This guide explains how AI providers use your data, why it matters, and the concrete tools and settings that let you use AI while keeping your data private.
Reading time: ~15 min
- Introduction
- How Your Data is Used by AI Providers
- Why It Matters
- Handling Your Data When Using AI
- Private Ways to Use LLMs
- Choosing Your Approach
AI tools rely on large amounts of data to function, and much of that data is yours: your prompts, the files you upload, and how you use the service. Privacy here is not about secrecy; it is about keeping control over your information and knowing how it is used.
That control matters for several reasons:
- Personal security: limiting where your data goes reduces exposure to identity theft, fraud, and leaks.
- Trust and transparency: you should be able to know how a service handles your data.
- Legal compliance: regulations such as the GDPR in Europe set requirements for how organizations process personal data.
- Ethics: respecting privacy is a basic principle that protects individual rights.
To go deeper, Neil Richards' book Why Privacy Matters is a thorough introduction. For practical starting points:
- Privacy Guides — approachable, for non-technical users.
- Awesome Privacy (pluja) and Awesome Privacy (lissy93) — large lists of privacy-focused tools.
- Personal Security Checklist — basic-to-advanced steps to protect your devices and accounts.
When you use a hosted AI service, your information typically moves through three stages:
- Processing: your input goes to the provider's servers to answer your request.
- Storage: it may be retained to monitor usage, debug, and improve the service.
- Reuse: depending on the provider and your settings, it may be used to train future models or shared with partners.
How user data can flow through an AI provider.
Providers use this data in two main ways. The first is personalization: your interactions help tailor responses to you. The second is model and product improvement: aggregated data from many users informs research, new features, and performance tuning. Whether your data is used for training depends on the provider, the tier you use, and your account settings — which is exactly what the rest of this guide helps you control.
Using a hosted AI service exposes your data to several risks, and you often have limited visibility into how it is handled:
- Hidden collection: services may log more than your direct inputs, including usage patterns and the edits you make to their suggestions.
- Breach risk: large stores of user data are attractive targets, and a single breach can expose many users at once.
- Profiling: behavioral data can be combined into a detailed profile of how and when you work.
- Re-identification: "anonymized" data can sometimes be linked back to an individual by combining data points.
- Weak legal guarantees: privacy law lags behind AI, so compliance today does not guarantee your data won't be used in unexpected ways tomorrow.
- Data as an asset: inputs can be treated as a commercial asset — shared, sold as insights, or used to build new products — often without clear disclosure.
A few habits go a long way:
- Read the privacy policy before you commit to a service. Check what is collected, how long it is retained, whether it is used for training, and how to opt out.
- Prefer transparent or local-first tools. Open-source and on-device options keep your data under your control; for cloud services, choose ones with clear, strong privacy commitments.
- Share only what's needed. Don't paste secrets, credentials, or personal details unless the task truly requires them, and separate personal from professional use.
- Use strong account security: HTTPS, up-to-date software, a password manager, and two-factor authentication.
- Review regularly. Re-check app permissions and privacy settings periodically, and revoke access you no longer need.
Beyond general habits, these concrete settings have the biggest impact on whether your data is retained or used for training:
- Turn off training in consumer apps. ChatGPT, Claude, Gemini, and Copilot all have a setting to exclude your conversations from model training. It is often on by default on free consumer tiers, so check it.
- Use temporary / incognito chats. Features like ChatGPT's Temporary Chat aren't saved to your history and aren't used for training — useful for sensitive one-off questions.
- Prefer the API or Team/Enterprise tiers. By default, the major providers (OpenAI, Anthropic, Google) do not train on data sent through their APIs, and their Team/Enterprise plans contractually exclude training. If you have the option, these are more private than free consumer apps.
- Remember that "no training" is not "no logging." Even when training is disabled, providers may retain data for a limited period to monitor abuse. For the strongest guarantees, run locally or use a provider built around zero-retention.
Running a model on your own hardware is the most private option: your data never leaves your computer, and you have full control over which model you use and how. The trade-off is that capable models need capable hardware, and there is some setup involved.
A step-by-step guide is in our dedicated tutorial: How to Run LLMs on Your Machine.
Duck.ai is a free, anonymous AI chat from DuckDuckGo. It requires no account, removes identifying metadata (such as your IP address) before forwarding your prompt, and stores recent chats locally on your device. DuckDuckGo has agreements with the underlying providers that prohibit them from using your prompts to train their models.
- Models: a selection of hosted models from Anthropic, Meta, Mistral, and OpenAI (including open-weight options), available for free.
- Notable: no sign-up, anonymized requests, local chat history, and a voice mode.
- Best for: quick, anonymous access to capable models with zero setup. See its privacy details.
Proton Lumo is a privacy-first assistant from Proton (the maker of Proton Mail). It keeps no logs of your prompts or replies, and saved conversations use zero-access encryption, so they can only be decrypted on your device. Lumo is fully open-source and never uses your data to train its models, which are open-weight models run on Proton-controlled servers.
- Models: open-weight models (such as GPT-OSS, Qwen, and Kimi K2) hosted by Proton.
- Notable: zero-access encrypted chat history, open-source code, encrypted "Projects" for grouping files and chats, and a free tier with a paid Lumo Plus upgrade.
- Best for: a confidential, account-based assistant from an established privacy company. See its security model.
The right choice depends on your priorities, hardware, and how much setup you'll accept:
- Maximum privacy and control → run locally. Your data never leaves your machine and you can use any compatible open-source model, at the cost of needing capable hardware and some setup. See the hardware guidance.
- Anonymous, zero-setup access → Duck.ai. Free and account-free, with capable hosted models and no training on your prompts. The trade-off is trusting DuckDuckGo's anonymization.
- A confidential, persistent assistant → Proton Lumo. Encrypted, open-source, and account-based, with synced history. The trade-off is trusting Proton's implementation.
- Hosted models in general → use the practical privacy levers above (API or Enterprise tiers, training turned off, temporary chats) to reduce what's retained.