The Jagged Frontier

In the [YOU] on AI Field Guide

[YOU] on AI describes the orange pill moment as a recognition of capability that exceeds expectation—the encounter with a system more powerful than the discourse had prepared the practitioner to expect. The jagged frontier explains why the orange pill moment is so often followed by a chastening encounter with confident failure: the same system that produced the extraordinary capability sits adjacent to a zone of confident wrongness, and the adjacency is invisible until you walk into it. The cycle's prescription of ascending friction—using the tool to reach harder problems while maintaining the judgment to supervise the output—is a prescription for navigating the frontier deliberately: staying inside the wall on high-stakes tasks where confident failure would be catastrophic, pushing toward the edge on lower-stakes explorations where surprising capability is the discovery sought.

The frontier also explains the structure of the Boston Consulting Group study that grounds Mollick's empirical claims. Inside the frontier, augmented consultants dramatically outperformed unaugmented ones. Outside the frontier—on a task designed to fall beyond reliable AI capability—augmented consultants performed worse than unaugmented ones, led astray by fluent plausible wrongness. Both results were necessary for an honest account. The frontier is not a limitation to be apologized for; it is the defining structural feature of the technology, and understanding it is the precondition for using it well.

Origin

The jagged frontier metaphor was developed by Mollick through his sustained personal experimentation with AI systems beginning in 2022, formalized through the Boston Consulting Group field experiment he co-authored in 2023, and named and articulated in Co-Intelligence (2024). The empirical grounding distinguishes it from purely impressionistic accounts of AI's uneven capability: the BCG study provided controlled conditions in which the inside-the-frontier and outside-the-frontier performance could be measured separately, confirming the jaggedness in professional-grade tasks with a large sample.

The concept builds on the broader observation, made by practitioners across many domains, that AI capability is uncorrelated with human intuitions about difficulty. Tasks that appear to require sophisticated reasoning sometimes fall inside the frontier; tasks that appear routine sometimes fall outside it. The reason is structural: the frontier reflects the distribution of the training data and the optimization objective, neither of which tracks human intuitions about difficulty. A model trained on a corpus that contains many examples of strategic business analysis and few examples of arithmetic verification will be more reliable on the former than the latter, regardless of which a human would consider harder.

Key Ideas

The invisibility of the boundary. The frontier cannot be seen from the outside. A practitioner cannot know in advance which of two apparently similar tasks falls inside the wall and which falls in the treacherous zone just outside it where failures wear the costume of success. This invisibility is the reason Mollick insists on experimentation as a personal discipline: the knowledge of where the frontier lies in a specific domain is tacit, local, and idiosyncratic, and it cannot be acquired by reading or by watching others. It must be earned through contact.

The peculiar danger of fluent failure. AI failure outside the frontier is often more dangerous than the obvious mechanical failure of previous software systems. A spreadsheet that divides by zero produces a visible error; a language model that invents a plausible but false citation produces a fluent document that a non-expert cannot easily distinguish from a correct one. The confidence that makes the system's inside-the-frontier outputs so useful is precisely what makes its outside-the-frontier failures hard to detect. The human in the loop must maintain enough domain expertise to sense the small wrongness that the machine cannot sense because it has no purchase on truth, only on plausibility.

The frontier expands but does not smooth. As models improve, the wall expands, swallowing tasks that used to fall outside it. But the frontier does not become smooth as it expands. New capabilities bring new adjacencies with the zone of confident failure; the dangerous edge moves outward with the capability boundary. This means the skill of sensing where the edge lies never becomes obsolete—it simply needs to be continuously updated as the frontier moves. The fourth of Mollick's rules—assume this is the worst AI you will ever use—implies that the frontier tracking must be an ongoing practice, not a one-time map that can be memorized and retired.

Falling asleep at the wheel. The most precise failure mode the frontier generates is what Mollick calls falling asleep at the wheel: the human in the loop, lulled by the system's consistent competence inside the frontier, relaxes supervision at the moment the system crosses into the outside-the-frontier zone. The very competence that makes the AI valuable inside the frontier is what makes it dangerous near the edge: the fluency is constant, the confidence is constant, and only the truth-tracking drops, invisibly. The BCG study documented this inversion empirically: the same access to AI that dramatically improved performance on in-frontier tasks degraded performance on the deliberately out-of-frontier task.