A new MIT study on AI productivity for research scientists just produced the most important split result in workforce science.

Source: Toner-Rodgers, MIT. arXiv:2412.17866. December 2024.

And did almost nothing for the rest.

This is not an opinion. It is a peer-reviewed MIT study, and it is one of the most important papers on the future of work published in the last five years.

The paper is called Artificial Intelligence, Scientific Discovery, and Product Innovation. Written by Aidan Toner-Rodgers at MIT. Published December 2024.

Not a survey. Not a simulation. A randomized study of 1,018 real scientists in the R&D lab of a large US firm. Half got the AI tool. Half did not. Then the researchers measured everything.

AI-assisted researchers discovered 44% more materials. Filed 39% more patents. Produced 17% more downstream products.

The numbers are extraordinary. But here is what is buried inside them.

AI did not lift the floor. It raised the ceiling.

Here is why. The AI automates 57% of idea-generation tasks, the part of scientific research that involves generating hypotheses, imagining possibilities, making conceptual leaps. The human’s job is no longer to think of ideas. It is to evaluate the ideas the AI produces.

Top scientists are exceptional at evaluation. They can rapidly identify which AI-generated candidates are worth pursuing.

Scientists in the bottom third? Their primary contribution was generating ideas. That work is now being done by the AI. The reallocation leaves them with less to do. Not more. Less.

The AI Productivity Split Nobody in Research Is Discussing

Every major research university has thousands of scientists across the full productivity distribution. Career paths are structured around the assumption that junior researchers grow into senior ones through the practice of idea generation.

If AI now generates 57% of the ideas, and the primary beneficiaries are the scientists already at the top, the middle and bottom of the scientific workforce is not being augmented.

It is being made redundant.

And unlike the layoffs in tech and customer service, nobody is publicly acknowledging it yet.

That silence is a leadership failure. Not a technology problem.

What This Means for Your Organization

The MIT study is not just about scientists. It is a data-precise preview of what AI does to any knowledge work pipeline with a productivity distribution.

The workers at the top of your distribution, the ones with the strongest evaluative judgment, will see their output multiply. The workers at the bottom, those whose primary contribution is task execution, face a different equation entirely.

The question is not whether AI will affect your talent pipeline.

The question is, which end of the distribution are your people on? And what are you building that develops evaluative judgment, not just task completion?

Because the organizations that answer that question are the ones whose people end up at the top of the split.