Privacy means protecting information people share with an LLM system from unnecessary collection or exposure. A response can be accurate yet unreliable if the system leaks sensitive data. Exposure can be lasting: information that has escaped cannot simply be taken back. section 11.5 section 1.4.3
Privacy protection must be proactive. A safety layer that catches a problematic response after generation does not prevent sensitive information from entering the system. LLMs add a further risk: they can memorize, recombine, and reproduce information in unexpected contexts later. section 11.5
Build privacy into the data pipeline and operating practices:
- Minimize collection and storage; do not retain data you do not need. section 11.6.6
- Add personally identifiable information (PII) detection and redaction to pipelines. section 11.6.6
- Determine applicable regulatory requirements early, give users control of their data, and explain what is collected and why. section 11.6.6
Sensitive inputs to an external service
In early 2023, Samsung employees submitted source code, internal presentations, and confidential business strategies to ChatGPT while seeking help with code reviews and meeting summaries. The service’s default settings at the time allowed user inputs to be used for model improvement. Samsung subsequently banned its use companywide. The case illustrates why privacy controls must address what enters a service, not only what its responses reveal. section 11.5
