The thing that makes AI hard to use well is that it’s uneven in a way you can’t see from the outside. It’s not bad, exactly. It’ll nail a hard task and then botch an easier-looking one sitting right next to it, and it’ll sound exactly as confident either way. Learning where that line falls is quietly becoming one of the more valuable skills a person can have.
The jagged frontier
The idea comes from an experiment run by researchers at Harvard Business School and BCG, involving hundreds of consultants. They gave people tasks and let some use AI. On tasks that fell inside the tool’s strong zone, the results were great: consultants using it finished about 12% more tasks, and the quality was roughly 40% higher than the group working without it.
Then the researchers gave people a task designed to sit just outside that zone, a problem where the tool’s confident answer was actually wrong. Here the effect flipped. People using AI were about 19% more likely to land on the wrong answer than people working without it. The tool didn’t just fail to help. It actively led capable people astray, because it was persuasive and wrong at the same time.
They called this shape the “jagged frontier.” The tool’s ability isn’t a clean line where everything below is easy and everything above is hard. It’s jagged, strong here, weak an inch over, with nothing on the surface telling you which is which. That’s what makes it dangerous. The failures don’t look like failures.
The failures cluster where things get long
There’s a second pattern that helps you predict the jagged edges instead of just stumbling into them. Anthropic’s data shows model performance falls as tasks get longer and more complex. A quick, self-contained task is where the tool is most reliable. A long, multi-step problem with a lot of moving parts is where the errors accumulate and compound.
OpenAI’s own evaluation work points the same way. Their GDPval study found models did best on shorter tasks and got shakier as the work stretched out. So one rough rule of thumb: the longer and more tangled the task, the more you should treat the output as a first draft to be checked rather than an answer to be trusted.
That’s not a hard boundary, but it’s a useful prior. It tells you where to slow down and where you can move fast.
Knowing the edges is the paid skill
All of this adds up to something concrete for a career. The valuable person in an AI-heavy workplace is the one who knows when to trust the tool and when to double-check it. Using it constantly doesn’t make you that person, and refusing to touch it doesn’t either. The judgment sits in between.
That judgment is a skill, and it’s harder to build than it sounds, because the tool is engineered to sound convincing whether it’s right or wrong. Building the instinct takes reps: using it on problems where you can actually verify the answer, getting burned a few times, and gradually developing a feel for the kinds of questions where it tends to go off the rails. You won’t get there by reading about it. You get there by doing it and paying attention to the misses.
The people who have this skill are worth a lot right now, because most of their colleagues don’t. Plenty of people either trust the tool too much and ship its mistakes, or distrust it too much and forgo the help. The person in the middle, who knows the frontier, gets the speed without the errors.
What makes this hard to hire for, and therefore valuable, is that it doesn’t show up on a resume. You can’t list “knows when the model is lying to me” as a credential, and a certificate won’t prove you have it. It lives in the reps, in the accumulated memory of the times the tool sounded certain and was wrong. That kind of judgment is slow to build and hard to fake, which is exactly why the people who have it stand out.
It shows up when you’re asked to defend a call
Employers have started testing for exactly this. It comes up in interviews now, usually as a question about a time you used AI and then caught it being wrong, or a scenario where you have to say how you’d verify something the tool told you. What they’re probing is whether you have the judgment to supervise the output rather than just accept it.
That’s worth practicing before you’re doing it for real. Talking through how you’d check a suspicious answer, in a conversation where someone follows up and pushes on your reasoning, is the kind of thing that sounds obvious until you have to do it live and realize you’ve never put it into words. The candidate who can articulate where they’d distrust the tool signals judgment. The one who treats every AI answer as gospel signals the opposite.
The honest read
AI is a strange tool because it’s excellent and unreliable at once, often within the same hour, and it never tells you which mode it’s in. The frontier between the two is jagged and invisible, and the failures are the confident, polished kind that slip through.
The skill that matters, then, isn’t using the tool. It’s knowing its edges. That comes from real reps and honest attention to where it misled you, and it’s exactly the kind of judgment that’s getting more valuable as everyone else races to adopt the tool without learning where it breaks. Get good at that, and you’re the person others come to when they can’t tell if the answer is right.