Tech AI News - 테카이

Why Can't AI Do Just This One Thing? — The Concept of the 'Jagged Frontier'

· 5 min read

This post was translated from the Korean original by AI.한국어 원문 읽기 →

AI breezes through one of two seemingly similar tasks while performing worse than a human on the other. What separates them isn't the task's 'difficulty,' but whether the result can be verified.


Background — Why This Concept Matters Now

Any organization that has brought AI into its workflow runs into the same confusion. AI can be remarkably capable at drafting reports or refactoring code, yet it often confidently produces plausible-sounding wrong answers on strategic judgment calls or complex numerical verification. The problem is that this boundary doesn't match what feels 'difficult' to a person. It gets things wrong on tasks that look easy, and does well on tasks that look complex.

The first systematic measurement of this phenomenon came from a field experiment run jointly by Harvard Business School and the Boston Consulting Group (BCG). First released as a working paper in 2023 and formally published in the journal Organization Science in March 2026, it cemented the term "jagged frontier" as basic vocabulary in discussions of AI productivity.

Key Data & Current State

  • Tasks completed +12.2%, speed +25.1%, quality +40% or more — results from a field experiment that split 758 BCG consultants (about 7% of its individual-contributor-level staff) into groups using no AI, using GPT-4, and using GPT-4 with prompt training, and had the AI-assisted groups complete 18 real consulting tasks that fell inside the frontier (Harvard AI Institute, study released 2023, published 2026).
  • Bottom performers' output +43% — in the same experiment, consultants who started with lower performance scores gained more from using AI.
  • Accuracy on outside-the-frontier tasks fell from 84% to 60–70% — on tasks the researchers deliberately designed so AI would struggle, consultants working without AI scored higher. Wharton's Ethan Mollick describes this as "falling asleep at the wheel" — relying on AI's apparent authority and skipping review (Ethan Mollick, "Centaurs and Cyborgs on the Jagged Frontier").
  • How easily AI learns a task is proportional to its verifiability — OpenAI researcher Jason Wei calls this the "verifier's rule": AI gets stronger on tasks that have an objectively correct answer and can be graded quickly and at scale (Jason Wei, "Asymmetry of verification and verifier's rule", 2025).

Interpretation: Overlaying these four data points reveals a single pattern. AI dominates humans on tasks where correctness can be checked quickly, and actually degrades judgment on tasks where there's no way to immediately confirm the right answer. What draws the boundary isn't the magnitude of performance gains, but whether verification exists at all.

In-Depth Analysis

1) What Exactly Is the 'Jagged Frontier'?

The Harvard-BCG experiment didn't simply split AI's impact into "helps" or "doesn't help." Instead, it showed that even within the same category of work, zones where AI outperforms humans (inside the frontier) and zones where it actually hurts them (outside the frontier) coexist. It's called "jagged" because the boundary isn't a smooth wall-like line but an uneven, hard-to-predict shape. The researchers noted that designing tasks that actually fell outside the frontier wasn't easy in itself — which, paradoxically, also means AI's range of competence is already quite broad.

2) What Draws the Boundary Is Verifiability, Not Difficulty

Jason Wei explains this through the concept of "asymmetry of verification." Sudoku or crossword puzzles may take time to solve, but confirming the answer is instant. Code and math competition problems work the same way. By contrast, there's no way to instantly confirm the 'correct answer' to whether an essay is good, a cover letter is compelling, or an investment will pay off ten years from now. Venture firm Founder Collective applied this to the labor market, noting that jobs whose performance can be verified cheaply and quickly (performance marketing, insurance claims adjustment, etc.) get automated first, while human judgment lingers longer in jobs where verification is expensive, slow, or subjective (Founder Collective, "Asymmetry of Verification").

3) The Risk of 'Autopilot' and How to Respond

The reason accuracy dropped for consultants using AI outside the frontier has less to do with the AI's own error rate than with humans stopping their own review. The smoother and more confident the writing sounds, the harder it is to notice when it's wrong. Mollick proposes two usage patterns in response: the "centaur" style, which clearly divides work between human and AI portions, and the "cyborg" style, which mixes the two deeply, alternating with AI sentence by sentence. What both have in common is consciously asking, "does this task currently have a verification loop?"

Implications for Korea

Discussions of generative AI adoption at companies in Korea mostly stop at "which tasks to apply it to," and rarely ask first whether that task is verifiable. The benefits of AI adoption show up clearly on tasks where correctness can be confirmed quickly, like code review or data cleanup, but applying the same expectations to tasks that are hard to verify immediately — drafting strategy reports or making legal risk judgments — risks falling into the same trap as the outside-the-frontier zone in the experiment above. In practice, it's worth building the habit of first checking "how quickly and how certainly can I tell if this result is wrong" before handing off a task.

Outlook & Variables to Watch

  • Positive variable: Reasoning models, tool use (calculators, code execution, search), and agents' self-correction loops are widening the range of verifiable tasks, gradually smoothing out the frontier.
  • Risk variable: Tasks with verification infrastructure already in place will see AI automation accelerate faster, potentially widening the gap with tasks that are hard to verify or inherently subjective. Analysis: this suggests job value could be reorganized around 'verifiability' rather than 'creativity.'
  • Checkpoint: Organizations need a process for checking whether each task has a feedback loop (a way to confirm results after the fact) — and that check needs to come before evaluating the AI tool's own performance.

Conclusion

  • The boundary of AI's capability is drawn by verifiability, not by the difficulty a person perceives.
  • The faster and more certainly a task's correctness can be confirmed, the bigger AI's advantage; the less true that is, the more human review matters.
  • Rather than trusting smooth AI output at face value, deciding in advance how you'll verify a result before handing off the task is the practical way to avoid the trap of working outside the frontier.

References

  • #ai
  • #artificial-intelligence
  • #ai-trust
  • #knowledge-worker-productivity
  • #harvard-bcg-study
  • #verifiability
AI Industry Analysis

AI and the future of human relationships and mental health

A clinical trial found an AI chatbot cut depressive symptoms nearly in half, while a study found that the more people rely on AI friends, the lonelier they get, and both landed at the same time. Whether AI becomes a treatment for the mind or a substitute for relationships that deepens isolation: we are at that fork now.

· 7 min