AI is making your people worse at their jobs

The output is there — the reports are getting written, the code is being produced, the analysis is landing in inboxes. What is harder to see is what is happening underneath. The person writing the report is spending less time thinking about it. The engineer producing the code is less certain about what it does. The analyst summarising the research has retained less of it than if they had worked through it themselves.

This is not speculation. In April 2026, researchers from Carnegie Mellon, Oxford, MIT, and UCLA published a controlled study of 1,222 participants completing reasoning tasks with and without AI assistance. After ten to fifteen minutes of working with an AI that answered their questions directly, participants performed measurably worse on the same class of problems unassisted — and were more likely to give up when the problems became hard. The effect appeared within a single session. It was consistent across three separate experiments.

The finding that changes the practical conversation is this: the people who used AI for direct answers showed the largest decline. The people who used it for hints and prompts did not show the same pattern. The degradation is not in using AI. It is in the specific way of using it — outsourcing the thinking rather than augmenting it.

The same effect has been documented outside knowledge work entirely. A 2025 study in The Lancet found that endoscopists who routinely relied on AI assistance during colonoscopies saw their detection rate for precancerous lesions drop from 28.4% to 22.4% when the AI was removed. These are experienced medical professionals. The tool had not replaced their capability. It had quietly substituted for it — and the substitution was invisible until the tool was gone.

The challenge for any CEO reviewing an AI rollout is that the metrics organisations use to track AI adoption are measuring exactly the wrong thing. Licence utilisation. Tasks completed. Time saved per process. All of these can be rising while the foundation underneath them is eroding. The quarterly numbers stay green. The people delivering them are, in a specific and measurable sense, becoming less capable of delivering them without the tool.

This is not an argument against AI tools. It is an argument about how you deploy them. The distinction the research draws — between AI that answers and AI that prompts — is one that organisations can act on. Whether your people are using AI to do their thinking or to sharpen it is a management question, not a technology question. It is also one that I rarely hear asked.

The cost of not asking it is not visible this quarter. It accumulates in the gap between what your people can do with the tool and what they can do without it — a gap the organisation does not discover until the tool is unavailable, the contract expires, or the model changes in a way that breaks the workflow built around it.

The right question is not whether your teams are using AI. It is whether anyone is tracking what they are getting better at — and what they are no longer able to do without it.