Anthropic Education Report: How University Students Use Claude — Key Findings
Anthropic has published an Education Report analyzing approximately one million anonymized student conversations on Claude.ai, one of the first large-scale studies of how students actually use AI in real academic settings rather than in surveys or controlled experiments. The headline result: STEM students, and Computer Science students in particular, are the earliest and heaviest adopters, and students most often delegate higher-order cognitive work — Creating and Analyzing — to the model.
Why this report matters
Most prior discussion of AI in education has relied on surveys and controlled experiments rather than direct evidence of how students naturally integrate AI into academic work. Anthropic frames its Education Report as filling that gap by studying real usage at scale.
"AI systems are no longer just specialized research tools: they're everyday academic companions."
The report's stated purpose is to inform questions about learning, assessment, and skill development as AI becomes more deeply embedded in educational environments.
Methodology: Clio and the dataset
The study used Claude Insights and Observations ("Clio"), Anthropic's automated analysis tool that provides insights into how people use Claude. Clio works bottom-up, distilling user conversations into high-level usage summaries such as "troubleshoot code" or "explain economic concepts."
Clio uses a multi-layered, automated process that removes private user information from conversations. Anthropic states that the process was built to minimize the information that passes from one layer to the next, and describes Clio's privacy-first design in an earlier blog post.
The analysis pipeline was:
- Approximately one million anonymized conversations from Claude.ai Free and Pro accounts tied to higher education email addresses.
- Filtering for student and academic relevance — for example, whether a conversation pertained to coursework or academic research — which yielded 574,740 conversations.
- Grouping by Clio to derive aggregate education-related insights on subject representation, student–AI interaction styles, and delegated cognitive tasks.
What students use Claude for
Students primarily use Claude to create and improve educational content across disciplines (39.3% of conversations) — designing practice questions, editing essays, or summarizing academic material. The second-largest category is technical explanations or solutions for academic assignments (33.5%), including debugging and fixing errors in coding assignments, implementing programming algorithms and data structures, and explaining or solving mathematical problems.
The remaining usage breaks down as follows:
| Use case | Share of conversations |
|---|---|
| Creating and improving educational content | 39.3% |
| Technical explanations or solutions for assignments | 33.5% |
| Analyzing and visualizing data | 11.0% |
| Supporting research design and tool development | 6.5% |
| Creating technical diagrams | 3.2% |
| Translating or proofreading content between languages | 2.4% |
Anthropic notes that some of this usage might also be cheating, which the report discusses separately.
Discipline-level adoption: Computer Science dominates
STEM students are early adopters of AI tools like Claude. Computer Science students are particularly overrepresented: they account for 36.8% of students' conversations while comprising only 5.4% of U.S. degrees, according to the report's key findings. In the discipline-level analysis section, the report states that Computer Science accounted for 38.6% of conversations on Claude.ai.
To measure disproportionate use, Anthropic compared Claude.ai usage patterns against the number of U.S. bachelor's degrees awarded, using National Center for Education Statistics (NCES) subject areas.
| Subject area | Share of Claude.ai conversations | Share of U.S. bachelor's degrees |
|---|---|---|
| Computer Science | 38.6% | 5.4% |
| Natural Sciences and Mathematics | 15.2% | 9.2% |
| Business | 8.9% | 18.6% |
| Health Professions | 5.5% | 13.1% |
| Humanities | 6.4% | 12.5% |
Business, Health, and Humanities students show lower adoption rates relative to their enrollment numbers. Anthropic suggests these patterns indicate that STEM students, particularly in Computer Science, may be earlier adopters of Claude for educational purposes, while students in Business, Health, and Humanities may be integrating these tools more slowly into their academic workflows. The report offers two possible explanations: higher awareness of Claude in Computer Science communities, and AI systems' greater proficiency at tasks performed by STEM students relative to those performed by students in other disciplines.
Four patterns of student–AI interaction
Anthropic identified four distinct patterns of student–AI interaction, categorized along two axes. The first axis is mode of interaction: Direct conversations, where the user wants to resolve a query as quickly as possible, versus Collaborative conversations, where the user actively seeks dialogue with the model. The second axis is desired outcome: Problem Solving, where the user seeks solutions or explanations, versus Output Creation, where the user seeks to produce longer outputs such as presentations or essays.
Combining the axes produces four patterns:
- Direct Problem Solving
- Direct Output Creation
- Collaborative Problem Solving
- Collaborative Output Creation
These four interaction styles were represented at similar rates — each between 23% and 29% of conversations — which the report says shows the range of uses students have for AI. Anthropic contrasts this with traditional web search, which typically only supports direct answers, and argues that AI systems enable a much wider variety of interactions and with them new educational opportunities.
Selected positive learning examples cited in the report include:
- Explain and clarify philosophical concepts and theories
- Create comprehensive chemistry educational resources and study materials
- Explain muscle anatomy, physiology, and function concepts for academic assignments
The cheating question
Anthropic states directly that it is hard to answer how much students use AI to cheat, because the specific educational context of each Claude response is unknown. The report gives paired examples to illustrate the ambiguity: a Direct Problem Solving conversation could be cheating on a take-home exam, or a student checking their work on a practice test; a Direct Output Creation conversation could be creating an essay from scratch, or creating summaries of knowledge for additional research. Whether a Collaborative conversation constitutes cheating may also depend on specific course policies.
Nearly half (~47%) of student–AI conversations were Direct — seeking answers or content with minimal engagement. Anthropic notes that many of these serve legitimate learning purposes, such as asking conceptual questions or generating study guides, but the report also found concerning Direct conversation examples:
- Provide answers to machine learning multiple-choice questions
- Provide direct answers to English language test questions
- Rewrite marketing and business texts to avoid plagiarism detection
The report adds that even Collaborative conversations can have questionable learning outcomes. It cites the example "solve probability and statistics homework problems with explanations," which may involve multiple conversational turns between AI and student but still offloads significant thinking to the AI.
Subject-specific interaction patterns
Interaction styles differ by discipline, which the report says suggests that educational approaches to AI integration would likely benefit from being discipline-specific.
- Natural Sciences & Mathematics conversations tended toward Problem Solving, such as "solve specific probability problems with step-by-step calculations" and "solve academic homework or exam problems with step-by-step explanations."
- Computer Science, Engineering, and Natural Sciences & Mathematics leaned toward Collaborative conversations, whereas Humanities, Business, and Health were more evenly split between Collaborative and Direct conversations.
- Education showed the strongest preference for Output Creation, covering 74.4% of conversations. Education made up 3.8% of all conversations. Anthropic cautions that this usage might stem from imperfections in its filtering methods: many of these conversations involved "creat[ing] comprehensive teaching materials and educational resources" and "creat[ing] detailed lesson plans," indicating that teachers are also using Claude for educational support.
Cognitive tasks students delegate to AI
Anthropic adapted Bloom's Taxonomy — a hierarchical framework used in education to classify cognitive processes from simpler to more complex — to analyze Claude's responses when conversing with a student, rather than the student's own thinking.
The report found an inverted pattern of Bloom's Taxonomy domains exhibited by the AI:
| Bloom's Taxonomy domain | Share of operations |
|---|---|
| Creating | 39.8% |
| Analyzing | 30.2% |
| Applying | 10.9% |
| Understanding | 10.0% |
| Remembering | 1.8% |
This distribution varied by interaction style. Output Creation tasks, such as generating summaries of academic text or feedback on essays, involved more Creating functions. Problem Solving tasks, such as solving calculus problems or explaining programming fundamentals, involved more Analyzing functions.
Anthropic notes that AI systems exhibiting these skills does not preclude students from also engaging in the skills themselves — for example, co-creating a project together or using AI-generated code to analyze a dataset in another context. But the report says this points to potential concerns about students outsourcing cognitive abilities to AI, including legitimate worries that AI systems may provide a crutch that stifles the development of foundational skills needed to support higher-order thinking.
"An inverted pyramid, after all, can topple over."
Limitations of the study
Anthropic lists several limitations that affect the scope of the findings:
- The dataset likely captures early adopters and might not represent the broader student population.
- It is unclear how representative Claude use is relative to overall AI usage in education; many students use AI tools beyond Claude.ai, so the report presents only a partial view of their overall AI engagement patterns.
- There are likely both false positives and false negatives in how conversations were classified. The analysis relied on accounts tied to higher education email addresses, so some conversations classified as student-related may actually be from staff or faculty, and other student conversations are likely on accounts tied to non-university email addresses.
- Due to privacy considerations, Anthropic only analyzed Claude.ai usage within a single 18-day retention window. Students' usage likely differs across the year as their educational commitments fluctuate.
- The study only examines what tasks students delegate to AI, not how they ultimately use AI outputs in their academic work or whether these conversations effectively support learning outcomes.
- The categorization of student–AI conversations into academic disciplines may not fully capture interdisciplinary work, where AI usage patterns may differ significantly.
- Applying Bloom's Taxonomy to the cognitive processes of an AI, as opposed to a student, is imperfect. Skills like Remembering are harder to quantify in the context of AI systems.
Anthropic also notes that institutional policies regarding AI use in education vary widely and might significantly impact the observed patterns in ways that cannot be measured within the dataset.
Anthropic's approach to education and what comes next
Anthropic describes the report as a bird's-eye view of where and how students are using AI in the real world, and states that it is only at the beginning of understanding AI's impact on education. The report cites examples of AI empowering learning, including supporting a student's nuclear fusion reactor project and facilitating better communication between students and teachers in classrooms.
At the same time, Anthropic acknowledges that AI is making educators' lives more challenging in ways the research does not fully capture. The report raises open questions: how to ensure students still develop foundational cognitive and meta-cognitive skills; how to redefine assessment and cheating policies in an AI-enabled world; what meaningful learning looks like when AI systems can near-instantly generate polished essays or rapidly solve complex problems that would take a person many hours; and whether homework design and assessment methods will fundamentally shift as model capabilities grow.
Beyond the report, Anthropic says it is partnering with universities to better understand the role of AI in education. As an early step, the company is experimenting with a Learning Mode that emphasizes the Socratic method and conceptual understanding over direct answers, and it plans to collaborate with universities on future research studies and on directly studying the effects that AI has on learning.