# What is Training Opt Out?

Also called Model Training Exclusion.

A training opt out is a setting or contractual term under which a provider agrees not to use a customer's inputs and outputs to train or improve its models. Defaults differ substantially by provider, product tier and region, so whether content is used for training is a question to verify in current terms rather than assume.

The concern behind the control is that training data can influence a model in ways that are difficult to reverse or inspect. Research has demonstrated memorization and extraction of training examples under some conditions, particularly for rare or repeated strings such as credentials and identifiers. Even where extraction is unlikely, customers in regulated sectors often cannot justify contributing confidential material to a third party's training corpus.

Business and consumer offerings commonly differ. Many providers state that content submitted through business, enterprise or developer channels is not used for model training by default, while consumer products may use content unless a user opts out. These positions have changed repeatedly as products and regulation evolve, which is why the reliable source is the provider's current documentation and the signed agreement.

An opt out is narrower than it sounds. It typically addresses training specifically and says nothing on its own about retention, human review for abuse monitoring, or use in aggregate analytics. A complete data posture pairs the training term with explicit retention scope, access controls on staff review, subprocessor disclosure and deletion commitments.

For the party deploying the AI system, the obligation runs downstream too. Users whose data is processed generally need transparency about whether their content reaches a model provider and under what terms. Purpose limitation applies with equal force to data the deploying organization holds, so reusing customer content to fine tune an internal model raises the same questions it raises for a vendor.

## Key points

- Controls whether inputs and outputs feed model training
- Defaults vary by provider, plan tier and region
- Training terms are separate from retention and human review
- Memorization research is why the control exists
- Reusing customer data for internal fine tuning raises the same issues

## In practice

Before rolling out an assistant to a support team, an operations lead checks the provider's current terms for their plan, confirms in writing that submitted content is excluded from model training, and records the reference in the vendor file. Separately they confirm the retention window and who at the provider can view content during abuse investigations, since the training term alone would not have answered either question.

## Related terms

- [Zero Data Retention](/en/glossary/zero-data-retention)
- [Purpose Limitation](/en/glossary/purpose-limitation)
- [Data Processing Agreement](/en/glossary/data-processing-agreement)
- [Model Provider Data Policy](/en/glossary/model-provider-data-policy)
- [GDPR](/en/glossary/gdpr)

[Back to the AI Glossary](/en/glossary)