Does AI assistance build lasting skills or temporary abilities?

The exoskeleton metaphor (Wiles et al., on non-technical consultants performing data-science work with AI assistance) captures a category that the standard productivity-vs-learning vocabulary obscures. An exoskeleton is a temporary capability extension: it makes the wearer more capable while it is on, and the wearer reverts to baseline capability when it is removed. The capability resided in the device, not in the wearer, the whole time.

AI-enhanced work performance often has this structure. A worker without programming skill can produce code with AI assistance. A worker without analytical skill can produce analyses with AI assistance. The output looks like a skilled performance. But the capability resides in the AI plus the prompting interface, not in the worker. Remove the AI and the performance vanishes. The exoskeleton came off.

This is structurally different from skill. Skill, by definition, is a durable capability that persists when supports are removed. A skilled programmer can program without an IDE; the IDE makes them faster, but the underlying skill is theirs. An exoskeleton-augmented worker has no underlying skill to fall back on — the IDE was holding them up, not just speeding them up.

The distinction matters for how AI-augmented productivity should be measured and managed. A team that hits a productivity target with AI assistance has two possible interpretations: (a) they have learned to work with AI as a tool and have accumulated skills along the way, or (b) they are exoskeleton-dependent and would collapse if AI access were lost or degraded. The two look identical in productivity metrics; they look very different in resilience metrics. Organizations that do not distinguish them are accumulating dependency without knowing it.

The exoskeleton framing also clarifies what AI-augmented work should and should not be deployed for. Tasks where the work is the goal (output matters, durable capability does not) are well-served by the exoskeleton model. Tasks where the worker's capability development is the goal (apprenticeship, training, professional development) are harmed by it. Treating both task-types the same is a category error that the exoskeleton frame helps name.

Does AI assistance help workers learn lasting skills? is the empirical companion claim — Wu et al.'s finding that AI-improved performance on writing tasks does not transfer to subsequent independent performance is exoskeleton dynamics measured in-the-moment.

The strongest counterargument: with sufficient time, exoskeleton use produces some skill transfer through observation and adaptation. Possible, but the empirical findings consistently show no significant transfer. The exoskeleton remains an exoskeleton even after extended use.