Private AI Chat: A Consumer's Guide

You probably wouldn't say it out loud at a dinner party. You typed it into ChatGPT instead. So where does it actually go?

Where your AI chats actually go

The honest version is structural, not dramatic. Your chat isn't being actively read by a stranger right now. But it's sitting somewhere, and that "somewhere" creates real risk. Three things to understand.

1. They get stored

On consumer products like ChatGPT, Claude.ai, and Gemini, your chats sit on the company's servers attached to your account, often indefinitely. On Free and Pro plans, they're also used by default to train the next generation of the model unless you've opted out in settings. The chat is data, and the data has a long life.

2. They can be subpoenaed

Your chat history is a record. Records can be subpoenaed. In 2024, OpenAI was ordered by a federal court to preserve every ChatGPT conversation indefinitely as part of the New York Times copyright lawsuit. Including deleted ones. Including the ones in temporary chat. The order has not been lifted at the time of this writing.

3. They can be breached

If a hacker gets your password, your phone number, or your laptop, they get the chat history. If the AI company itself gets breached (it's happened to OpenAI in a 2023 incident), your chats are part of the breach.

The full list of what's flagged

Every major AI provider (Anthropic, OpenAI, Google) runs automated safety classifiers on every prompt. This is universal across the industry. The good news: the categories they flag are narrow, specific, and published. Here's the complete list from Anthropic's Usage Policy (other providers' lists are nearly identical):

• Child sexual abuse material (CSAM) and any sexual content involving minors
• Non-consensual sexual content depicting real people
• Weapons of mass destruction: chemical, biological, radiological, nuclear
• Manufacturing instructions for explosives or conventional weapons
• Cyberweapons, malware creation, instructions for exploiting software vulnerabilities
• Attacks on critical infrastructure (power grids, water systems, financial systems, healthcare systems)
• Incitement to violence, terrorism, or genocide
• Human trafficking, exploitation, and forced labor
• Fraud, scams, and large-scale deception schemes
• Doxxing, stalking, and unauthorized surveillance of individuals
• Election manipulation and coordinated political disinformation
• Discrimination in high-stakes decisions (employment, housing, credit, healthcare access)
• Unauthorized impersonation of real people
• Generating defamatory content about identifiable individuals
• Unauthorized practice of regulated professions (medicine, law, financial advice) without disclaimers

The incognito illusion

Both OpenAI and Anthropic have shipped "private" or "incognito" or "temporary" chat modes. They are real features. They are also genuinely misunderstood.

What incognito modes do: hide the chat from your visible history, opt the chat out of training, and remove cross-chat memory.

What they don't do: stop the AI company from receiving the prompt, encrypt the message end-to-end, or prevent retention during the company's own retention window. See the full breakdown of incognito modes, including which ones actually clear the data and how fast.

If your threat model is "I don't want this chat showing up next time my partner uses my laptop," incognito is fine. If your threat model is "I don't want this conversation discoverable in a lawsuit two years from now," incognito does nothing.

Three real paths to private AI chat

If you actually want the chat to be private, not theatrically private, you have three real options. Each has tradeoffs. Pick the one that matches your situation.

Path 1: Anonymous web tools (cheapest, weakest)

Sites like DuckDuckGo AI Chat, GPTAnon, AnonChatGPT, and Uncensored Chat let you talk to AI without an account. They proxy your message to an underlying model (GPT-4o mini, Claude, Mistral, etc.) and strip identifying info before forwarding.

Best for: one-off questions you don't want associated with your identity. Quick stuff.

Tradeoffs: you have to trust whoever runs the proxy. DuckDuckGo is reputable. The others vary. You generally don't get logged-in features (memory, history, premium models). Model quality is often capped at a free tier.

Path 2: Connection-password chat (best for daily use)

Here's how this works. You make an account directly with Anthropic at console.anthropic.com (it's free to sign up). They give you a connection password, which is a long secret that lets apps talk to Claude on your behalf. You paste that password into a privacy-respecting chat interface like Private Claude. The interface gives you a normal chat experience, but the conversation goes through the developer version of Claude with its stronger privacy rules, not through your Claude.ai account.

For Anthropic specifically: developer-version messages are never used for training, operational logs auto-delete after 7 days, and there's no saved chat history. Claude.ai is the opposite: training on by default, indefinite chat history, full account record.

Best for: people who use AI daily, want frontier models, and want their conversations to not exist as a permanent record anywhere.

Tradeoffs: there's a free tier (50 Haiku + 25 Sonnet messages, no card) for trying it out. Paid tiers are unlimited messaging (you pay only for your own Anthropic usage on top, about $3 to $5 a month if you're not using it every day). Private Claude Basic is $17 a month. Most people don't cancel their ChatGPT or Claude.ai subscription. They add Private Claude on top for the conversations they don't want sitting in their main account. You also don't get chat history on any tier, by design. Close the tab and the conversation is gone.

Path 3: Self-hosted open-source AI (strongest privacy, lower model quality)

Tools like Ollama, Jan, GPT4All, and LM Studio let you run an open-source model on your own computer. The data never leaves your machine. Period. There is no provider.

Best for: people with strong privacy threat models (journalists, activists, lawyers, security researchers), people who already have a decent GPU, or anyone who wants full data isolation as a principle.

Tradeoffs: the open-source models you can run on a laptop are meaningfully less capable than Claude or GPT-4. They're getting better. But for now, if you need the smartest answer, you're paying a real intelligence tax.

Side-by-side

Which is right for you

Most people don't need the strongest possible privacy. They need privacy that matches the kind of stuff they actually ask AI about. Match the tool to the question.

• Daily use, sensitive personal stuff: Private Claude or similar. Real Claude, stronger privacy rules, no chat history saved. Read the Private Claude post.
• One-off anonymous question: DuckDuckGo AI Chat. Reputable proxy, no account, throwaway chat.
• Maximum privacy, willing to pay in model quality: Ollama with Llama 3.3 or Mistral. Local, free, no provider involved.
• You don't actually care, you just want the best model: ChatGPT or Claude Pro. Just opt out of training in settings.

The question isn't "what's the most private tool." The question is "what's the right tool for the kind of stuff I actually ask." Most people end up with two: a daily driver that respects privacy, and an anonymous backup for the occasional question they don't want anywhere near their account.