Privacy means governing information that can identify a person throughout its use in an AI web application. Personally identifiable information (PII) includes names, addresses, phone numbers, and email addresses. A handling plan should address user consent, secure storage, and requests to delete data—not just access to the application. section 1.6.1
Chat histories need particular care: they may contain PII, financial details, or medical information. Before storing them or sending their contents to an LLM provider, assess their sensitivity and the provider’s privacy policy. Avoid putting PII in prompts or logs, especially when the provider does not share your PII policies. section 9.5 section 9.6.2
Turn that assessment into controls across the data lifecycle:
- Encrypt sensitive chat histories, and protect sensitive data both at rest and in transit. section 9.5 section 9.6.2
- Restrict access to people who need it. If someone must review another user’s history through impersonation, log the action and keep its scope and duration limited. section 9.5
- Set retention periods against legal and business needs; delete expired histories securely and permanently, and give users a way to download them. section 9.5
- Anonymize or pseudonymize histories where possible, particularly before using them for model training or analysis. Audit access to stored histories. section 9.5
Make privacy a production-deployment check alongside security, cost, and latency. Check the applicable rules—including GDPR, HIPAA, or CCPA as relevant—rather than assuming that a locally working application is ready to process users’ data publicly. section 9.1 section 9.6.2
