The jagged frontier is Ethan Mollick and Fabrizio Dell'Acqua's framing that AI capability is uneven across tasks within a single model: frontier systems are superhuman on some tasks, human-parity on others, and weak on tasks that appear similar to the first two. The boundary is not smooth or intuitive, and users cannot reliably predict where it falls without empirical testing. Andrej Karpathy's term "jagged intelligence" describes the same phenomenon. See Ethan Mollick for the principal.
Naming and origin
Two parallel coinages describe the same phenomenon. The academic framing, "Jagged Technological Frontier," originates in the BCG and Harvard Business School field experiment by Dell'Acqua, McFowland, Mollick, et al. (Navigating the Jagged Technological Frontier, September 2023). The study assigned 758 BCG consultants to tasks inside and outside the model's capability boundary. On tasks inside the frontier, GPT-4-assisted consultants outperformed controls on speed, quality, and creativity. On tasks outside the frontier, GPT-4 users performed worse than controls, trusting model output on tasks where it was systematically wrong.
The mainstream framing, "Jagged Intelligence," was coined by Andrej Karpathy, an OpenAI founding researcher and former Tesla Autopilot head, in a 2024 social media post:
"Some things work extremely well (by human standards) while some things fail catastrophically (again by human standards), and it's not always obvious which is which."
Karpathy contrasts this with the human brain, "where a lot of knowledge and problem solving capabilities are all highly correlated and improve linearly all together, from birth to adulthood."
Cade Metz's NYT explainer (April 2026) offers a canonical illustration of the contrast. On the high-capability side, AI answered 5 of 6 questions at the International Math Olympiad (summer 2025). On the low-capability side, leading chatbots told a user to walk to a repair shop 50 meters away, failing a basic spatial and common-sense judgment. The practical upshot is that a task being similar to one the model handles well does not predict that it will handle the new task well; the boundary is jagged, not a smooth gradient. (Source: nytimes.com)
Why the frontier is jagged
Several mechanisms are proposed for the unevenness. Training data density is non-uniform: some topics are overrepresented, such as common programming languages, mainstream English prose, and widely-discussed history, while others are sparse, such as niche legal doctrines, specialized lab protocols, and non-English minority languages. Task structure matters more than task appearance: a reasoning task that looks like a math problem may actually require retrieval, and a retrieval task that looks like reading may actually require reasoning, with models handling these asymmetrically. Reinforcement post-training shapes the frontier deliberately, as RLHF and reasoning post-training reward certain patterns while adjacent patterns inherit capability unevenly. Tool access changes the shape: with code execution and web search, the frontier expands in specific directions such as verified computation and current facts while remaining unchanged elsewhere. Finally, prompting sensitivity contributes, because near the boundary small prompt changes can flip outcomes, which is itself a form of jaggedness.
Implications for benchmarks
Jaggedness is offered as the central reason aggregate benchmarks can mislead. Benchmark overfitting means capability on MMLU, HumanEval, SWE-bench, and similar tests describes the model on those distributions, not at the task a given user cares about. Average scores hide holes: a model at 85% average might be 99% on most items and 5% on a predictable slice, a productivity problem if a workflow lands in that slice. Saturation of one benchmark tends to spawn another, including ARC-AGI, FrontierMath, GPQA, and HLE, each exposing a different facet of the same underlying jaggedness.
This is the reason GDPval Paper (OpenAI, Oct 2025) and similar economic-value benchmarks are complicated: real tasks draw from many cells of the frontier, and a model's value depends on which cells a job lands in. See AI Benchmarks and Evaluation.
Metz's April 2026 NYT explainer notes the release of ARC-AGI-3 by François Chollet, described as an upgraded benchmark "asking for solutions to hundreds of gamelike puzzles without providing a single instruction for how to solve them. All of the puzzles can be solved by an average, untrained person, but the leading AI systems fail to master any of them." This extends the ARC-AGI series (ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems) and exemplifies the same jaggedness: models that answer math olympiad problems cannot solve puzzles a child can solve. (Source: nytimes.com)
Implications for deployment
Mollick's derived prescription is that humans must learn the boundary empirically, task by task. Effective users develop task-specific priors about where to trust the model through trial-and-error adoption. Mollick's "centaur" and "cyborg" taxonomy of human–AI task division depends on correctly identifying which side of the frontier a task sits on. Organizations that invest in this learning loop are said to outperform those that either over-trust or under-trust the model, carrying training and adoption costs; see AI and Productivity and Enterprise AI Deployment Gap. A deskilling risk arises near the frontier: when humans stop checking because the model usually gets it right, errors concentrate at the boundary. See AI Deskilling.
Dynamics over time
The frontier moves outward, but unevenly. A new model might expand capability on reasoning tasks while barely moving on something like long-context coherence. This implies that the shape of the frontier is as important as its size, that generation-over-generation comparisons should disaggregate by task type, and that forecasts of when AI will be able to do a given task are more tractable when the task is specified precisely than when framed as generic capability.
Tension with uniform-capability narratives
The jagged frontier sits awkwardly with several framings that treat capability as smoother. AGI framings treat capability as a scalar (AGI Timelines). Scaling-law extrapolations (Scaling Laws) forecast aggregate loss curves. AI as Normal Technology arguments model deployment as smooth industrial adoption. Jaggedness is compatible with average progress but adds variance, so the framings can be reconciled, though they pull in different directions on policy and forecasting.
Debates and positions
Two questions remain contested. On whether the frontier becomes smoother as models scale, some evidence from reasoning post-training and multimodal unification suggests it does, while Emergent Misalignment and AI Is Really Weird evidence suggests it gets weirder before it gets smoother. Whether jaggedness is a transient artifact of current training recipes or an enduring property of learned systems is unresolved.
Relationships
- depends-on: AI Benchmarks and Evaluation
- supports: AI Deskilling, Enterprise AI Deployment Gap, AI and Productivity
- contradicts: strong-form uniform-AGI framings in AGI Timelines
- related: Ethan Mollick, GDPval Paper (OpenAI, Oct 2025), AI as Normal Technology, AI Is Really Weird
See also
- Ethan Mollick — principal.
- Cade Metz, NYT, April 2026 — Karpathy attribution; canonical NYT explainer; ARC-AGI-3 mention (Source: nytimes.com)
- GDPval Paper (OpenAI, Oct 2025) — economic-value benchmark complicated by jaggedness.
- AI Is Really Weird — companion framing (irregular capability surface).
- A Guide to Which AI to Use in the Agentic Era — practitioner implications.