The Hidden Cost of Trusting AI

There’s no shortage of advice about how to write better prompts. But I’ve started to think prompting isn’t the skill that will separate great AI users from average ones. Judgment is.

One idea I’ve come across that really stuck with me is something researchers call the “Jagged AI Frontier”(https://mitsloan.mit.edu/ideas-made-to-matter/working-definitions/what-is-jagged-ai-frontier). It’s based on a simple observation: AI isn’t consistently intelligent. Instead, its abilities are uneven.

Sometimes it can solve remarkably complex problems in seconds. Other times it struggles with tasks that seem almost trivial. That’s why so many people have contradictory experiences with AI. One person uses it to write an excellent business proposal and is amazed. Another asks it to verify a few facts, spots several mistakes, and concludes the technology is useless. In reality, both experiences can be true.

The problem begins when we assume that because AI performs brilliantly on one task, it will perform equally well on another. For example, a consultant became faster and produced better work when AI supported tasks that suited its strengths. But once the work moved beyond those strengths, performance actually declined because people trusted AI more than they should have.

When AI consistently produces high-quality work, it’s easy to become less critical of its outputs. Instead of carefully reviewing every response, we begin to trust that the next one will probably be correct. Over time, this can subtly shift our role from active evaluator to passive reviewer.

I’ve been thinking about what this means for the future of knowledge work. First, what happens to expertise if AI can immediately elevate the performance of less experienced employees? AI can be an incredible learning tool, but it can also become a shortcut. If junior employees rely on AI before developing their own mental models, will they build the judgment needed to recognize when AI is wrong?

Second, there’s the hidden cost of verification. Reviewing AI-generated work isn’t free. It requires concentration, domain knowledge, and critical thinking. In some cases, checking an AI-generated output may take just as long or even longer than completing the task independently. If that happens, the productivity gains we often associate with AI begin to look far less straightforward.

The more capable AI becomes, the easier it is to lower our guard. We stop questioning outputs, stop checking facts, and gradually move from active collaborators to passive reviewers. That’s why I don’t think the future belongs to the people who can write the cleverest prompts. It belongs to the people who know when AI is likely to be right, when it’s likely to be wrong, and when human expertise should take over.