Model training on your data
Also: training on your data, does it train on your data
Model training on your data is whether an AI vendor uses the content you submit, meaning prompts, uploads and generated outputs, to train or fine-tune its models for other users, as opposed to keeping it isolated to your own session.
This is the single most-asked question about any AI tool, and the answer is rarely a flat yes or no. Vendors describe their posture along a spectrum: training on everything by default with no opt-out; training by default with an opt-out available; not training unless you explicitly opt in; or not training at all on a given tier. The same product can sit in different places on that spectrum depending on which plan you are on, which is why a tool's overall brand reputation is a poor substitute for checking the specific tier.
The common tier-based pattern
The most common pattern among major assistants is 'tier-based': the free or consumer plan trains on conversations by default (sometimes with human review as part of that process), while the paid business, enterprise or API tier is excluded from training entirely as part of the commercial terms. ChatGPT is a clear example of this split: Business, Enterprise and API usage is excluded by default, while the free and Plus consumer product trains unless you turn that off yourself.
'Not trained on' is not 'not retained'
'Not trained on' and 'not retained at all' are also different claims worth separating. A vendor can exclude your data from model training while still retaining it for a period for abuse monitoring, debugging or legal holds. If retention length matters for your use case, check the data-retention specifics, not just the training answer.
Treat the answer as dated
Because this fact changes as vendors update their terms, and because it depends on plan tier, treat any answer as time-stamped. A sourced, dated fact from the vendor's own trust center or enterprise-privacy page beats a general impression, however confident that impression feels. The AI Tool Risk Directory states each tool's sourced training posture with a link to the vendor's own documentation, and the no-training compliance hub lists tools that do not train on your inputs.