# AI Privacy The set of concerns around what happens to your data when you use AI platforms. Every prompt you send, every file you upload, and every conversation you have with a cloud AI service is data that goes somewhere. ## The core tension Using AI effectively requires giving it context: your documents, your code, your ideas, your business data. But sending that context to a cloud provider means trusting them with it. The more context you provide, the better the output; the more you share, the greater the exposure. ## Key risks - **[[AI Training Data Collection]]**: your prompts and responses may be used to train future models, effectively making your data part of the model's knowledge - **Data retention**: providers may store your conversations for varying periods - **IP leakage**: proprietary code, business strategy, or trade secrets sent to AI become data you no longer fully control - **Employee exposure**: staff using consumer AI tools may inadvertently share confidential information - **Third-party access**: data may be accessible to provider employees, subcontractors, or through legal requests ## Mitigation strategies - Use **API access** instead of consumer chat interfaces (APIs typically don't train on your data) - Use **enterprise plans** with explicit data handling agreements - Run models locally with [[Running AI Models Locally|local inference]] ([[Ollama]], [[LM Studio]]) - Use [[AI Open Weight Models]] to keep everything on your infrastructure - Review and configure opt-out settings on every platform - Establish clear AI usage policies for teams and organizations ## The tradeoff spectrum | Approach | Privacy | Capability | Cost | |----------|---------|------------|------| | Consumer chat (free tier) | Low | High | Free | | API access | Medium | High | Per-token | | Enterprise plan | High | High | Subscription | | Local [[Small Language Models (SLMs)|SLMs]] | Maximum | Limited | Hardware | | Local large models | Maximum | Good | Expensive hardware | There's no free lunch. Maximum privacy with maximum capability requires significant hardware investment. Most people land somewhere in the middle: API access for sensitive work, consumer tools for general use. ## References - ## Related - [[OpenAI Privacy Filter]] — open-weight PII detection model; pre-prompt sanitization layer - [[AI Training Data Collection]] - [[Running AI Models Locally]] - [[AI Open Weight Models]] - [[AI Safety]] - [[AI Governance]] - [[Responsible AI]] - [[Small Language Models (SLMs)]] - [[Ollama]] - [[LM Studio]] - [[On-Device Machine Learning]] - [[Browser-Provided Language Models]]