How we read the twelve documents

Question and sample

Which AI activities and practical access or exit questions do twelve accessible first-party documents expressly address?

This is a declared convenience sample, chosen for accessible, relevant public text. It is not representative of the data market. The final sample contains six access agreements, one participation-terms document, one access policy, two dataset cards and two license explanations. Document types remain separate in the CSV.

We excluded private negotiated contracts, inaccessible pages, generic website terms and documents visibly scheduled for a future effective date. We recorded one inspected document per named offer. Referenced licenses and linked terms were not all separately coded.

Reading and reconciliation

Two AI agents read identical saved copies independently using the same codebook, then compared their coding. Both agents assigned the same label in all 72 activity fields. One practical price-description difference and three scope refinements were resolved against sources. The reconciler was also the first coder. This was not an independent third adjudication.

Matching labels are not a statistical reliability score. Many cells use N because the inspected text does not separately name an activity. General licenses and linked restrictions may still govern that use.

What the four codes mean

A

The inspected text expressly allows this specific activity. Ordinary general legal obligations may still apply.

C

This specific activity requires a separate agreement, written authorization, named approval or stated additional eligibility.

R

The inspected text expressly forbids this specific activity under the stated access route.

N

No explicit permission decision for this specific activity was identified in the inspected document. General grants, intended-use descriptions or risk clauses are recorded separately. This does not mean allowed, forbidden or absent everywhere.

What each row records

Each row identifies the publisher, document, type, scope, source URL, collection time, saved-source SHA-256 hash and stated version. The hash identifies the saved copy we read; it does not promise that today’s page remains identical.

Six fields record express decisions about training, fine-tuning, evaluation, retrieval, contractor access and affiliate access. Notes cover updates, ending access, obligations afterward, price wording and the sections inspected. Training means changing model parameters. Retrieval means looking up records to answer a request, rather than training on them.

Changes to the initial sample

World Bank Microdata and News API replaced two initially proposed sources that failed the access or relevance checks. The inspected OpenAlex source was its maintained license explanation. The CSV lists the final twelve documents, rather than the initial proposal.

Reproduce and use the findings

Download the CSV and follow each source link. Compare the exact document scope and version before using its notes. Public pages, dataset cards, underlying content rights and negotiated agreements are different evidence. Ask a supplier to resolve questions for the actual records and intended use.

The report contains source-linked summaries, not copies of the full third-party texts. Source changes after October 26, 2026 are outside this edition. We made no market-wide estimate, provider ranking, legal clearance or claim of measured search traffic.