Insights
The AI Contribution Penalty at Work
An analyst I read about recently spent a year on one problem. She sat with the people who ran the process, collected the complaints nobody had written down, drafted alternatives, learned which of them would break something downstream, and eventually produced a change worth a great deal to her employer. Somewhere in that year she used an AI assistant for a few days of drafting.
When the work went to senior leadership, her manager asked her to foreground the AI. In the meeting he interrupted her presentation to say the tool had done it.
I keep returning to that moment, because it is not really a story about a bad manager. It is a story about an organisation that had lost the ability to see a human contribution once a machine had touched any part of it. There is now a surprising amount of evidence that this error is systematic.
In four preregistered experiments with 4,439 participants, researchers found that people who use AI at work both anticipate and receive worse evaluations of their competence and motivation than people who do the same work without it. The penalty was not imagined. It showed up in real assessments of job candidates.
A separate programme of eleven experiments with 3,846 participants found that evaluators pay AI-assisted workers less — across task types, employment statuses, payment formats, and even when output quality was held constant or statistically controlled. The mechanism was measured directly: people judged that AI-assisted workers deserved less credit, and that judgment accounted for roughly sixty per cent of the pay reduction. In thirteen further experiments on disclosure, actors who declared their AI use were trusted less than those who did not, across supervisors, subordinates, professors, analysts and creatives. The explanation was not accuracy but legitimacy: disclosure marked the work as socially inappropriate.
Adopt these tools
Put those together and you have an organisation quietly running two incompatible instructions. Adopt these tools. Do not be seen using them.
The rational response, for anyone paying attention, is concealment. And concealment is expensive in ways that never appear on a balance sheet: nobody shares what worked, nobody reports what failed, and the organisation’s collective learning about its most important new capability happens entirely in private.
The accounting error underneath
What interests me more is the accounting error underneath. We have built evaluation systems that measure output and infer effort from it. When output arrives faster, the inference machinery concludes that less was contributed. But in the analyst’s story, almost nothing she actually contributed was executional. She contributed a year of attention. She contributed knowing which stakeholder’s objection was real and which was territorial, and the judgment to discard three alternatives that looked good on paper. The AI drafted. She decided.
An evaluation system that cannot distinguish between those two things will systematically underpay exactly the capability that is becoming scarce.
One hopeful finding in this literature
There is one hopeful finding in this literature. The penalty softens considerably when the human remains visibly involved — reviewing, adapting, contextualising, taking responsibility for the result. Which suggests the problem is not that people refuse to credit AI-assisted work. It is that most of what a person does in AI-assisted work is currently invisible, and organisations have not built any way to make it visible.
That is a design task, not a values problem: describing contributions in terms of judgment exercised, not hours displaced. Asking, in reviews, what the person decided rather than what they produced. A manager being able to say what a colleague added, specifically, and being embarrassed if they cannot.
An organisation that cannot name what its people contribute will eventually stop paying for it. Not out of malice. Out of blindness.