The report opens by noting that much of the discussion of AI in education concerns how students use it. It then turns to the people who teach them. It follows an earlier report on how university students use Claude (April 2025) and asks what university educators use Claude for, which tasks they collaborate on and which they hand over, and how teaching is changing in response. The report is a blog-length piece in Anthropic's Societal Impacts series, combining a privacy-preserving analysis of conversations with survey and qualitative data from one university. It was designed and written by Drew Bent and Kunal Handa, with feedback from seven colleagues (Acknowledgements).
How the data were gathered
Conversation analysis
The analysis used Clio, Anthropic's privacy-preserving tool for studying usage patterns (Clio). Because Claude.ai does not record users' occupations, the authors started from Claude.ai Free and Pro accounts associated with higher-education email addresses and then applied a Claude-powered filter to keep only conversations about educator tasks (section "Identifying educators' use of Claude"). Footnote 2 prints the filter, which asks whether a conversation is "likely to be with an educator (teacher, professor, or instructor) seeking help with instructional content, grading, research, or administrative duties," excludes "students doing their own coursework," and instructs the classifier to "Err on the side of conservatism." The filter yielded about 74,000 conversations from 22 May to 2 June 2025, an 11-day period (footnote 1).
Each conversation was matched to the most appropriate task in O*NET, the US Department of Labor's database of occupations and their tasks, restricted to tasks associated with postsecondary teaching or administrative occupations. Similar tasks were grouped into higher-level categories.
Augmentation and automation
Following the Anthropic Economic Index, each conversation was classified by how the work was divided. Augmentation means the AI collaborates with the user on a task; automation means the AI directly performs it. Conversations whose mode could not be identified were excluded from the chart (note to the augmentation chart).
Faculty survey
The authors partnered with Northeastern University and gathered survey data and qualitative responses from 22 faculty members described as early adopters of AI.
What it finds
Most common uses
In the conversation data the most common use was developing curricula (57% of conversations analyzed), followed by conducting academic research (13%) and assessing student performance (7%) (section "Common uses among educators"). Curriculum development was also the leading use reported by the Northeastern faculty. Those faculty also said they spent on average 29% of their AI time on their own learning, a use the conversation analysis did not study because the filter could not separate educator learning from student learning. Other uses named in the data include mock legal scenarios for teaching, vocational and workforce training content, recommendation letters, and meeting agendas.
Faculty gave three kinds of reason for these uses: automating tedious tasks, using AI as a thought partner, and building personalized learning experiences for students.
Building tools with Artifacts
Educators used Claude's Artifacts feature to build working resources rather than only holding conversations (section "How educators are building custom tools with AI"). The report lists interactive games such as escape rooms and simulations; quizzes with automatic feedback and rubric tools; data visualizations; subject-specific tools such as chemistry stoichiometry games and computational physics models; academic calendars; budget tools; and academic documents such as meeting minutes and recommendation letters. One faculty member wrote that what had been "prohibitively expensive (time) to do [before] now becomes possible."
Which tasks are delegated
The report gives augmentation and automation shares for task categories (section "The augmentation-automation spectrum"):
| Task category | Share | Mode |
|---|---|---|
| University teaching and classroom instruction | 77.4% | augmentation |
| Writing grant proposals | 70.0% | augmentation |
| Academic advising and student organization mentorship | 67.5% | augmentation |
| Supervising student academic work | 66.9% | augmentation |
| Managing institution finances and fundraising | 65.0% | automation |
| Maintaining student records and evaluating academic performance | 48.9% | automation |
| Managing academic admissions and enrollment | 44.7% | automation |
The authors read the pattern as educators delegating routine administrative and financial work while collaborating on tasks that need context, creativity or direct contact with students, and say this matches the survey.
Grading
The category that includes grading was the second most automation-heavy, with 48.9% of grading-related conversations classified as automation, although it made up only 7% of conversations studied. Surveyed faculty rated grading as the task where AI was least effective. The report calls the automation share "concerning" and notes that it is not clear how far AI-generated grades or feedback enter final grades. A quoted professor objected to AI grading on both accuracy and ethical grounds, and the authors conclude that most educators "seem to agree" grading should not be close to fully automated.
Changes to teaching
Quoted faculty described changing how and what they teach: redesigning assignments that AI can complete, spending less time on debugging syntax and more on concepts in an analytics course that now uses AI-based coding, and teaching students to evaluate AI output. One professor said they "will never again assign a traditional research paper." The authors note that raising the difficulty of assignments is a moving target as AI improves and can burden educators, and that students still need foundational skills learned without AI.
Limitations
The report lists its own caveats (section "Limitations and considerations"):
- Narrow filter. The filter kept only about 1.5% of conversations from higher-education addresses, so it captures tasks explicitly tied to educators and likely misses many others, such as help explaining a concept.
- Higher education only. K-12 teachers are excluded.
- Early adopters. Users in the data may be more comfortable with AI than educators generally.
- One institution. The Northeastern sample is small and may not generalize.
- One platform. Patterns may differ on other AI products.
- Short window. The May to June period misses seasonal variation over the academic year.
The report presents its analysis as "an exploration" of profession-specific uses rather than a comprehensive view. It also gives percentages only for the categories named in the text; the full set of task categories appears in a chart.
Research questions
- What does grading automation look like in detail? The report cannot tell how AI-generated grades or feedback are used. Linking a sample of grading conversations to instructor-reported use (final grade, draft feedback, formative comments) would show what the 48.9% automation share means in practice.
- How sensitive are the results to the educator filter? The filter was conservative and kept about 1.5% of conversations. Re-running the analysis with a looser filter, or validating it against accounts with known roles, would show how much the task mix depends on it.
- Do the delegation patterns hold for K-12 teachers and other countries? The data cover higher education. Repeating the analysis with appropriate account identification would test generality.
- How do educator and student use interact? The report raises this itself. Matched studies of courses where instructors use AI, measuring student awareness and their own AI use, would address it.
- Do Artifact-built tools improve learning? The report describes tools educators built but not their effect. Classroom studies comparing outcomes with and without such tools would measure it.
- Is the gap between survey opinion and observed grading use stable over time? Repeating both measures in later terms would show whether practice moves toward or away from faculty views.
How it connects
- Companion report. Anthropic Education Report: How university students use Claude (April 2025) is the student-side study the report says it follows.
- Method. The analysis uses Clio and the augmentation and automation categories of The Anthropic Economic Index.
- Cited by the September Economic Index report. Anthropic Economic Index report: Uneven geographic and enterprise AI adoption links to this report when explaining a rise in tasks about developing instructional materials.
- Later education work. Anthropic Education Report: The AI Fluency Index (February 2026) continues the education report series.
References
- Bent, D., Handa, K., Durmus, E., Tamkin, A., McCain, M., Ritchie, S., Donegan, R., Martinez, J., and Jones, J. (2025, August 27). Anthropic Education Report: How educators use Claude. Anthropic. https://www.anthropic.com/research/anthropic-education-report-how-educators-use-claude
- The Gallup survey (5.9 hours saved per week) and the Economic Index paper (arXiv:2503.04761) are cited in the report and were not read for this note.