Maier and colleagues at LMU Munich and the Munich Center for Machine Learning just ran a preregistered experiment on 704 adults practising fraction arithmetic with an LLM assistant. They pitted a reward that incentivised restrained AI use against a small UI signal that made each offload visible to the user in the moment.
The reward did nothing
Every enterprise I have worked with in the last year is designing the reward. Badges for asking fewer copilot questions. Dashboards that praise "smart AI use." Metrics that gamify restraint. This working paper measured that design pattern directly. It moved the offloading rate by an amount indistinguishable from zero (OR=0.66, p=.139) and did nothing to what participants could do without the tool afterwards (OR=0.76, p=.890).
The signal did the work. Metacognitive feedback halved the odds a participant would ask the assistant for the full answer (OR=0.47, 95% CI [0.22, 1.00], p=.026, one-sided). It raised the odds of getting the next unaided item right (OR=1.51, 95% CI [0.98, 2.33], p=.030, one-sided). Both intervals brush 1, so I would not sell precision. The direction and the mechanism are the point.
"Metacognitive feedback changed how the AI assistant was used, whereas the reward mainly reduced how often it was used."
The head of L&D who wanted a policy
Last month I sat with a head of L&D at a European insurer. She wanted a copilot policy for the underwriting team. Something firm. Something the CRO could sign. Half the graduates on the desk were pasting entire risk narratives into the tool and lifting the output back into the file. She was worried they would never learn the craft they were hired to do.
I asked what the interface told them when they pasted. Nothing. It accepted the text and returned the narrative. She wanted a rule that said "no more than 200 words per prompt." I asked whether she thought a rule would survive the deadline before quarter-end.
The paper matches what I have watched happen in rooms like hers. Restriction rules and reward schemes shift the dashboard while the skill stays where it was. What shifts the skill is a workflow that reflects the specific move back at the person at the moment they make it. "This is the fourth time you have handed off the full narrative on this class of risk this week." The message is boring. It costs nothing to build. It sits closer to a browser autofill notice than to a training programme.
The finding held in a task the study designed to make offloading measurable. About one in five items in the AI-only condition got asked for the complete answer. Fraction arithmetic is not underwriting. The mechanism is the same. A workflow where the model can hand you the answer. A person who has to retain the skill anyway.
The lever is what the workflow tells the person about what they are doing.
Executives keep reaching for policy because policy is the tool they have. Design is the tool that works.