How do I keep ChatGPT, Gemini, Copilot, and Claude conversations private?

Start with the honest privacy model

Hosted AI assistants process your prompt somewhere you do not control. Privacy controls are still worth using, but they are not magic: they change retention, training, memory, or workspace handling; they do not turn a cloud chatbot into an offline encrypted vault.

The practical rule is to separate everyday brainstorming from sensitive material. If a prompt contains passwords, private keys, unreleased product strategy, confidential customer data, health records, legal facts, financial records, or proprietary code, assume it needs a stricter workflow than a consumer chat window.

ChatGPT: use data controls and Temporary Chat

OpenAI documents a Data Controls setting called Improve the model for everyone. Turning it off means conversations remain in history but are not used to train ChatGPT. OpenAI also documents Temporary Chats: they are not used to train models, are not saved in history, do not create memories, and are deleted after 30 days, though they may be reviewed for abuse monitoring.

Use Temporary Chat for one-off questions you do not want saved into history or memory. Use the model-improvement opt-out as the default if you do not want ordinary chats used for training. Still avoid secrets: OpenAI's privacy policy warns that no Internet or email transmission is fully secure.

Gemini: check Activity, connected apps, and browser context

Google's Gemini Privacy Hub is explicit that Gemini can collect what you say, upload, share, or ask it to do, and can also receive context from connected apps, browsers, devices, remote-browser sessions, cookies, screen captures, and other page content depending on the feature you use. That is useful for assistant workflows, but it is also a privacy boundary.

Before using Gemini with personal or work data, review Gemini Apps Activity, connected apps, browser or Chrome context, location, audio/Live, and any feature that can see screen content or act in another app. The privacy decision is not just Gemini or not; it is which data surfaces you connect.

Claude: use Incognito and model-improvement controls

Anthropic says consumer Claude chats and coding sessions may be used for model improvement when you choose to allow it, when content is flagged for safety review, or when you explicitly opt in. It also says Incognito chats are not used to improve Claude even if Model Improvement is enabled.

For Claude Code and connector-heavy workflows, pay attention to what is copied into the conversation. Anthropic states raw connector content, including remote and local MCP server content, is not included in training data unless it is directly copied into the conversation. That still means an agent can surface sensitive material into chat if your tool permissions are too broad.

Copilot: distinguish work-protected Copilot from consumer chat

Microsoft's strongest public privacy boundary in the checked sources is for Microsoft 365 Copilot Chat. Microsoft says prompts and responses are processed inside the Microsoft 365 service boundary and are not used to train underlying foundation models. It also logs prompts and responses for auditing/eDiscovery in the Microsoft 365 tenant.

That does not mean every Copilot-branded surface has identical handling. For work data, use the tenant-managed Microsoft 365 Copilot experience with enterprise data protection rather than pasting business content into an unmanaged personal assistant. Check your organization's policy and the exact Copilot surface before assuming a protection applies.

A practical checklist before you paste

Use this sequence when the answer matters: first decide whether the data belongs in a hosted assistant at all; then choose the most protected account or workspace available; then turn off training/model-improvement settings where available; then use temporary or incognito mode for sensitive one-off prompts; then strip identifiers, secrets, customer names, and proprietary context that the model does not need.

If the task is work-related, ask whether your organization has an approved AI workspace. If not, use synthetic examples or local/offline tooling instead of pasting real private material into a consumer chat.

• Never paste passwords, API keys, private keys, seed phrases, customer records, unreleased financials, private medical/legal facts, or confidential source code unless your approved policy explicitly allows it.
• Disable or review memory/personalization if you do not want the assistant to reuse facts across chats.
• Review connected apps, browser context, file uploads, screen sharing, remote computer/browser features, and MCP tools separately; each one expands what the assistant can see.
• Use enterprise/work plans for company data and confirm retention, training, audit, and subprocessor terms before relying on them.

A useful next action

Open each AI assistant you use today and find its data controls. Turn off model-improvement training if that is your preference, confirm whether memory is on, and learn how to start a temporary or incognito chat. Then write a simple personal rule: what data is never allowed in consumer AI chat. The rule matters more than the brand.