A new Anthropic study offers a detailed look at how university users are bringing Claude into academic work. The central finding is not simply that students are using AI. It is that many are turning to it for the kinds of tasks that sit near the top of the learning process: creating, analyzing, solving, and generating polished work.
The study raises a practical question for higher education. If AI can help students learn, but can also complete the work that learning is meant to build, institutions need clearer ways to separate support from substitution.
What Anthropic Studied
Anthropic’s research team began with one million conversations from users with university email addresses. After removing material outside the academic scope, the researchers analyzed 574,740 academic chats across an 18-day window.
The sample showed a strong tilt toward STEM users. Computer science students represented 38.6 percent of all users in the data, while making up only 5.4 percent of the U.S. student population. That imbalance matters because it suggests the study may capture the behavior of early and heavy adopters more than the full student population.
Even with that caveat, the dataset gives universities a useful signal. It shows not just that AI is present in academic routines, but that its use is already distributed across different types of student work.
Four Ways Students Use Claude
Anthropic grouped the conversations into four broad interaction patterns. Students either used Claude directly or collaboratively, and those interactions were aimed either at solving problems or generating content.
Each category accounted for between 23% and 29% of conversations. The striking detail is that nearly half, 47%, were direct, minimal-effort interactions. In those cases, students appeared to give Claude the task with limited back-and-forth.
That pattern is important because it points to a difference between using AI as a tutor and using it as a replacement for effort. A student who asks for step-by-step help may still be trying to understand the process. A student who asks for a finished answer may be removing the very work the assignment was designed to require.
The examples Anthropic highlighted show why the line is hard to police. Some students used Claude for multiple-choice questions on machine learning. Others asked it to produce direct answers for English tests, or to rewrite marketing and business assignments in ways meant to avoid plagiarism detectors.
At the same time, not every academic use is automatically suspect. The source notes that using AI to check practice work can be a legitimate form of self-study, while using it to complete graded homework is a different case. Without the assignment context, intent, and course rules, the same-looking AI interaction can mean very different things.
Higher-Level Thinking Is Being Handed Off
The most consequential part of the study is what students are asking AI to do. Using Bloom’s Taxonomy, Anthropic found that 39.8% of prompts fell into the "Create" category and 30.2% fell into "Analyze." Simpler categories were less common: "Apply" accounted for 10.9%, and "Understand" accounted for 10%.
That distribution matters because creating and analyzing are not minor academic chores. They are often the point of the assignment. When a language model drafts, restructures, reasons through, or generates the final shape of a response, it can take on a large part of the cognitive load.
This does not mean every such use blocks learning. A student might learn from comparing Claude’s reasoning with their own, or from asking for an explanation after attempting a problem. But the study suggests that many interactions may not be structured that way. Direct, low-effort use makes it harder to know whether the student is building skill or simply obtaining output.
The patterns also vary by discipline. Science and math students were more likely to use AI for problem-solving, including guided explanations. Educators used Claude mainly to create teaching materials and lesson plans, with 74.4% of use cases in that area focused on creation.
What Universities Still Do Not Know
Anthropic’s study comes with limits. The research covered a short time frame, and the sample may be shaped by early adopters. The classification process could also assign some interactions incorrectly. For example, chats from university staff may have been included alongside student activity.
Those limits do not erase the findings, but they should shape how universities read them. The study is a strong snapshot, not a complete map of student behavior. It shows where concern is warranted, while also making clear that more research is needed.
The most difficult institutional problem is context. A conversation log can show what Claude was asked to do, but not whether the work was for practice, tutoring, lesson planning, a graded assignment, or an attempt to bypass academic rules. That missing layer is exactly where academic integrity decisions usually live.
Universities therefore face several connected challenges:
- Defining when AI assistance supports learning and when it replaces required work.
- Designing assignments that make student reasoning visible.
- Giving students practical guidance instead of relying only on broad warnings.
- Understanding how different disciplines use AI in different ways.
The Campus AI Race
The study arrives as AI companies are moving deeper into higher education. Anthropic has launched Claude for Education, a campus-focused product with special learning modes. Northeastern University, the London School of Economics, and Champlain College are already rolling it out, and Anthropic plans to connect it with platforms such as Canvas LMS.
OpenAI has also entered the higher education market with ChatGPT Edu, launched in May 2024. The offering gives universities discounted access to the latest models and includes capabilities such as data analysis and document summarization. Oxford, Wharton, and Columbia are already using it for tutoring, assessments, and administrative work.
The broader direction is clear from the source: AI providers want students and institutions to build these tools into everyday academic routines. If students become familiar with a platform during university, they may carry that habit into their professional lives after graduation.
That makes the education debate larger than a classroom policy issue. AI is becoming part of the infrastructure around learning, teaching, assessment, and administration. The urgent question is not whether students will use these tools. The question is how universities will set boundaries that preserve learning while acknowledging that AI is already embedded in academic work.