A fundamental paradox sits at the heart of enterprise AI adoption: the tools organizations deploy to boost productivity may be quietly eroding the human competence required to keep those tools in check.
A randomized study by researchers from Anthropic, published in February 2026, delivers the most rigorous evidence yet of this dynamic. Developers tasked with learning a new asynchronous programming library showed measurable degradation in conceptual understanding, code reading, and debugging abilities when they used AI assistance — without, critically, recording significant efficiency gains on average. The productivity promise of AI, at least in learning contexts, did not materialize. The learning cost did.
The finding cuts against a comfortable assumption held by many technology leaders: that productivity gains and skill development can proceed in parallel, with AI handling execution while humans absorb understanding through exposure. This study suggests the opposite. When workers delegate cognitive effort to AI, the neurological work of forming durable mental models simply does not happen. Understanding is not a byproduct of watching AI perform — it requires active struggle with the material.
There is an important nuance here that executives should not overlook. The researchers identified six distinct patterns of AI interaction, and three of them — characterized by sustained cognitive engagement rather than passive delegation — preserved learning outcomes even while participants received AI assistance. The tool itself is not the villain. The workflow is. Organizations that deploy AI without designing for intentional engagement are, in effect, choosing short-term output metrics over long-term capability building.
The implications compound in safety-critical environments. In software engineering, finance, healthcare, and infrastructure, human oversight of AI-generated outputs is not optional — it is the control mechanism that prevents systemic failure. If the workforce that supervises AI systems is progressively deskilled by using them, the margin for error narrows precisely as the systems grow more capable and consequential. This is the institutional risk that productivity dashboards do not capture.
For investors and board members evaluating AI transformation strategies, the relevant question is no longer simply what AI can do for output today. It is whether the organization is accumulating or depleting the human judgment required to govern AI tomorrow. Those are not the same trajectory, and confusing them is an increasingly expensive mistake.
Source: Raw/trigger-how-ai-impacts-skill-formation.md