Skip to content
Private AI

Private AI: what it means, and how to judge one

Private AI is artificial intelligence you can use without your words being stored, read by others or used to train a model. It comes in three forms — a model on your own device, a deployment inside your company, or a private AI service in the cloud — and each draws the line in a different place. Evanesca is a private AI agent and chat of the third kind: it runs its models in Europe, keeps no copy of your messages, and shows under every answer where it ran.

Facts checked on 6 Nov 2026

What private AI means

“Private” is not a switch on a settings page; it is the answer to four questions about your words once you have sent them: who keeps them, for how long, who can read them, and whether they train the next model. Mainstream AI chats answer most of them in their own favour by default. ChatGPT, for one, stores conversations “on OpenAI systems and our trusted service providers’ systems in the US and around the world” until you delete them source, and on its plans for individuals may use them to train its models unless you opt out source. The same questions for Gemini, Claude, Copilot and Meta AI are answered on the sources page.

Three quite different things are sold as private AI:

  • On-device, or local, AI — an open-weight model downloaded and run on your own computer or phone. Your words need not leave the device, but the model has to fit in its memory, so it is usually smaller than the models served online what open-weight means.
  • Private deployment — a company runs models on its own servers, or on a cloud it rents, for its own staff. Privacy then means the company's own rules: its logs, its administrators, and the jurisdiction of its cloud provider own cluster or API?.
  • Private AI services — a chat or an agent in the cloud that commits not to keep your conversations or train on them: Lumo by Proton, Venice AI and Evanesca among them. You keep the convenience of the cloud, and rely on the provider's design and on the company that actually runs the model.

A checklist to judge any private AI

Whatever form it takes, a private AI should answer each of these in writing, with a date. Where it cannot, assume the less private answer.

  1. Retention — what is kept after the answer, and by whom. Ask twice: the service you use, and the company that runs the model, which often has its own rule for abuse monitoring or errors.
  2. Training — whether your conversations can train a model, and whether that is a setting you have to find and switch off, or simply how the service works.
  3. Jurisdiction — which company runs each link, and under which law. Servers in the EU are not enough: the US CLOUD Act lets US authorities require a provider under US jurisdiction to disclose data in its “possession, custody, or control”, wherever it is stored source — why EU hosting is not enough.
  4. Who can read it — staff reviewing abuse or feedback, the operator's support, contractors. A credible answer names them, and says when.
  5. Logs and metadata — even without the text, timestamps, IP addresses and usage records say a lot about you. Look at what is logged around the content.
  6. Encryption — encrypting the connection is common; encrypting the prompt so that the provider itself cannot read it is rare, and worth knowing about.
  7. Tools — when the AI searches the web or acts as an agent, what leaves for those tools: the whole conversation, or only a search query?

Private AI options side by side

The Lumo and Venice AI cells come from their own pages, read on 6 October 2026, and link to them. The full comparisons are on Evanesca vs Lumo and Evanesca vs Venice AI.

FIVE WAYS TO USE PRIVATE AI
OptionWhere the model runsKept after the answerWho can read the promptWhat you give up
A local model on your deviceYour computer or phoneWhatever your device and the app keepYou, and whoever can use the deviceModels small enough for its memory; you install and update them
A deployment in your companyIts servers, or the cloud it rentsIts own logging policyIts administrators, and its cloud providerHardware, staff and upkeep; the cloud provider's jurisdiction
Lumo, by Proton“Proton’s fully European infrastructure, protected by Swiss privacy laws” sourceErased: no logs, no metadata; an encrypted history synced to your devices sourcePrompts are encrypted so that only Lumo's GPU servers can decrypt them sourceLumo picks the model for each request source; a Lumo Plus subscription for more source
Venice AINot stated on its privacy page sourceHistory on your device: “Venice never stores it on its servers” sourceDepends on the mode: Venice-controlled GPUs or zero-retention partners by default; end-to-end encryption on the Pro plan, which Venice itself cannot read sourceA US company, under the laws of Wyoming source
EvanescaIn France by Scaleway and OVHcloud (as a fallback), in Mistral AI's own EU region, in Germany by Verda and AKI.IO and in Switzerland by Infomaniak sovereigntyNot stored by Evanesca; of your messages, only billing data — tokens and cost; each operator's own rule below privacyNo one at Evanesca; the operator processes the prompt in clear to write the answerNo end-to-end encryption; history stays in the browser of the device you use

Private AI on your own machine or servers

Running a model yourself is the most private option on paper, because no third party sees the prompt. In practice it is a trade. The weights must fit in your memory, the largest open models need several data-centre GPUs, and you maintain the software, the updates and the security yourself how much memory open weights need.

For a company, the question is cost and use: a cluster pays off only if it is kept busy. Self-hosting an LLM vs paying per token does the sums, and your own cluster or an API? compares the two routes link by link.

Local and cloud also mix well: a local model for the most sensitive notes, and a private service in the cloud for what a small model handles poorly — long documents, web research, larger models.

Where Evanesca fits as a private AI

Evanesca is a private AI service built around jurisdiction and records rather than encryption. Every model it serves runs in Europe — in France by Scaleway and OVHcloud (as a fallback), in Mistral AI's own EU region, in Germany by Verda and AKI.IO and in Switzerland by Infomaniak — and it serves none outside Europe models.

  • Your messages are held in memory while the answer is written and are not stored by Evanesca. Of your messages, only billing data — tokens and cost — is kept; your account and payments are kept as the law requires privacy.
  • Your history lives in your browser. Memory is off by default; switched on, it is a profile of at most 4,000 characters, kept in the browser too. Read-aloud uses your device's own voices.
  • Evanesca never trains a model on what you type.
  • Web search is off by default. Switched on, it sends Linkup, a French company, only the query the model writes and the addresses of the pages it reads — never the conversation source web search.
  • A passport under every answer names the model, the operator, the country, the law and what was kept passport.

The company that runs each model has its own rule, given here as each one states it: OVHcloud keeps “only the data required for billing purposes” source; Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days source; Mistral AI keeps no input or output beyond the time needed to answer, as Evanesca's account has been on zero data retention since 17 November 2026 source, and lets its customers object to training — Evanesca's account objects source source; AKI.IO says prompts and outputs are neither logged nor stored, and keeps usage metadata for 90 days source source; Verda says it stores no request and holds the context in an in-memory inference cache removed within 2 hours source; Infomaniak states that queries are “neither recorded nor used to train models” source.

Two limits, said plainly. Evanesca does not encrypt your prompt end to end: the operator reads it to write the answer. And Evanesca's own servers — the website, the chat and the agent's loop — run at Vercel, a US company, in its Paris region, with no copy of your messages kept there privacy. OVHcloud, Scaleway, Mistral AI and Infomaniak are European companies with no US parent, out of the CLOUD Act's direct reach; AKI.IO and Verda do not publish their full shareholding, so Evanesca cannot say that they have no US parent sovereignty.

For teams, Evanesca signs an Art. 28 DPA and offers SSO business. To hand over whole tasks rather than questions, see the private AI agent page; for the European side of the question, European AI.

When another private AI is the better choice

  • Nothing may leave your premises, or you work offline: a local model, or a deployment of your own.
  • You want the provider unable to read your prompt: Venice AI's end-to-end encrypted mode, on its Pro plan source.
  • You want an encrypted history synced across devices, use without an account, and apps for iOS and Android: Lumo source source.
  • Privacy is not the point and you want OpenAI's own models, voice or scheduled tasks: ChatGPT source — Evanesca vs ChatGPT.

Evanesca fits when the material is confidential, should be processed in Europe by named companies, and you want a record of what happened to each answer — paid per use, with no subscription required and credit from €5 pricing.

Frequently asked questions

What is private AI?

Private AI is artificial intelligence you can use without your words being stored, read by others or used to train a model. It can run on your own device, on your company's servers, or as a cloud service that commits not to keep your conversations or train on them.

Is ChatGPT private?

Not by default on its plans for individuals: OpenAI stores conversations in the US and around the world until you delete them, may use them for training unless you turn off “Improve the model for everyone”, and authorised staff and service providers may access them for abuse, support, legal matters or model improvement.

Is private AI the same as local AI?

No. Local AI runs on your own device, which is one way to be private. A private AI service runs in the cloud and commits not to keep, read or train on your conversations; how far that holds depends on which companies run each part, and where.

Where does Evanesca keep my conversations?

In your browser, on the device you use. Evanesca stores no prompt and no answer: of your messages, only billing data — tokens and cost — is kept. The operators that run the models have their own rules; Scaleway, for instance, may keep requests that cause errors or look abusive for up to 14 days.

Can I use private AI for work documents?

Yes, if the service says in writing what it keeps, who can read it and under which law, and, for a company, signs a data processing agreement. Evanesca runs every model in Europe and signs an Art. 28 DPA for teams.