Anthropic Education Report: How university students use Claude

educationstudentsusage dataClioBloom's taxonomyacademic integrity

Handa, Bent and colleagues use Clio on about one million Claude.ai conversations from higher-education accounts, 574,740 of them judged academic, and find Computer Science heavily overrepresented relative to U.S. degrees, four interaction styles (direct or collaborative, problem solving or output creation) at similar rates, about 47% of conversations direct, and Claude mostly performing the higher-order Bloom's Taxonomy functions of creating and analyzing.

Kunal Handa, Drew Bent, Alex Tamkin et al. (Anthropic) · 2025-04-08 · original

The post observes that, until then, most discussion of how students use AI had relied on surveys and controlled experiments, which record what students say or what they do in a study setting. This post, the first Anthropic Education Report, looks instead at what students actually ask an AI system, using a large sample of real Claude.ai conversations from university accounts. It asks which subjects use Claude most relative to enrollment, how students interact with it, and which kinds of thinking they hand to it. It was written by Kunal Handa and Drew Bent with twelve co-authors. All numbers below are the post's; there is no accompanying paper.

What the report does

The sample

The authors used Clio, Anthropic's system in which Claude summarizes conversations into aggregate patterns without researchers reading them, on about one million anonymized conversations from Claude.ai Free and Pro accounts with higher-education email addresses, such as .edu and .ac.uk domains, over an 18-day period set by data-retention policy (section "Identifying educational AI usage", footnotes 1 and 2). Because such accounts also belong to staff, Clio then filtered the conversations to those relevant to students' academic work, leaving 574,740.

Two classification schemes

Interaction styles. Conversations are classified on two axes (section "How students interact with AI"):

  • Mode: Direct, where the user wants to resolve the query as quickly as possible, or Collaborative, where the user engages in dialogue with the model to reach the goal.
  • Desired outcome: Problem Solving, seeking solutions or explanations, or Output Creation, producing longer outputs such as essays or presentations.

The four combinations are Direct Problem Solving, Direct Output Creation, Collaborative Problem Solving and Collaborative Output Creation. In the classification itself the authors used the labels "Transactional" and "Dialogic," which they found best described the patterns Clio surfaced, and renamed them for the report (footnote 5).

Bloom's Taxonomy. To ask which cognitive work is delegated, the authors classify Claude's responses using Anderson and Krathwohl's 2001 revision of Bloom's Taxonomy, an educational hierarchy of cognitive processes from Remembering through Understanding, Applying, Analyzing and Evaluating to Creating (section "Cognitive tasks students delegate to AI", footnote 6). The framework was designed for student thinking; the authors adapt it to the AI's side of the conversation.

What it finds

What students ask for

The most common requests are creating and improving educational content across disciplines (39.3% of conversations), such as practice questions, essay editing and summaries, and technical explanations or solutions for assignments (33.5%), such as debugging code, implementing algorithms and solving mathematics problems. Smaller shares go to analyzing and visualizing data (11.0%), research design and tool development (6.5%), technical diagrams (3.2%) and translation or proofreading (2.4%) (section "What are students using AI for?"). The post notes that some of this use "might also be cheating."

Subjects relative to enrollment

Comparing each subject's share of conversations with its share of U.S. bachelor's degrees, using National Center for Education Statistics (NCES) data (section "AI usage across academic disciplines"), the post reports two overrepresented subjects and three underrepresented ones. Computer Science accounted for 38.6% of conversations against 5.4% of degrees, and Natural Sciences and Mathematics for 15.2% against 9.2%. Business accounted for 8.9% of conversations against 18.6% of degrees, Health Professions for 5.5% against 13.1%, and Humanities for 6.4% against 12.5%. The post's chart covers the other subjects; the figures here are the ones its text states.

The Computer Science share is given as 38.6% in this section and as 36.8% in the post's opening list of key findings. The authors suggest STEM students may be earlier adopters, possibly because of greater awareness of Claude in computer science communities and greater AI proficiency at STEM tasks.

Interaction styles

The four styles each account for between 23% and 29% of conversations, and about 47% of conversations are Direct. By subject (section "Subject-specific AI usage patterns"):

  • Natural Sciences and Mathematics lean toward Problem Solving.
  • Computer Science, Engineering, and Natural Sciences and Mathematics lean toward Collaborative conversations; Humanities, Business and Health are more evenly split.
  • Education shows the strongest preference for Output Creation, 74.4% of its conversations, and makes up 3.8% of all conversations. Many involve lesson plans and teaching materials, which the authors take to mean teachers are in the sample, a possible imperfection of the filter.

Academic integrity

The post says how much students cheat is hard to answer, because the same pattern can serve different purposes: a Direct Problem Solving conversation could be cheating on a take-home exam or checking a practice test. Among Direct conversations the authors nevertheless found concerning examples, such as providing answers to machine learning multiple-choice questions and rewriting business texts to avoid plagiarism detection. They add that Collaborative conversations can also offload thinking, as with homework solved step by step over several turns.

Bloom's Taxonomy: an inverted pattern

Claude's responses mostly perform higher-order functions: Creating (39.8%) and Analyzing (30.2%), against Applying (10.9%), Understanding (10.0%) and Remembering (1.8%) (section "Cognitive tasks students delegate to AI"). The text gives no share for Evaluating. Output Creation conversations involve more Creating and Problem Solving conversations more Analyzing. The authors note that the AI exhibiting a skill does not stop the student from also using it, but that the pattern points to concerns about students outsourcing cognitive work and failing to build foundational skills: "An inverted pyramid, after all, can topple over."

The prompts and classifications used

The post prints no full prompts, but two pieces of wording define its sample and categories:

  • The academic filter. Clio kept conversations "likely to be a student seeking help with academics, school work, studying, learning a new concept, academic research, etc?" (footnote 3), a filter the post says was validated in the earlier Clio paper.
  • Interaction labels. Clio received a description of each of the four interaction patterns under the Transactional and Dialogic labels (footnote 5). The descriptions themselves are not published.

The report also mentions a product direction: as part of partnerships with universities, Anthropic was experimenting with a Learning Mode that "emphasizes the Socratic method and conceptual understanding over direct answers" (section "Anthropic's approach to education"). The report gives no results for it.

Limitations and open questions

The authors list these (section "Limitations"):

  • The data probably capture early adopters, and Claude.ai is only one of the AI tools students use.
  • Classification has false positives (staff or faculty classed as students) and false negatives (students on personal email accounts).
  • Only one 18-day window is analyzed, and use likely varies over the academic year.
  • The study sees what students delegate, not how they use the outputs or whether learning improves.
  • Subject categories may not capture interdisciplinary work.
  • Applying Bloom's Taxonomy to an AI's responses is imperfect; Remembering in particular is hard to quantify for an AI.
  • Institutional AI policies vary and may shape the patterns in ways the data cannot show.

The report presents no human validation figures of its own for the interaction-style or Bloom's classifications.

Research questions

  1. How reliable are the interaction-style and Bloom's classifications? The report gives no validation figures. Having human raters label a sample of conversations under the same definitions, and measuring agreement with Clio, would establish their accuracy.
  2. Which Direct conversations involve academic misconduct? The post cannot tell cheating from legitimate use. Pairing usage patterns with course policies and assignment types, with consent, would show how often Direct use breaks the rules that apply.
  3. Does the inverted Bloom's pattern affect learning outcomes? The authors raise the worry that delegating creating and analyzing could weaken foundational skills. A longitudinal study linking usage patterns to assessed performance would test it.
  4. Does a Socratic Learning Mode change interaction styles? The report mentions the experiment without results. Comparing interaction styles and Bloom's distributions for the same students with and without the mode would show its effect.
  5. How does use vary across the academic year? The single 18-day window is a stated limitation. Repeating the analysis at several points in a term would show whether proportions shift near exams.
  6. Why do Business, Health and Humanities students use Claude less? The authors suggest awareness and AI proficiency. Surveys of students in those fields, combined with task-level success rates, would separate the two explanations.

How it connects

References

  • Handa, K., Bent, D., Tamkin, A., McCain, M., Durmus, E., Stern, M., Schiraldi, M., Huang, S., Ritchie, S., Syverud, S., Jagadish, K., Vo, M., Bell, M., and Ganguli, D. (2025, April 8). Anthropic Education Report: How University Students Use Claude. Anthropic. https://www.anthropic.com/news/anthropic-education-report-how-university-students-use-claude
  • Anderson, L. W., and Krathwohl, D. R. (Eds.). (2001). A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom's Taxonomy of Educational Objectives. Cited as the post uses it; not read for this note.