Research Data
Datasets we publish in open research repositories. Each has a permanent DOI and a CC BY 4.0 licence.
Research papers
A Functional Taxonomy for AI Governance: 18 Categories, 165 Subcategories and the Training Policies of 11,890 AI Tools
How the AI tool market breaks down by function, and what vendors state about training on customer data: 18 categories and 165 subcategories applied to 11,890 active AI tools, with the stated training policies of 9,853 tools.
Same Tools, Different Rules: Generative AI Acceptable-Use Policies Across Fifteen Sectors, with a Complete Policy Matrix for K-12 Schools
Fifteen sector acceptable-use profiles, from K-12 schools to defense, applied to the same 11,890 AI tools: the share of tools blocked ranges from 10.0% to 81.6%, and two sectors disagree on 41.3% of tools on average. Includes the complete K-12 policy matrix.
Where Over-Blocking Comes From: Simulating CIPA School Filtering Policies on 103 Million Categorised Domains
Five school filtering policies applied to 103,367,510 categorised domains: the content CIPA names blocks 0.92% of domains and 0.048% of education, health and civic sites, while a restrictive 29-category district policy blocks 19.1% and 7.6% of them. Over-blocking comes from what districts add, not from what the law requires.
Datasets
Collateral blocking under CIPA: aggregate tables from a census of 103 million web domains
The CIPA over-blocking tables, archived at ICPSR, the social science data archive of the University of Michigan, with a full study description.
K-12 Web Filtering Policy Outcomes: Exact Block and Over-Block Counts for 103 Million Domains
Exact counts of blocked domains and of blocked education, health and civic domains for five school filtering policies and three multi-label rules, by domain class and link popularity, with a cost-ordered blocking curve. The data behind the CIPA over-blocking paper above.
K-12 Web Filtering Policy Outcomes: Exact Block and Over-Block Counts for 103 Million Domains (OSF project)
The same tables as an open OSF project, with a project wiki that explains the law, the five policies, the results and every file.
Category Co-occurrence and Collateral Blocking in School Web Filtering: Exact Counts for Five CIPA Policies Across 103 Million Domains
How the 57 web filtering categories overlap across 103 million domains, and what that overlap means for five school filtering policies: shared domains, conditional shares, lift and Jaccard index for every category pair.
CIPA School Web Filtering: Policy Simulations over 103 Million Categorised Domains
The aggregate tables of the CIPA over-blocking study: policy simulations, results by domain class and popularity, category prevalence and the full 57 by 57 co-occurrence matrix.
AI Tools Taxonomy and Training-Policy Statistics (version 1.3.0)
The current version of the taxonomy: 11,890 AI tools, every one reviewed individually, with stated training-policy statistics for 9,853 tools. This is the data behind the SSRN paper above.
AI Tools Taxonomy and Statistics
Version 1.2.0: a functional taxonomy of 11,959 AI tools in 18 categories and 165 subcategories, a 200-tool sample with risk ratings, and aggregate statistics on the stated training policies of AI tools.
AI tool classification snapshot, September 2026
The same functional classification of AI tools with counts per category and subcategory, category profiles and stated training policies, with its own sample of 200 example tools.
Generative AI tool categories, a 200-tool stratified sample, and vendor training-data disclosures
Snapshot of 26 September 2026: 11,959 AI tools in 18 functional categories and 165 subcategories, a 200-tool stratified sample with risk and data-sovereignty ratings, and how often vendors state a policy on training with customer inputs.
AI Agent Page-Type Statistics: Login, Pricing, Documentation and 17 Other Page Types across 40.8 Million Domains with URL-Rule Results
How often login, pricing, documentation and 17 other page types appear across 40.8 million domains, with the results of URL rules for detecting each page type.
Also open, on our product sites: Shadow AI Sector Policy Statistics and the Web Filtering Research Dataset.