How University Students Really Use Claude: Insights from Anthropic’s Landmark Education Report

Redacción HC
10/09/2025

Generative AI is rapidly becoming a fixture in higher education, yet most of the debate so far has been driven by surveys and small-scale experiments. Anthropic’s Education Report: How University Students Use Claude changes that by offering one of the first large-scale analyses of actual student–AI interactions, drawn from anonymized conversations on Claude.ai linked to university email accounts. This dataset provides a rare window into not only what students ask AI to do, but also how they integrate it into academic work. The findings are already shaping discussions among educators, policymakers, and institutional leaders about both the risks—academic integrity, cognitive outsourcing—and the opportunities—enhanced learning support, content creation—of AI in the classroom.

From Hypothesis to Hard Data

Most prior studies on AI in education have relied on self-reported data, which can be incomplete or biased. Anthropic set out to observe real usage “in the wild,” processing over 1 million conversations and narrowing them to 574,740 that were clearly academic in nature. The analysis relied on Clio, an internal classification tool designed to categorize interactions without exposing private details. This multilayered pipeline tagged conversations by subject area, mapped interaction styles, and linked cognitive tasks to Bloom’s taxonomy.

Methodologically, the report also compared usage patterns across disciplines against U.S. National Center for Education Statistics (NCES) graduation data. This highlighted which fields were adopting AI more aggressively—and where uptake lagged. Importantly, the study acknowledges its own limits: early-adopter bias, focus only on Claude (excluding ChatGPT, Gemini, etc.), 18-day data retention, and no direct measurement of learning outcomes.

What Students Do with Claude

The single largest category—39.3%—was creating and improving educational content: drafting questions, editing essays, summarizing material. A close second, 33.5%, involved technical problem-solving: debugging code, working through algorithms, tackling math problems. Other uses included data analysis and visualization (11.0%), research design support (6.5%), technical diagramming (3.2%), and translation or proofreading (2.4%).

These numbers challenge the perception of AI as a tool for quick fact lookups. Instead, many students appear to use it for complex, multi-step tasks that combine domain knowledge with generative capabilities.

Disciplinary Divide: STEM Dominance

The adoption gap across disciplines was stark. Computer Science accounted for 38.6% of all academic conversations—far above its 5.4% share of U.S. degrees awarded. Natural Sciences and Mathematics also punched above their enrollment weight. Meanwhile, Business, Health, and Humanities were underrepresented, suggesting slower adoption outside STEM.

These disparities could reflect both the immediate utility of AI for programming and quantitative tasks and a lack of tailored integration strategies in non-STEM fields.

Four Interaction Styles—And Why They Matter

Anthropic mapped conversations into a Direct vs. Collaborative and Problem-Solving vs. Output Creation matrix. All four styles were common, each comprising 23–29% of interactions. However, almost half (≈47%) were direct—students asking for ready-made answers or outputs. While some of these uses are legitimate (e.g., generating study guides), others raise red flags: requesting answers to multiple-choice exams or rewriting text to evade plagiarism detection.

Discipline-specific patterns emerged too. In Education, 74.4% of conversations focused on output creation, hinting that instructors as well as students might be active users—particularly for lesson plan and material development.

Cognitive Load: Bloom’s “Inverted Triangle”

Claude’s most frequent cognitive functions were Create (39.8%) and Analyze (30.2%), with lower shares for Apply (10.9%), Understand (10.0%), and Remember (1.8%). This “inverted triangle” suggests students may be outsourcing higher-order thinking more than foundational skills. Depending on instructional design, this could either enrich learning—through co-creation and iterative feedback—or erode fundamental competencies if not balanced with scaffolded practice.

Implications for Policy and Practice

For Institutions: The prevalence of direct solution requests underscores the need for clear AI use policies that distinguish between acceptable support (feedback, brainstorming) and academic misconduct. Assessment strategies such as oral exams, portfolios, and process-based rubrics could mitigate overreliance on AI-generated answers.

For Policymakers: The heavy skew toward STEM fields indicates the need for AI literacy initiatives that are discipline-specific and inclusive of the humanities and social sciences. Equity concerns are also paramount—students without access or training risk being left behind.

For Educators: Designing assignments that require critical engagement with AI outputs, combined with explicit instruction on responsible use, can help transform AI from a shortcut into a cognitive partner.

The Road Ahead

Anthropic’s report doesn’t prescribe a single policy solution, but it points to promising directions: discipline-specific integration strategies, further research on learning outcomes, redesign of assessments, and university–AI provider partnerships for “learning modes” that foster deep understanding.

As AI’s role in higher education grows, the question may no longer be whether to allow it, but how to ensure it supports—not supplants—the intellectual growth universities are meant to cultivate.

Topics of interest

Academia Technology