Private AI chat: what actually stays private
"Private" AI chat can mean no saved history, no training, no provider retention or on-device inference. These are four separate controls. What each service promises, including ours.
Short answer
"Private" is not one setting, and no hosted chat can prove it stores nothing. A hosted AI chat must send your text to a server to generate a reply. Check four things separately: is history saved, who processes the prompt, is it used for training, and what metadata is kept. Only a local model keeps prompts on your device.
The four layers
| Layer | Question to ask | Example |
|---|---|---|
| History | Is the conversation saved to an account or server? | Temporary or private chats skip saved history |
| Processing | Which provider runs the model, and what does it keep? | OpenRouter shows each provider’s retention policy |
| Training | Can prompts be used to train models? | OpenRouter lets you exclude providers that train |
| Metadata | Are timestamps, token counts and billing logged? | Almost always yes |
What each service says (checked 8 Nov 2026; their claims, not audits)
- Unrestricted AI (us): signed-in chats are saved by default to our database (Neon) so you can reopen them on any device; they are not available through any admin interface. Private chat is a switch: its messages stay in this browser’s local storage and never reach our conversation database. Prompt-text collection for product review stopped on 8 Nov 2026; older records may remain in storage or backups until deleted. Venice AI, NanoGPT or Featherless processes every request, and we keep account-linked usage metadata. Private chat is not anonymous and not end-to-end encrypted. Details in our privacy policy.
- Venice: "Your conversation history stays on your device." The default Private mode runs inference "on Venice-controlled GPUs or zero-data-retention partner infrastructure", which Venice says is "trust-based — relies on Venice and partners honoring contracts". TEE and E2EE modes add hardware isolation on fewer models. Anonymous mode: "assume the provider is storing your content."
- OpenRouter: "Prompt and completion are not logged by default", and you can opt in to logging for a 1% discount. It logs basic request metadata. Upstream providers have their own retention policies.
- decuff: "Your words aren’t stored after the reply." Accounts can have no name or email.
- Featherless: "Prompts and completions are never stored."
Red flags
- "No logs" with no mention of metadata or upstream providers.
- No named model providers.
- No retention schedule. Our own retention schedule is not published yet either; our privacy policy says it will be before paid general availability.
- "Anonymous" claims from a service that requires a verified email or phone. We require a verified account.
FAQ
Is temporary or incognito chat private?
It usually means "not saved to your history". The provider still processes the text. See Are temporary AI chats private?.
Can admins read my chats on Unrestricted AI?
Our privacy policy says saved conversations are not available through any admin conversation or prompt-review interface, and that Private chat messages never reach our database.
What is the most private option?
A model running on your own hardware. See local vs hosted.
Should I paste secrets or sensitive personal data?
Not into any hosted chat. Our privacy policy says the same during early access.