Generative AI has arrived in education with a promise that sounds almost too good to resist: students can get better answers, complete assignments faster, and receive personalized assistance whenever they need it. But a new study by David Strömberg, Victor Lei, and Yanhui Wu presents a striking challenge to that narrative. Drawing on 30 months of data from 26,811 Chinese secondary-school students, the study finds a sharp divergence between task performance and learning. Students using generative AI complete homework faster and earn substantially higher homework scores, yet their performance on closed-book examinations deteriorates—and the decline becomes even more consequential over time.
The implication extends well beyond schools. The study forces educators, managers, and technology leaders to confront a fundamental question about artificial intelligence: When AI makes a task easier, are we improving productivity—or eliminating the very effort through which capability is developed? This is the central management problem of generative AI. Productivity and learning are not always complements. In some settings, they can be substitutes.
The Productivity Paradox
The most revealing finding is also the simplest. After approximately six months of generative AI use, homework scores increase by 18 percent while homework completion time falls from roughly 64 minutes to 45 minutes. Yet monthly examination scores decline by about 20 percent. On the surface, the student appears to be performing better. Homework is completed faster, grades are higher, and the student has more discretionary time. Conventional productivity metrics would celebrate the result. But education is not fundamentally a productivity system. Its objective is human-capital formation.
A student who spends 45 minutes producing an excellent answer with AI may be demonstrating impressive task productivity. But if that student subsequently struggles to reproduce the reasoning without assistance, the productivity gain has come at the expense of capability. The assignment has been completed, but the learner has not necessarily been developed. This distinction is critical for organizations adopting generative AI. Businesses face a similar phenomenon when employees use AI to prepare analyses, write code, produce presentations, summarize documents, or make recommendations. A reduction in task time is immediately visible. A reduction in human capability is often invisible until the employee encounters a situation in which the AI is unavailable, unreliable, or unable to solve a novel problem. The educational setting therefore provides a powerful early-warning system for the broader economy.
The Hidden Cost of Outsourcing Cognitive Effort
The researchers identify homework outsourcing as the principal mechanism behind the learning penalty. Approximately 81 percent of AI users display behavior consistent with outsourcing: they complete assignments unusually quickly, receive high homework scores, and subsequently perform poorly on closed-book examinations. By contrast, students who continue to spend approximately the same amount of time on homework as non-AI users experience much smaller learning losses.
That finding changes the debate about AI in education. The question is not simply whether students should use AI. The more important question is how they use it. AI can function as a tutor, critic, explainer, or intellectual partner. It can also function as an outsourcing engine. The same technology can therefore produce radically different outcomes depending on the user’s incentives and behavior. This distinction has an important managerial parallel. Organizations frequently introduce AI with a supply-side mindset: Which model should we purchase? Which chatbot should employees use? How can we automate more tasks? The study suggests that these questions are incomplete. Leaders must also understand the demand side: What behavior will employees adopt when the technology is placed in their hands? If incentives reward speed and output volume, workers will naturally discover ways to minimize cognitive effort. If organizations want AI to develop human capability, they must redesign workflows so that employees remain cognitively engaged.
Figure: Adapted from Stromberg et al., 2026
The Six-Month Warning—and the Two-Year Problem
Perhaps the most important contribution of the study is its time horizon. The negative effect on monthly examinations emerges gradually and reaches its full magnitude after roughly six months. The consequences for high-stakes entrance examinations take considerably longer, reaching their full magnitude only after approximately two years. This creates a dangerous managerial illusion. Short-term measurements may suggest that AI is working exceptionally well. Students submit better assignments. Employees complete reports faster. Programmers produce more code. Analysts generate presentations more quickly. Managers see productivity rising and conclude that the technology is delivering value. The cost appears later.
The researchers interpret the gradual deterioration partly through the dynamics of skill formation: when foundational knowledge and practice are repeatedly bypassed, students may become less capable of absorbing subsequent material. Skills are cumulative; today’s learning creates the foundation for tomorrow’s learning. Organizations should recognize the same dynamic. An employee who routinely delegates analysis to AI may gradually lose analytical fluency. A junior programmer who accepts generated code without understanding its architecture may become less capable of debugging complex systems. A manager who relies on AI-generated strategic summaries may become less practiced at synthesizing ambiguous evidence independently. The danger is therefore not simply that humans become dependent on AI. It is that dependency can become self-reinforcing.
AI Can Compress the Skill Distribution—But Not in the Way We Expect.
Another surprising finding concerns who loses the most. The learning penalty is larger among high-achieving students. The difference between the highest and lowest pre-AI achievement groups is substantial: the most negative effect is approximately 24 percent for the highest tercile compared with 16 percent for the lowest tercile. This produces an unusual form of skill compression. Much of the broader AI productivity literature emphasizes how generative AI can help lower-skilled workers perform more like their higher-skilled colleagues. In education, however, the study suggests that excessive AI reliance may erode the capabilities of stronger students more sharply.
For institutions, this is strategically important. The students—or employees—who appear most capable of exploiting AI may also be among those with the most to lose if AI displaces rather than augments their cognitive development. The lesson is not to protect high performers from AI. It is to ensure that AI does not substitute for the difficult work through which high performance is created.
The Homework Problem Is Really a Measurement Problem
Perhaps the most uncomfortable implication is that AI undermines the informational value of homework. Historically, homework has served two functions. It provides practice, and it gives teachers and parents a signal about whether students understand the material. With generative AI, these two functions can separate. A student can produce excellent homework while understanding considerably less. The study finds that among AI users with above-median homework performance, higher homework scores can actually be associated with lower examination performance. This is a profound challenge for educational measurement.
The same problem is emerging inside organizations. If AI produces a polished report, does that report demonstrate the employee’s analytical capability? If an AI coding assistant generates functioning software, does the developer understand the architecture? If an AI system prepares a market analysis, can the manager challenge its assumptions? Organizations have traditionally measured outputs. The AI era requires them to measure capability as well.
The Answer Is Not to Ban AI
The study does not support a simplistic anti-AI conclusion. Indeed, its evidence suggests that students who use AI while maintaining similar levels of homework effort can achieve similar examination performance to non-users, while still receiving higher homework scores. That is an important distinction. The challenge is therefore not AI itself. It is AI without an appropriate learning architecture.
The researchers explicitly distinguish their findings from research showing that pedagogically designed AI tutors can improve learning. Their argument is that when teenagers freely choose general-purpose AI tools, the dominant behavioral pattern is often outsourcing rather than tutoring. For educators, the implication is to move from AI prohibition toward AI design. Students should be required to explain reasoning, defend answers, work through intermediate steps, and demonstrate knowledge without AI assistance. AI can become part of the learning process without becoming a replacement for it. For business leaders, the equivalent is to redesign jobs around AI augmentation rather than AI substitution.
Answer: A New Productivity Metric for the AI Era
The deepest lesson from this research is that organizations may need to rethink what productivity means. Traditional productivity asks: How much output can we generate with how little time and labor? An AI-enabled economy needs another question: How much human capability are we building while generating that output? That distinction could become one of the defining management questions of the next decade.
A company that uses AI to reduce every employee’s analytical effort may initially report spectacular productivity gains. But if junior employees no longer learn how to analyze, write, code, negotiate, research, or solve problems independently, the organization may gradually lose its internal capability pipeline. The best organizations will therefore treat AI not merely as an automation technology but as a capability architecture. They will deliberately decide which cognitive tasks AI should perform, which tasks humans must continue to practice, and where human judgment must remain visible.
The Chinese education evidence provides a warning that should not be ignored. Generative AI can make people look more capable before it actually makes them more capable. The strategic objective, therefore, should not be to maximize the amount of work AI can eliminate. It should be to maximize the amount of human capability AI can amplify. The difference may determine whether generative AI becomes the greatest productivity technology of the era—or a technology that quietly trades tomorrow’s capabilities for today’s efficiency.
Stromberg, David and Lei, Victor and Wu, Yanhui, The Generative AI Learning Penalty: Evidence from Chinese Secondary Education (June 02, 2026). Available at SSRN: https://ssrn.com/abstract=6868618