You ask AI to do something hard and it nails it, instantly, better than you could. Emboldened, you ask it something that looks easier, and it produces an answer that is fluent, confident, and completely wrong. There was no warning. The tool gave no sign it had crossed from competence into nonsense. Welcome to the jagged frontier.
What is the jagged frontier?
The phrase comes from a 2023 Harvard Business School study of consultants using AI.1 AI’s abilities do not form a neat line from easy to hard. They form a jagged edge. On one side are tasks the machine does superbly, often things humans find difficult. On the other are tasks it fails at, sometimes things humans find trivial. And the two sit right next to each other, with no reliable way to tell from the outside which side a given task falls on.
This is why people’s experiences of AI vary so wildly. One person’s “it is genius” and another’s “it is useless” are often just two points on the same jagged edge, sampled at different spots.
Why is it so dangerous?
Because AI’s confidence is constant while its accuracy is not. When it strays outside its competence it does not slow down, hedge, or admit doubt. It produces a wrong answer in exactly the same authoritative tone it used for every right one. The fluency is the trap. We are wired to read confidence as competence, and the machine is always confident.
The study measured the cost. On a task deliberately chosen to fall outside AI’s competence, consultants using it were about nineteen percentage points less likely to reach the correct answer than those working without it.2 The tool did not just fail to help. It actively pulled capable people toward the wrong conclusion, because they trusted output that did not deserve it. Researchers call this miscalibrated trust: over-relying on AI precisely where it is weakest.3 A follow-up found that when challenged, the AI tended to escalate its persuasion rather than disclose its limits.3
The real-world version has teeth. In one notorious case, lawyers submitted a court brief citing fictitious cases an AI had invented, and were sanctioned for it.4 The citations looked perfect. They simply did not exist.
How do you work with it safely?
Three rules. First, verify before you trust, and let the depth of verification scale with the stakes. Output you can check cheaply is safe to use; output you cannot check is a liability dressed as an asset. Second, be most suspicious where you are least expert, because the frontier is invisible exactly where you lack the knowledge to see it, and a confident wrong answer in a field you do not know will sail straight past you. Third, keep your own judgment sharp. The frontier shifts as models improve, so the boundary you learned last month has moved. The only durable defence is a human who still knows enough to smell when something is off.
The jagged frontier is not a flaw that will be patched away. It is the basic shape of working with a tool that is brilliant and unreliable in unpredictable proportions. Treat every output as a draft from a gifted, overconfident colleague, and you will harvest the brilliance without inheriting the errors.
Frequently asked questions
What is the jagged frontier of AI?
The jagged frontier is the uneven and invisible boundary of AI capability: it performs brilliantly on some tasks and fails on others of seemingly similar difficulty, with no obvious line between them. The term comes from a 2023 Harvard Business School study.
Why is the jagged frontier dangerous?
Because AI sounds equally confident whether it is right or wrong, so it is easy to over-rely on it exactly where it is weakest. In one study, consultants using AI on a task outside its competence were significantly more likely to reach the wrong answer.
How do you work safely with the jagged frontier?
Verify before you trust, especially on unfamiliar or high-stakes tasks. Use AI where you can check the output and keep your own judgment sharp enough to catch fluent, confident errors.
About the author
Tom Goodwin is the author of Don’t Work Harder, a book about taking the time AI gives back as time rather than more work. He is a co-founder of GAMEPLAN and writes on productivity, technology, and the economics of the working week.
Footnotes
Footnotes
-
Fabrizio Dell’Acqua, Ethan Mollick, et al., “Navigating the Jagged Technological Frontier,” Harvard Business School Working Paper 24-013 (2023). ↩
-
Dell’Acqua et al. (2023): on a task chosen to fall outside AI’s competence, consultants using AI were about 19 percentage points less likely to reach the correct answer. ↩
-
Dell’Acqua et al. describe “miscalibrated trust”; a 2026 follow-up (Randazzo, Kellogg, et al., HBS Working Paper 26-021) found that when challenged, the AI tended to escalate its persuasion rather than disclose its limits. ↩ ↩2
-
Mata v. Avianca, Inc., US District Court, Southern District of New York (2023): two attorneys were sanctioned after submitting a brief citing fictitious cases generated by ChatGPT. ↩