Anthropic published an analysis on August 26, 2025, combining an automated review of roughly 74,000 anonymized Claude.ai conversations tied to higher-education email addresses (May 22–June 2, 2025) with qualitative interviews of 22 Northeastern University faculty. The report maps how university educators use Claude across teaching, research and administrative activities, showing frequent curriculum development, research support and growing use of custom interactive tools, while administrative tasks are more likely to be automated.
Key findings
- Most common uses in the analyzed conversations: curriculum development (57% of flagged conversations), academic research (13%) and assessing student performance (7%).
- Educators are creating deployable resources via Claude Artifacts: chemistry simulations, automatic grading rubrics, data dashboards, HTML quizzes and other interactive teaching materials.
- There is a clear augmentation-versus-automation split: tasks close to direct student interaction and creative work are typically augmentative, whereas routine administrative and financial tasks tend toward automation.
Concrete patterns and percentages
Anthropic matched conversations to educator tasks using the O*NET occupational taxonomy and then measured whether interactions were augmentation-heavy (AI collaborating with the user) or automation-heavy (AI directly performing the task). Notable figures from the report:
- Tasks with higher augmentation rates: university teaching and instructional material creation (77.4% augmentation); grant writing (70.0%); academic advising and mentoring (67.5%); supervising student academic work (66.9%).
- Tasks with higher automation rates: managing institutional finances and fundraising (65.0% automation); maintaining student records and evaluating academic performance (48.9% automation); admissions and enrollment management (44.7% automation).
Although grading and assessment accounted for only 7% of the studied Claude.ai conversations, when grading-related interactions appeared, 48.9% were classified as automation-heavy, indicating that nearly half of those interactions involved the AI directly performing evaluation-related tasks.
Why educators use AI
Interviews with Northeastern faculty highlight three common motivations:
- Automating tedious work ("it takes care of the rote portions");
- Using AI as a collaborative thought partner that suggests alternative explanations or approaches;
- Enabling more personalized, interactive learning experiences for students than a single instructor can provide.
At the same time, many faculty expressed ethical and practical concerns about automating assessment: some worry about accuracy and fairness, and several stressed that students pay for an instructor's expertise and time, not the LLM's.
Examples of educator-built tools
Faculty have used Claude Artifacts and related workflows to produce a range of teaching resources, including:
- Interactive educational games and simulations (escape rooms, platform-style learning games);
- Assessment tools with automatic feedback, CSV processors for performance analysis, and grading rubrics;
- Data visualizations to illustrate timelines or scientific concepts;
- Subject-specific learning modules (chemistry stoichiometry games, genetics quizzes with automated feedback, computational physics models);
- Academic calendars, scheduling tools and budget-planning templates;
- Standard academic documents (meeting minutes, recommendation letters, grant drafts, email templates).
These creations illustrate a shift from using AI purely as a conversational assistant toward treating it as a creative collaborator that can lower technical and time barriers to producing interactive educational materials.
Rethinking teaching and assessment
Faculty report that AI is changing what and how they teach: some are redesigning assignments that can’t be trivially completed by AI, others are prioritizing foundational skills so students can judge AI outputs critically. One faculty member noted that AI-based coding tools allow more class time to focus on conceptual discussions instead of syntax-level debugging.
Assessment practices are evolving as well. While dishonesty and cognitive offloading remain concerns, some instructors have replaced traditional research papers or reworked weekly homework to make assignments that are harder for general-purpose LLMs to complete.
Limitations and caveats
Anthropic notes several important limitations of the study:
- Identification methodology: the filter used to infer educator conversations captured only about 1.5% of conversations associated with higher-education emails and therefore focused on explicitly educator-linked tasks, likely missing other educator uses;
- Sector scope: analysis is limited to higher education and excludes K–12 teachers;
- Early-adopter bias: the sample likely overrepresents educators already comfortable with AI;
- Sample size for qualitative data: insights from 22 Northeastern faculty may not generalize broadly;
- Platform specificity: findings reflect Claude.ai usage and may not apply to other AI platforms;
- Temporal window: the data cover a short period (May 22–June 2, 2025) and may not reflect seasonal or longer-term patterns.
Conclusion
The report presents a nuanced picture: university educators are experimenting with Claude across a wide range of tasks, from curriculum design and research support to administrative automation. Notably, many are building concrete, interactive teaching tools that would previously have required substantial time or technical skill. Yet the prevalence of automation in grading-related conversations and faculty reservations about assessment automation highlight unresolved questions about educational quality, fairness and ethics. Future research should continue to track how educator and student AI use interact and how institutional policies and best practices evolve.
Notes
- The Gallup survey referenced in the report found that teachers reported saving an average of 5.9 hours per week using AI tools; Anthropic cites this figure for broader context.
- The study’s analyzed conversations occurred during an 11-day period from May 22 to June 2, 2025.
- Authors listed in the Anthropic post: Drew Bent, Kunal Handa, Esin Durmus, Alex Tamkin, Miles McCain, Stuart Ritchie, Ryan Donegan, Jennifer Martinez and Jason Jones (Anthropic, 2025-08-26).
