Jagged Frontier – Part 3: Blind Trust at the Edge

The jagged frontier is that unpredictable boundary between what Generative AI (GenAI) can and can’t do well. Two tasks that feel equally hard to a person can fall on opposite sides; a large language model (LLM) might ace one and perform poorly on the other.

In the Harvard study on the GenAI jagged frontier, this distinction mattered. On tasks inside the frontier, where GenAI thrives, people got more done, faster, and with higher quality. But on tasks outside the frontier, relying on GenAI hurt: performance dropped substantively compared to humans working alone.

One clue was something the researchers called “retainment;” how much of the LLM’s output people copied straight into their work. Inside the frontier, high retainment often meant better results. Outside it, the same behavior meant confidently producing the wrong answer.

That’s a reminder that navigating this frontier isn’t just about prompts. It’s about judgment. Knowing when to lean on GenAI – and when to pause, question, and take back the wheel.

You can find Part 1 here and Part 2 here.