How it works in a data room
AI features send document text to a model to produce summaries, answers or redaction suggestions. Some AI services retain inputs to train future models unless told otherwise. A training opt-out, ideally written into the data processing agreement, confirms that the provider does not train on customer content and that its model suppliers are bound the same way. Strong versions also limit how long prompts and outputs are retained for abuse monitoring and say whether staff can review them.
Why it matters in a deal
Deal documents contain trade secrets, unannounced plans and personal data. If they feed a shared model, fragments could in theory resurface for other users, and the seller may breach its own confidentiality and privacy obligations. Many corporate policies now require a written no-training commitment before any AI tool touches confidential material. A settings toggle is helpful, but the contract is what holds.
Example
Before switching on AI summaries, a seller’s general counsel asks the provider three things: whether customer content is used for training, which model suppliers see it and how long they keep it. The provider confirms no training in the processing agreement, names its model host as a subprocessor and states a short retention period. The feature is approved for the room.