What is zero data retention (ZDR)?
Zero data retention means an AI provider processes your request and keeps nothing: no stored prompts, no stored outputs, no training on your text. What ZDR covers, and what it does not.
Updated June 11, 2026
The definition
Zero data retention, shortened to ZDR, is a processing guarantee from an AI provider: your request is processed, the response is returned, and nothing persists afterward. No prompt logs, no stored outputs, and no use of your content to train models. The data exists on the provider's machines only for the seconds the computation takes.
ZDR matters most when the text being processed is valuable before publication: an unpublished manuscript, a confidential script, a client's product copy. For that material, a provider's default logging is a quiet leak even when nobody ever reads the logs.
What ZDR is not
ZDR describes the model host in the processing path, not your account. A product can offer a saved library, history, and re-downloads (all of which are retention) while still requiring ZDR from the inference providers it calls. The first kind of retention is a feature you control and can delete; the second is a side effect you usually cannot see.
ZDR is also not anonymity. The provider still receives the text in order to process it; what it promises is to keep nothing afterward. Who receives your text in the first place is a separate, equally fair question.
How Cantari handles retention
Your library is retention you choose: scripts and generated audio are stored privately in your account, never published, never trained on, and deletable item by item or with the whole account. The named list of who processes your text on the way to an engine is published on the privacy page, in plain language.
Behind the scenes, small text-helper jobs (cue tagging, manuscript chapter detection) follow a hard routing rule: the budget model in that chain may only run on zero-data-retention endpoints, with rerouting to other hosts disabled. If that constraint cannot be met, the job moves to a different model entirely rather than loosening the constraint.
The questions worth asking any AI vendor: who processes my text, is it logged, is it trained on, and can I delete what is stored? If an answer is vague, treat it as a no.
Related reading
Retention is about what happens to your inputs; ownership is about your outputs. The second half is covered in ownership and rights and on the ownership page.