How can I tell whether AI training data was used with permission?
Look for evidence about the specific model: named datasets, licenses or agreements, public-domain sources, creator controls, and a way to request removal or challenge use. Broad source categories or an open model download do not reveal a complete rights record.
What does a useful training disclosure show?
Look for source lists, dataset versions, licenses or permission statements, web exclusions, and a process for rights requests. Three catalog profiles show different levels of detail: OLMo 3 7B Instruct publishes stage-by-stage data and training recipes but leaves source-level permission checks incomplete; Apertus 1.5 8B publishes scripts and data artifacts with opt-out safeguards, while source-level rights clearance remains unverified; Pleias-Pico documents the base model sources and removal process, but not the exact later training dataset version.
A page being public does not tell you whether it was included or under what permission. Open weights describe how a model can be run; they do not provide a complete rights record. Compare compare models and read the source list on each model profile.
Is training-data permission the same as chat privacy?
No. Training-data provenance concerns the material used to build a model. Chat privacy concerns how a service handles what you type now: storage, human access, model-improvement use, and third-party routing. A chatbot can have a useful prompt opt-out while its model’s training sources remain partly undisclosed; a well-documented training data does not tell you how the chat service stores prompts. See private AI chat controls and the compare providers comparison.
What can creators check or do?
Read the developer’s current documentation for source descriptions, licensing claims, opt-out or takedown procedures, and whether the mechanism applies to past training or future crawls. Keep a copy of the policy date and the model version you checked. If the disclosure names only broad categories, record that as a limit rather than treating it as proof of permission or misuse.