Generative AI adoption yields uneven productivity gains, new student experiment shows
New experimental study finds access to generative AI like ChatGPT improves student output on average, but adoption and benefits vary widely across tasks and users.
Study finds uneven productivity gains from generative AI
A controlled experiment conducted by researchers at Leuphana University Lüneburg and the Fraunhofer Institute shows that generative AI raises average task performance, but the benefits are far from uniform. The research, involving 210 university students, tested creative problem solving, text comprehension and quantitative reasoning with and without access to ChatGPT. The results highlight that task type and individual user choices shape how much generative AI actually boosts productivity.
How the experiment was structured
Participants were randomly assigned to two groups: one with direct access to a large language model and one without, with strict controls to prevent cross-use in the control group. Researchers logged whether and how students used the AI tool, then had 1,176 external raters evaluate anonymized answers to measure quality. This design allowed the team to separate the effect of mere access from real-world usage and to quantify performance differences across task categories.
Measured gains concentrated in language-heavy tasks
The clearest improvements from generative AI appeared in tasks that demanded synthesis, organization and clear written expression. Students with AI assistance produced higher-quality responses on text-based assignments and on quantitative problems that required estimation or structured reasoning. By contrast, gains were small or negligible for the creativity task and for exercises that required translating visual chart information into concise prose.
Actual use lagged behind access
Despite direct access and explicit encouragement to use the tool, many students did not rely on the AI consistently. Only 29 percent employed the generative AI across all four assigned exercises, while 16 percent never used it at all. Use varied by task: roughly three quarters tapped the AI for text comprehension but fewer than half consulted it for diagram-based questions.
User background and skill profiles matter
The study found clear selection effects: students with prior experience using language models were more likely to integrate the tool into their workflow. Individual ability profiles also influenced adoption and impact, suggesting that the productivity effects of generative AI depend as much on user competence as on model capability. These patterns indicate that unequal uptake could amplify or mitigate existing performance gaps depending on who uses the technology and how.
Implications for labor markets and organizations
Researchers caution against equating the existence of powerful models with immediate, economy-wide productivity surges or mass job displacement. Historical evidence and recent economic literature show that technology requires organizational change, institutional adaptation and time to generate broad gains. Generative AI may reduce costs and change job content, but the net effect on employment will depend on complementary investments, new demand and how firms reorganize work processes.
Education and skills as levers for adoption
The authors argue that unlocking AI’s potential will require deliberate skill-building, not just tool deployment. The LAICA project at Leuphana, funded by the Stiftung Innovation in der Hochschullehre, will pursue curricular strategies to teach students how to use AI effectively across disciplines. Building complementary competencies in higher education could narrow the gap between what models can do and what users actually do with them.
Equity concerns and preliminary gender patterns
Unequal adoption risks reinforcing existing disparities if certain groups use AI more effectively than others. The study’s findings echo broader research, including a meta-analysis from Harvard Business School, which reported lower AI usage among women in professional settings. However, initial evidence from the experiment suggests that generative AI can help weaker performers more, potentially reducing some performance inequalities when properly targeted.
Generative AI is reshaping how tasks are completed, but this study underscores that access alone does not determine impact. Adoption patterns, skill development and institutional change will determine whether generative AI becomes a broad productivity engine or a tool that benefits a subset of users and tasks.
