ModelRefs / AI Privacy — AI Glossary
AI Privacy — AI Glossary
The practices and controls that prevent LLM systems from leaking, memorizing, or exposing private user or organizational data. Privacy is not security.
Overview
AI privacy concerns: PII exposure in prompts, model memorization of training data, embedding inversion attacks, and vendor data retention. Mitigations: PII redaction, differential privacy in training, enterprise no-training contracts, on-premise model deployment.
Reference details
| Topic | safety |
|---|---|
| Last reviewed | 2026-06-24 |
Related terms
Example: Three surfaces, and a contract covers only one
Data reaches a model three ways, and they need different controls. It arrives in the prompt — mitigated by redaction before the call. It can be retained and trained on — mitigated by a zero-retention or no-training agreement, which is a contractual control, not a technical one. And it can already be inside the weights from pretraining, where no contract you sign now helps, because the memorisation already happened. Teams routinely sign the agreement and treat all three as solved. The third surface is why differential privacy at training time and output-side scanning both exist.
Commonly confused with
Privacy is not security. Security keeps out people who should not have access; privacy governs what happens to data from people who legitimately do. A perfectly secure system that trains on customer prompts has a privacy problem and no security problem. Confidentiality agreements sit on the privacy side and are enforced legally, not technically.
When to use it
Reach for it when:
- Any deployment handling personal, customer or regulated data
- Vendor assessment, where retention, training and residency terms are the substance
- Designing what gets logged: traces and evaluation datasets are copies of user data
Reach for something else when:
- Treating a no-training clause as covering memorisation from pretraining
- As a substitute for access control, retention limits and deletion paths
- Assuming on-premise deployment resolves it — your own logs are still a copy
Referenced by
This term is used by the following ModelRefs references:
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Privacy — AI Glossary.
Frequently asked questions
What is AI Privacy?
The practices and controls that prevent LLM systems from leaking, memorizing, or exposing private user or organizational data.
What concepts are related to AI Privacy?
Closely related concepts include guardrails, responsible ai, open weights.