Anthropic Research: How Educators Use AI and Claude Code's New Learning Mode
Summary
- • Anthropic analyzed ~74,000 educator conversations with Claude via Northeastern University partnership
- • 77% of classroom instruction uses are collaborative human-AI efforts, not full automation
- • 65% of financial and fundraising tasks are fully delegated to AI by faculty
- • Claude Code launched a learning mode that teaches coding through explanation and hands-on practice
Details
74,000 educator conversations analyzed in Anthropic-Northeastern University collaboration
The dataset covers professors using Claude across a range of academic tasks. The scale and institutional collaboration give the findings unusual empirical grounding compared to typical AI adoption surveys.
Faculty are using Claude as a builder tool — creating simulations, rubrics, and dashboards with Claude Artifacts
Examples include interactive chemistry simulations and data dashboards. This positions AI as a prototyping and development partner within academia, beyond simple Q&A or summarization.
77% of teaching and classroom instruction uses are collaborative, involving both human and AI input
This suggests faculty are not simply offloading work to AI but actively co-creating instructional content. It challenges the narrative that AI reduces human involvement in teaching.
65% of financial and fundraising tasks are fully delegated to AI by educators
Administrative and financial tasks see the highest full-delegation rates, suggesting faculty trust AI most for structured, low-stakes outputs where consequences of minor errors are manageable.
49% of grading conversations show automation patterns, yet faculty rate grading as AI's least effective application
This contradiction is the study's most notable tension. Professors appear to automate grading despite skepticism about quality — possibly driven by workload pressure, even when they doubt the output.
Claude Code learning mode launches with Lite and Learn-by-doing tracks
Lite mode explains best practices and trade-offs as users code. Learn-by-doing mode pauses at key moments and asks the user to write critical sections, flagged as TODO(human). The feature targets both CS students and experienced developers.
Claude Code's learning mode positions AI as a coding tutor rather than a code generator
By deliberately withholding completions and requiring user input at critical junctures, Anthropic differentiates on skill-building — a response to criticism that AI coding tools create dependency and erode developer competency over time.
Research = empirical study findings, Stat = specific quantitative data point, Insight = notable pattern or tension, Product Launch = new feature release, Strategy = business or product positioning
What This Means
Anthropic is building an evidence base for how AI is actually used in education — and the findings reveal a more nuanced picture than simple automation: faculty are collaborating with AI far more than they are delegating to it, except in administrative work. The grading paradox, where professors automate a task they distrust, points to a gap between AI capability and faculty confidence that could shape how institutions govern AI use in assessment. Meanwhile, Claude Code's learning mode signals a broader industry reckoning with whether AI tools should optimize for output speed or human skill development — a tradeoff with long-term implications for how the next generation of developers learns.
Sources
- Wrote down some reflections after 1 year at AnthropicThreadreaderapp