대학생들은 AI챗봇을 어디에 쓰는가? (2025.4)

English summary

This video discusses Anthropic's April 2025 education report, How University Students Use Claude, to describe how university students use the Claude chatbot for academic activities. Using Clio, a privacy-preserving analysis system, the report examined approximately 547,740 university-related conversations selected from about one million higher-education conversations and described usage purposes, disciplinary distributions, interaction styles, and the delegation of cognitive tasks.

Key points

• The main purposes were creating and improving educational content at 39.3%, explanations and assignment problem-solving at 33.5%, data analysis and coding at 11.0%, research design and tool development at 6.5%, diagram and flowchart creation at 3.2%, and translation and language revision at 2.4%.
• Computer science was strongly over-represented, accounting for 38.6% of Claude use compared with about 5.4% of US bachelor's degrees. Natural sciences and mathematics were also over-represented at 13.2% compared with about 9.2% of degrees. Business, health professions, and humanities were under-represented relative to their degree distributions.
• Interactions were classified by combining direct or transactional versus collaborative or dialogic modes with problem-solving versus output-creation goals. The four types were broadly balanced overall, while direct problem-solving was relatively prominent in natural sciences and mathematics and collaborative problem-solving was relatively prominent in computer science. Education showed a high share of output creation. The supplied notes also provide some discipline-specific percentages, but do not clearly explain the denominator or calculation basis for every figure. Based on actual conversations, 39.8% of requests were classified as creating, 30.2% as analyzing, 15.9% as applying, 5.5% as evaluating, 3.8% as understanding, and 1.8% as remembering. Thus, the submitted tasks included more higher-order activities such as creating and analyzing than lower-order activities such as remembering and understanding. The discussed responses include a learning mode that uses Socratic questions, emphasis on core concepts, and structured templates or scaffolding rather than immediate complete answers, as well as human-AI collaborative assignments and critique of AI outputs. The video also notes that oral or interactive assessments may become more important for checking authentic understanding.

Audience

The content is intended for higher-education faculty, instructional designers, education policymakers, educational technology researchers, and university administrators.

Scope and limitations

The analysis concerns Claude users and may not represent university students' generative AI use as a whole. Automated classification through Clio may have misclassified user intent or academic domains, and filtering by institutional email addresses may have included staff, faculty, or researchers in addition to students. The data were limited to a single 18-day retention window and cannot distinguish the tasks submitted to the chatbot from the amount of learning or independent cognitive effort before, during, or after the interaction. Applying Bloom's Taxonomy, originally developed to describe human learning progression, to the delegation of tasks to AI also has conceptual limitations.

Research context

The content connects thematically with education and public health through its focus on how generative AI affects higher-education learning activities and assessment design. It is particularly relevant to interpreting student AI-use data and designing collaborative learning practices involving humans and AI.