Co-Intelligence Summary

A second seat has appeared beside the human at the workbench, and it is already covered in fingerprints.

Ethan Mollick, a professor at the Wharton School, writes about artificial intelligence from close range. He is not watching from a safe tower. He has made students, founders, managers, and teachers use these systems while the paint is still wet.

His core claim is simple and unsettling: the best way to understand AI is to work with it, because its real shape only appears in use. Large language models can draft, code, tutor, brainstorm, and reason in strange flashes, but they can also make false claims with a straight face.

The weird part is not that AI can write; the weird part is that it can join a meeting without knowing what a meeting is.

Mollick’s answer is not panic or worship. It is a new kind of shared craft.

In 2023, researchers gave GPT-4 to more than 700 Boston Consulting Group consultants and asked them to complete real business tasks.

The results looked almost unfair when the work sat inside the model’s strengths. The consultants using AI finished faster and produced work that outside graders rated much higher. Then the same tool made people worse on tasks just outside its reach, because it gave confident help in the wrong direction.

Mollick calls this the “jagged frontier.” AI is brilliant at some tasks that feel hard to humans, and weak at some tasks that feel easy. It may write a decent marketing plan, then fail at a simple logic problem. It may explain a legal idea clearly, then invent a case that never existed.

The danger is not that AI is useless. The danger is that it is useful in patches.

This matters because most organizations want a neat map. They want to label jobs as safe or unsafe, automated or human. Mollick says the real line cuts through tasks, not job titles. A lawyer, teacher, designer, or analyst may have parts of their work boosted and other parts quietly poisoned.

The frontier is jagged, which is a polite way of saying the monster trips over carpet.

The workbench image changes here. AI is not just another tool lying near the hammer. It is a tool whose handle moves when you grab it. You need to test each task against the system, not guess from a distance.

That changes the practical question. The first question is no longer, “Can AI do my job?” The better question is, “Which parts of my job sit inside the bright zone, and which parts sit near the teeth?”