Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
AI productivity can hide a learning debt¶
Anthropic's December 2 study of its own workforce describes engineers and researchers completing more work, operating beyond familiar specialties, iterating faster, and addressing tasks that have previously gone undone. It draws on a survey, qualitative interviews, and internal usage data from an organization of unusually early artificial intelligence (AI) adopters.
The productivity story is real. So is a quieter organizational question: what happens to the learning that traditionally occurs inside the work now delegated?
Professionals develop judgment through exposure to difficult cases. They read unfamiliar code, trace failures, reconcile conflicting evidence, ask more experienced colleagues for help, and learn which tidy abstractions break in practice.
AI can accelerate that process when it explains, challenges, and supports exploration. It can also let a person bypass it.
Output can rise while capability thins¶
An engineer may complete a task in an unfamiliar part of the stack because an agent supplies the missing syntax and local pattern. That is useful boundary crossing. But successful completion does not establish that the engineer can diagnose the system during an incident, recognize a subtle error, or teach the approach to someone else.
The gap between output and internalized capability is a learning debt. Like technical debt, it may remain invisible until conditions depart from the happy path.
Organizations will miss it if productivity is measured only through volume and cycle time.
Keep desirable difficulty in the work¶
The answer is not to preserve inefficient work for its own sake. It is to distinguish toil from practice that builds judgment.
Teams can use different interaction modes deliberately:
- Delegate routine, well-understood transformations with strong tests.
- Pair on unfamiliar or high-learning-value work so the person sees the reasoning and verifies each stage.
- Reverse the usual pattern by having the person propose an approach before asking the system to critique it.
- Teach back consequential changes through reviews, walkthroughs, or incident exercises.
- Rotate ownership so expertise does not collapse around the few people who can still work without the tool.
Research on cognitive offloading shows that people routinely use external aids to reduce mental demand. The issue is not offloading itself. It is whether the environment preserves the knowledge needed when the aid is absent, wrong, or insufficient.
Measure resilience and growth¶
Alongside delivery measures, leaders should examine review quality, escaped defects, time to recover, ability to explain system behavior, mobility across roles, and the development of less-experienced staff.
Periodic unaided or lightly aided exercises can reveal whether important skills remain distributed. Incident reviews can identify where no one understands an agent-produced component. Career frameworks can reward verification, problem framing, and teaching—not merely visible output.
Design adoption across generations of workers¶
Experienced professionals can often use AI to compress work because they possess a rich mental model against which to judge the result. New professionals still need opportunities to build that model. If the organization removes the apprenticeship tasks without creating new learning paths, it borrows productivity from its future workforce.
AI can make work faster and broader. A thoughtful organization will ensure it also makes people more capable over time.
Sources and research trail¶
- Anthropic, “How AI Is Transforming Work at Anthropic” (December 2, 2025).
- Risko and Gilbert, “Cognitive Offloading” (2016).
- Parker and Grote, “Automation, Algorithms, and Beyond: Why Work Design Matters More Than Ever” (2022).
- Argote and Ingram, “Knowledge Transfer: A Basis for Competitive Advantage in Firms” (2000).