Read the evidence like a workforce leader, not a buyer

In May 2025 the Massachusetts Institute of Technology publicly withdrew its support for a paper by one of its doctoral students that had claimed large gains in scientific discovery from AI tools at a materials-science laboratory. The paper had circulated widely, been covered in major outlets, and been cited by senior figures in business and policy. MIT concluded that it had no confidence in the provenance or validity of the data. The author is no longer at the institution.

I raise this not to make a point about one paper but about the people who cited it. Most were not careless. They were doing what busy leaders do: taking a striking result from a credible institution and using it to support a decision they were inclined to make anyway. The decision in question was usually about the workforce. That is why the skill of reading this evidence belongs to whoever is responsible for the workforce, and why it has to be a skill rather than a reflex.

Three habits distinguish a workforce leader's reading from a buyer's.

The first is to ask who paid. A survey reporting that 86 percent of chief human resources officers consider integrating AI agents central to their role was commissioned by a company that sells AI agents. It may be accurate. It is also sentiment from an interested source, and it tells you what executives say when asked by a vendor, not what their organisations do. The same applies to the hours-saved slide in every demonstration. A workforce leader notes the source before the number.

The second is to ask what was measured. The strongest evidence we have on generative AI at work comes from randomised field experiments: customer support agents about 14 percent more productive, software developers completing roughly 26 percent more tasks, a person with AI at Procter & Gamble matching a two-person team without it. These are real and they are peer-reviewed. They also measure bounded tasks, over weeks or months, with human review built into the workflow and no one losing their job. They say what an assistant does for a worker. They do not say what an autonomous agent does to a workforce, because that experiment has not been published yet. A workforce leader uses them for what they show and refuses to let them stand for what they do not.

The third is to ask what has been replicated. The Danish finding of almost no labour-market effect and the American finding of a decline among young workers in exposed occupations are both single studies from 2025. Both are careful. Neither has been reproduced by an independent team. A workforce leader treats them as the first points in a series, consistent with a hypothesis, and not as a basis for closing a graduate programme or opening one.

Why does this belong to the Chief Workforce Officer rather than to a research function or to the technology leaders? Because the decisions the evidence feeds are workforce decisions: how many people to hire at the entry layer, which roles to redesign, where to deploy agents and where to hold back. The person accountable for those decisions must be the most careful reader of the evidence in the building, because everyone else who brings them a number has a reason for bringing it.

The retracted paper did its damage through people who wanted it to be true. The best defence against the next one is a role whose job includes not wanting any particular result.