Appendices

A. Top education tasks by volume

The 10 highest-volume education tasks, ranked by their share of total conversation volume (onet_task_pct). These 10 tasks account for approximately 58% of all education conversation volume.

Rank	Task	Subsector	Volume share (%)	Success %
1	Assist students who need extra help with their coursework outside of class	Postsecondary	2.73	74.5
2	Review class material with students by discussing text, working solutions to problems, or reviewing worksheets or other assignments	Other Teachers	1.47	81.3
3	Provide private instruction to individual or small groups of students to improve academic performance, improve occupational skills, or prepare for academic or occupational tests	Other Teachers	1.03	82.8
4	Develop instructional materials to be used by educators and instructors	Edu Support	1.00	65.5
5	Develop instructional materials, such as lesson plans, handouts, or examinations	Edu Support	0.79	68.3
6	Prepare and deliver lectures to undergraduate or graduate students on topics such as how to speak and write a foreign language	Postsecondary	0.55	91.4
7	Instruct through lectures, discussions, and demonstrations in one or more subjects, such as English, mathematics, or social studies	K-12 Teachers	0.39	87.9
8	Conduct classes, workshops, and demonstrations to teach principles, techniques, or methods in subjects such as basic English language skills, life skills, and workforce entry skills	Other Teachers	0.36	90.1
9	Develop teaching or training materials, such as handouts, study materials, or quizzes	Other Teachers	0.32	66.5
10	Evaluate and grade students' class work, assignments, and papers	Postsecondary	0.26	78.4

The full list of 266 education tasks with all metrics is available in the source data (see Appendix B).

B. Data access

The underlying data for this report comes from two sources:

• Anthropic Economic Index V4. The full dataset is available on HuggingFace: Anthropic/anthropic-economic-index. No API key is required. The dataset includes both the Claude.ai and API conversation samples.
• This analysis. The derived CSV files, analysis notebook, and this report are available on GitHub: trgallagher-research/AlignED-research-report-3.

C. Methodology details

SOC code filtering

Education tasks were identified by filtering the V4 dataset to rows where the SOC code begins with "25-" (Major Group 25: Educational Instruction and Library Occupations). This yielded 266 unique tasks after deduplication.

Renormalisation formula

For each task, the five substantive collaboration patterns (directive, task iteration, learning, validation, feedback loop) were renormalised:

This excludes the not_classified and none categories, making patterns comparable across tasks. The trade-off is that reported percentages overstate the share of conversations where a substantive pattern was identified.

Volume weighting

Aggregate statistics use volume weighting:

where onet_task_pct is each task's share of total conversation volume. This ensures that high-traffic tasks contribute proportionally to aggregate figures.

Subsector assignment

Tasks were assigned to subsectors based on their 4-digit SOC prefix: 25-1xxx (Postsecondary), 25-2xxx (K-12 Teachers), 25-3xxx (Other Teachers), 25-4xxx (Library), 25-9xxx (Education Support). Tasks mapping to multiple SOC codes were assigned to the first match after alphabetical sorting of task names.

D. Verification summary

An automated data verification was conducted on an earlier draft of this analysis using Claude and Gemini to check numerical accuracy, methodological consistency, and interpretive claims. The verification identified three critical issues and five important issues. Key findings:

Critical issues (corrected)

• Baseline numbers incorrect. The global collaboration pattern baseline figures contained errors due to inconsistent handling of the not_classified category. Baselines have been recomputed with explicit renormalisation.
• Sub-sector task counts inaccurate. Initial estimates of tasks per subsector were significantly off because 64 of 266 tasks map to multiple SOC codes. A consistent first-match deduplication has been applied and documented.
• Request-level facets not acknowledged. The V4 dataset contains 11 request-level facets that were not used in this analysis. This report uses only the onet_task-level data and states this scope explicitly.

Important issues (addressed)

• V2-to-V4 comparison dropped. Different classification models across versions make temporal comparison uninterpretable. This analysis has been excluded entirely.
• Weighting methodology clarified. The report now documents the volume-weighting approach and the renormalisation formula used.
• SOC 25 scope acknowledged. Education-related tasks outside SOC Major Group 25 (e.g., school counsellors in SOC 21) are not captured. This is stated as a limitation.

E. Citation

Data citation

F. Contact

For questions about this report, contact Tim Gallagher via GitHub.

This report is part of the AlignED Reports series.