How AI Is Reshaping Skill Formation

“AI decouples performance from competence. People can now produce correct outputs without constructing the mental models normally built through effort, error, and feedback.”

Isabelle Hau, Executive Director of the Stanford Accelerator for Learning

As AI assistants become embedded across education and technical work, a quiet but consequential inflection point is emerging: for the first time, productivity is scaling faster than learning. A new study from Anthropic reveals a striking tension at the heart of AI adoption. The findings suggest that while AI accelerates output, it can simultaneously erode the cognitive processes that underpin real skill formation.

Rather than presenting AI as inherently harmful or beneficial, the research surfaces something more structural: how humans interact with AI now determines whether expertise compounds or atrophies. Interaction patterns that encouraged explanation-seeking and conceptual inquiry preserved learning outcomes, while workflows that outsourced thinking entirely led to weaker competence. The implications extend well beyond software development, touching workforce training, education systems, and organizational design at a moment when AI-mediated learning is rapidly becoming the default.

To unpack what this research signals for AI literacy, workforce development, and institutional learning, we speak with Isabelle Hau, Executive Director of the Stanford Accelerator for Learning, about why this moment requires deliberate strategies to ensure AI strengthens human capability rather than eroding it.

Isabelle C. Hau is a visionary leader dedicated to transforming how we nurture and educate our children.

As executive director of the Stanford Accelerator for Learning, she leverages brain science and technology to champion innovative, effective, and inclusive learning solutions. 

Previously a successful impact investor, Hau led the US education practice at Omidyar Network and Imaginable Futures, where she invested in mission-driven organizations that have reached millions of learners. She is the author of Love to Learn: The Transformative Power of Care and Connection in Early Education. She also writes a popular weekly newsletter, Small Talks. 

At Stanford University, Hau teaches the class Design to Equip Learners in Under-Resourced Communities. She serves on the board of EDC and Design Tech High School, on the steering committee of the EdSAFE AI Alliance, the Brookings Institution Global AI Taskforce, and the World Economic Forum’s 4.0 education alliance.

Named one of the 100 most inspiring women by Harvard Business School and a finalist for the World Education Medal, Hau has also received distinctions in early childhood education and human-centered artificial intelligence. She co-starred with Grover of Sesame Street. Her lifelong professional goal is to bring the love of learning to each and every child.

Anthropic’s findings point to a deeper change in how learning occurs. Isabelle Hau explains that “AI decouples performance from competence. People can now produce correct outputs without constructing the mental models normally built through effort, error, and feedback.” A task can now be completed successfully even if the user never works through the reasoning that previously produced understanding.

This alters what it means to become skilled. As Hau puts it, “Skill formation shifts from executing tasks toward framing problems, evaluating outputs, and refining processes. The danger is mistaking fluent production for deep understanding.” Execution becomes easier, while judgment becomes central. The user must decide whether an answer is reliable rather than simply generate it. When institutions evaluate only the finished product, fluency can appear indistinguishable from mastery.

The experimental evidence reflects this mechanism. 

“In a randomized experiment led by Anthropic where developers learned a new Python library, participants using AI completed tasks faster, but scored about 17% lower on conceptual understanding, debugging, and code reading.” Speed improved, yet comprehension weakened. “The contrast is revealing: AI improved immediate performance but weakened independent capability. Students learned how to obtain answers, not how to generate them.”

Similarly, she noted that in a large Wharton-led study of nearly 1,000 high-school math students in Turkey, students using a standard AI assistant performed 48% better during practice, yet 17 percent worse on the exam once AI was removed.

The results also clarify that the problem is not the presence of AI itself. Hau notes, “Importantly, when AI was redesigned as a tutor providing hints rather than solutions, the learning loss disappeared, showing the issue is not AI itself but how cognition is structured around it.” When the interaction requires prediction, revision, and reasoning, understanding persists.

The shift, therefore, concerns how people practice thinking. AI changes the cognitive work users perform. If the system replaces reasoning, skill formation declines. If it supports reasoning, learning continues.

The separation between output and understanding creates a practical challenge for schools and workplaces. Both rely on observable results to judge progress: completed assignments, faster turnaround times, and higher work volumes. These indicators improve quickly with AI assistance, yet they do not show whether a learner can reason through a problem independently or apply knowledge in a new situation.

Hau explains why the shift is difficult to detect. “Faster work can conceal weaker cognition. In the experiments above, the most assisted participants were the fastest yet learned the least; productivity masked declining skill formation.” Speed becomes a misleading signal. A student who finishes quickly or an employee who delivers more output may appear more capable even as the underlying mental model remains thin.

To see the difference, evaluation must move beyond completion. Hau says, “Institutions should therefore measure cognition, not just completion:” looking for the ability to explain reasoning without AI, transfer knowledge to unfamiliar problems, diagnose and correct errors, and formulate meaningful questions. Each of these behaviors requires a person to understand the process rather than rely on a generated response. They reveal whether the individual can operate without assistance.

One indicator carries particular weight. “We at the Stanford Accelerator for Learning are particularly interested in the theme of learning through creation.” Creation requires organizing ideas, testing them, and revising based on feedback. A learner who builds something must make choices and justify them. By contrast, completion alone can occur without reflection if a system supplies the solution.

This distinction matters because institutions increasingly optimize for efficiency. Dashboards track productivity, not reasoning. As AI becomes routine, the absence of visible struggle can be misread as mastery. Understanding remains observable only when assessment captures how a person thinks, not just what they produce.

Whether AI strengthens or weakens learning depends less on the technology itself than on how people use it. Hau draws a clear boundary: “Key distinction: whether AI extends cognition (including social learning), or replaces it.” The same system can support understanding or bypass it depending on how a learner engages.

Some patterns keep thinking active. Hau points to “AI as coach (hints, scaffolds, critique)” along with “prediction before seeing AI output” and “iterative drafting and reflection.” Each requires a user to form an expectation, compare it with the system’s response, and revise their reasoning. “Comparing multiple solutions and different perspectives (critical thinking)” and “required justification” deepen the process by making the learner explain why one answer works better than another. Through “creation and iteration,” knowledge is applied and tested rather than copied. She also includes “social connection augmentation,” where discussion with peers remains part of the learning loop. The system supports thought, but does not complete it.

Other patterns move in the opposite direction. Hau lists “AI as answer engine,” “copy-edit-submit workflows,” and “passive summarization.” These remove the need to represent the problem internally. “Skipping problem representation” and “delegating reasoning entirely” eliminate the stage where understanding normally forms. She adds “substituting human relationships, hindering social learning,” noting that collaboration itself contributes to cognition. When interaction stops at accepting a generated response, the user practices retrieval rather than reasoning.

The distinction clarifies why outcomes vary across settings. The technology is constant, yet the cognitive work changes. When the system prompts prediction, comparison, and explanation, it reinforces learning. When it supplies final answers, it replaces the processes that build expertise.

The implications extend beyond classrooms. If interaction patterns determine learning, then educators, employers, and platform designers all shape whether AI develops expertise or weakens it. Hau frames the goal directly: “Design for cognitive participation, creation, and human connections, not convenience.” Systems optimized only for speed remove the very effort through which people learn.

Several practices follow from that principle. “Require explanation, not just output.” When users must justify an answer, they reveal whether reasoning is present. “Stage assistance (delay full solutions)” keeps the learner engaged long enough to attempt the problem before receiving help. She also advises to “track reasoning processes, not only results,” and to “evaluate learning improvement over time. Focus on learning as a journey rather than an end test.” These approaches shift attention from the finished product to the development of understanding.

Incentives matter as well. Organizations tend to reward efficiency, yet Hau argues they should “reward judgment, critique, and creativity/originality.” Work that involves evaluation and decision-making reinforces expertise more than work that simply reaches completion. She also recommends “[making] cognitive participation social (e.g., small groups),” preserving discussion and collaboration as part of learning rather than replacing it with automated interaction.

The objective is not to remove AI from learning environments. Instead, it is to design systems that make thinking visible. As Hau concludes, “The goal is not for AI to do the work/outsource cognition, but to make thinking visible and restore learning as a journey.”