world tasks.
Definition#
Capability-reliability gap: the difference between a system’s best performance on a task and the consistency needed to depend on it. The label is Arvind Narayanan and Sayash Kapoor’s, in “AI as Normal Technology” (Knight First Amendment Institute, 15 April 2025), where they write that this “‘capability-reliability gap’ shows up over and over.” The wording of the definition here is this research’s paraphrase.
What the authors say about it#
The passage sits in the first part of the essay, in a discussion of why the outside world limits how fast AI spreads. Narayanan and Kapoor say the gap has been a major barrier to building useful AI agents that can automate real-world tasks. Graded entry. They illustrate the same pattern in self-driving cars, where safety considerations limited how fast each stage of development could scale.
What kind of source this is#
It is an argued essay, published by a research institute and not peer reviewed. The gap is offered as an observation that recurs across domains. The part of the essay read for this page does not give a way to measure it. Evidence that bears on it, such as task-length benchmarks, is treated on its own page and was not tested against this label here.
Why it matters for the person overseeing the system#
This section is interpretation. A reviewer sees an agent succeed and has no way to tell from the output whether the success was reliable or lucky. The larger the gap, the more the human check carries, and the more it depends on the reviewer still being able to do the task. That is the point at which this term meets the argument about human oversight: a person approving work they could not check adds a signature and no safeguard. Narayanan and Kapoor make no claim about oversight in the passage read.
Key sources
- Narayanan, A. and Kapoor, S. (2025). AI as Normal Technology. Knight First Amendment Institute, 15 April 2025. Graded entry.
Related SuperSkills research#
See the METR time horizon, the delegation boundary map and the jagged frontier. The label is Narayanan and Kapoor’s, not a SuperSkills coinage.
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.
Explainer · SS-2026-423 · Graded against the published rubric · 1 argued perspective
Hirji, R. (2026). What is the capability-reliability gap?. The SuperSkills evidence base, SS-2026-423. https://thesuperskills.com/research/what-is-the-capability-reliability-gap. Last reviewed 12 October 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work