Before an AI data trial, check content rights and exit terms

An AI team arranging a data trial needs to know which agreement covers its records, which activities that agreement addresses and what happens when access ends. SnowRock examined twelve public first-party documents to make those questions easier to ask.

Under a strict wording codebook, two documents make an express training decision: Reddit’s terms require additional permission, and the NYT developer terms prohibit training through that public route. Ten documents have no individual training decision under this codebook. Broader licenses, linked terms and other restrictions still apply. N in the CSV means no express activity-specific decision found in the inspected document. It does not mean permission was refused or granted.

Wikimedia Enterprise’s inspected agreement covers trial API access, the service an app uses to fetch records. It places content under separate licenses. Common Crawl warns that collected pages can carry their owners’ terms. Its AI-training passage assigns responsibility for claims rather than granting training permission.

OpenAlex explains a broad CC0 data grant, with a separate MAG-format exception. OpenStreetMap explains ODbL reuse and attribution. Neither page separately names every AI activity. A buyer must examine the general grant and the actual use, rather than treating missing AI-specific wording as a ban.

Exit terms can change the cost of a trial. Reddit directs deletion of stored content and derived data or models when access ends. News API distinguishes ordinary cancellation from provider termination: access and refunds follow different rules. Ask what happens to licensed records, model artifacts and already-committed payments before work begins.

Recipient questions also need context. World Bank Microdata rules vary by access category and restrict onward sharing on several routes. Common Voice combines a CC0 description with a request to use Mozilla Data Collective for distribution. The CSV preserves those statements without turning them into invented contractor or affiliate clauses.

For a proposed trial, ask the supplier to identify the agreement covering the records, the permitted model activities, every outside recipient, the sample decision and the exit obligations. Separate the license charge, preparation, delivery, testing and any agreed sourcing fee. Record who approves the next purchase and what that decision requires.

Method and limits: this is a declared convenience sample, not a survey of the data market. It contains six access agreements, one participation-terms document, one access policy, two dataset cards and two license explanations. Two AI agents independently coded identical saved copies before comparison. All 72 categorical cells matched. One price-description difference and three practical scope refinements were reconciled against sources. No statistical reliability score or legal clearance is claimed. Dataset cards, public summaries, source-site rights and negotiated contracts are different evidence. Linked documents were not all recoded.

CSV codes: A = express activity allowance; C = separate permission or eligibility required; R = express activity prohibition under the inspected route; N = no express activity-specific decision found. Code meanings describe the text, not the full legal position.