What are the seven levels of AI privacy?
| Level | Setup | Where a copy can survive | Who runs it | Built for teams? |
|---|---|---|---|---|
| 1 | AI on your own device | Your device only | You | No, one person per machine |
| 2 | AI on your own servers | Your servers only | Your firm or its IT partner | Yes, with multi-user apps |
| 3 | Hosted in hardware you can check | Encrypted, where the provider says it cannot read it | The provider | Varies |
| 4 | Hosted Workspace that stores no content | No stored content, by design | The provider | Yes |
| 5 | Cloud AI with zero retention by approval or setting | Nowhere, for covered features | The provider | Developer platforms |
| 6 | Chat apps with a retention window | The provider, for a set period | The provider | Yes |
| 7 | Consumer AI apps | The provider, under consumer terms | The provider | No |
Higher levels hold less of your data in fewer places, and leave less to trust. Level 4 sits above level 5 because no content is stored by design and the provider's staff have no access, not because fewer companies are involved: Faradex itself runs on a level-5 service, AWS Bedrock. The lower the level, the less work it takes to start. Every level below rests on the providers' own published pages, checked November 10, 2026. None of them has been audited for this guide.
A note on subpoenas before the details. A subpoena or a preservation order can only reach data that exists. So for each level, the useful question is where a copy exists after the work is done, and who holds it. This is general information, not legal advice.
Level 1: What does it mean to run AI on your own device?
Prompts and files: The model is downloaded once and then runs on the computer in front of you. LM Studio's documentation says: "Nothing you enter into LM Studio when chatting with LLMs leaves your device." The internet is still needed to search for and download models (LM Studio). LM Studio also sells cloud services, which it says are zero data retention (LM Studio). Using those moves you down to level 5. The quote above is about models running on your own machine.
Who can see them: Only people with access to that device.
Reachable by subpoena: Chats saved on the device. Because you hold them, a demand for them comes to you, the same as for any document on your laptop.
The trade-off: You get only the open models your hardware can run, and those are usually smaller than what the large providers run in their clouds. Installing and choosing models takes some technical confidence. Colleagues cannot share a workspace.
Good fit for: Individuals and solo practitioners with a capable computer, or anyone whose rule is that confidential data stays on hardware they hold.
Examples: LM Studio, and Jan, which its makers describe as "an open source alternative to ChatGPT that runs 100% offline on your computer" (Jan on GitHub). Jan can also connect to cloud models from OpenAI, Anthropic and others, and a chat sent to one of them leaves your computer.
Level 2: What does self-hosted AI mean?
Prompts and files: A firm installs an AI chat app on servers it controls, and conversations stay on those servers. AnythingLLM's Docker edition is "both a single-user and multi-user application you can install on any webserver" (AnythingLLM). Open WebUI describes itself as "a self-hosted AI platform" that is "built to run entirely offline" (Open WebUI).
Who can see them: The firm's users and the people who administer the servers.
Reachable by subpoena: Anything the servers keep, including chat history, uploads and backups. The firm holds it, so the firm receives the demand.
The trade-off: Someone must install, patch, secure and back up the servers, and buy hardware good enough for the models. There is also an easy way to lose the benefit: Open WebUI supports "both local and cloud-based models," and a self-hosted app connected to a cloud model sends prompts to that model's provider. At that point, the setup belongs on a lower level.
Good fit for: Firms with an IT team or a technical partner, and firms required to keep data on their own equipment. AnythingLLM's multi-user edition adds "admin and rule-based access for workspaces and documents."
Examples: AnythingLLM, Open WebUI. Whether a firm needs its own server at all is on Do you need a private AI server?
Level 3: What is confidential computing for AI?
Prompts and files: The model runs in the provider's cloud, inside a sealed area of the hardware called an enclave, designed so the provider itself cannot read what is inside. Before sending anything, the user's software checks a hardware-signed record of what the enclave is running, a step called remote attestation. Privatemode says it "keeps your prompts encrypted in memory while they are processed, and your client checks that before it sends your data" (Privatemode).
Who can see them: The design aims to shut the provider out. Tinfoil says it "does not see prompts, completions, uploaded files," while it does see account, billing, usage and network details (Tinfoil).
Reachable by subpoena: Little or no readable content. Tinfoil says it does "not retain API prompt or response content after the response is returned, apart from temporary prompt caching in encrypted enclave memory," and that files saved in Tinfoil Chat are encrypted "with keys held on your device, so neither Tinfoil nor our storage providers can read them, including in response to legal process." Account and billing records remain.
The trade-off: Model choice is narrower: Tinfoil verifies "the exact open-source code and model weights we publish," so it runs open-source models. Some of these options are sold to individuals. Venice sells its TEE and E2EE modes on its Pro plan, and its pricing page shows no team plan (Venice). Venice calls E2EE "Verifiably and entirely private from all parties" (Venice). Tinfoil itself says: "Confidential computing reduces risk, but it does not eliminate all risk."
Good fit for: Technically minded buyers who want hosted AI and want to check the provider's claims with the hardware, rather than take them on trust.
Examples: Tinfoil, Privatemode, Venice (TEE and E2EE modes), and Apple's Private Cloud Compute, which Apple says "deletes the user's data after fulfilling the request" (Apple Security). Apple's service runs Apple's own models for Apple devices and is not sold as a business AI product.
Level 4: What is a hosted AI Workspace that stores no content?
Prompts and files: The provider gives each customer its own environment and designs it so that content is never stored. Faradex is the example here. Each customer gets an isolated, single-tenant Workspace on dedicated infrastructure. Files, conversations and the Workspace's database are held in memory (RAM), with no backups. Your conversations stay in your Workspace's memory while your team works, and nowhere else. Reset it, and they're gone. Faradex cannot recover a reset Workspace. AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them.
Who can see them: The customer's own team. Faradex says its staff have no access: SSH is disabled, with no remote access and no admin backdoor. On Bedrock, Anthropic has no access to prompts or answers.
Reachable by subpoena: No stored content. Faradex puts it this way: "You cannot subpoena what does not exist. You cannot preserve what was never stored." Faradex can be subpoenaed; it just has nothing to produce.
The trade-off: This level is hosted, so it is less private than levels 1 to 3. The customer relies on Faradex's design and on AWS, rather than holding the data itself (levels 1 and 2). The AI model runs on AWS Bedrock, so AWS's level-5 rules, including its exceptions, apply here too: AWS applies the zero data retention setting Region by Region, and it lists some models whose traffic it keeps for up to 30 days for abuse detection (see level 5). Faradex's security statement reads: "We are SOC 2 compliant. Specifics can be found at https://faradex.ai/security." Because no content is kept, a firm with record-keeping duties must keep its own copies of what it needs. All Faradex services are USA-only, with no choice of region. It is a paid product; pricing is at faradex.ai/pricing.
Good fit for: Small and mid-sized firms, roughly 50 employees, that handle confidential client work and have no AI infrastructure of their own.
Examples: Faradex. The closest neighbor is Proton Lumo for Business, a no-logs chat tool sold per user to teams. Proton says that after Lumo answers, "the data is erased." It differs in two ways: saved chat history is stored on Proton's servers with zero-access encryption, so that "no one but you can read your chat history" (Proton), and it runs open-source models only (Proton). Both are compared in more detail on Best Zero Data Retention AI.
Level 5: What is zero data retention by approval or setting?
Prompts and files: A large provider processes requests in its cloud and stops keeping them once zero data retention is switched on, for the features it covers. How it is switched on differs:
- OpenAI API: the controls are "subject to prior approval by OpenAI" (OpenAI).
- Anthropic API: "To request ZDR for your organization, contact the Anthropic sales team" (Anthropic).
- AWS Bedrock: a self-serve account setting, data_retention_mode: none (AWS).
- Google Cloud: the customer switches off several features, such as caching and request logging (Google Cloud).
- Venice, Private mode: "no prompt or response is stored," which Venice labels "Trust-based": it relies on Venice and its partners honoring their contracts.
Who can see them: The provider's systems, under exceptions each provider publishes. OpenAI keeps images flagged as possible child sexual abuse material for review. AWS says traffic to Claude Fable 5, 5.1 and 5.2 "will be retained for up to 30 days for automated offline abuse detection" (AWS).
Reachable by subpoena: Nothing from covered features. When a court ordered OpenAI to preserve user content in NYT v. OpenAI, API customers on zero data retention endpoints were outside the order. OpenAI explained: "Because it is not stored, this court order doesn't affect that data" (OpenAI). Uncovered features are different: OpenAI's assistants, files and batches keep data "Until deleted."
The trade-off: These are developer platforms, so someone has to build the tool staff will use. Approval can mean a sales process, and coverage is set feature by feature. Microsoft offers a related option, modified abuse monitoring, for customers "managed by a Microsoft account team or under an eligible program." It switches off storage for human review, although "automated review may still be conducted" (Microsoft).
Good fit for: Companies with engineers who build their own AI apps.
Examples: OpenAI API, Anthropic API, AWS Bedrock, Google Cloud, Venice in Private mode. For the setup details, see How Zero Data Retention Works.
Level 6: Which AI chat apps keep data for a set period?
Prompts and files: These chat products keep conversations for a time:
- ChatGPT Enterprise and Business: admins control retention, and deleted conversations are removed "within 30 days," with exceptions for legal requirements and, on Business, when OpenAI says longer retention is "reasonably necessary to protect our services or any third party from harm" (OpenAI).
- Claude Team and Claude Enterprise: Anthropic says these interfaces "are not ZDR-eligible" (Anthropic). Enterprise data "is retained indefinitely unless a custom retention period is set," and the shortest custom period is 30 days (Anthropic).
- Kagi Assistant: threads expire 24 hours after the last message or are kept permanently (Kagi). Kagi's own table lists 30 days at OpenAI, and 30 days at Anthropic and Bedrock with flagged prompts kept for two years (Kagi).
- Duck.ai: strips identifying details such as the IP address, then relies on model providers that delete data "at most within 30 days, with limited exceptions for safety and legal compliance" (DuckDuckGo).
Who can see them: The app's provider and, where prompts are passed on, the model providers behind it.
Reachable by subpoena: Anything inside the retention window. The 2025 preservation order in NYT v. OpenAI covered ChatGPT Team, though not ChatGPT Enterprise or Edu.
The trade-off: The firm gives up control over how long its data lives. A no-training promise does not change that. ChatGPT Enterprise and Claude Enterprise both say they do not train on your business data unless you opt in, but training and retention are separate promises; see Zero Data Retention vs. No Training.
Good fit for: Teams whose work is not confidential, or whose policies allow a retention window.
Examples: ChatGPT Business and Enterprise, Claude Team and Enterprise, Kagi Assistant, Duck.ai.
Level 7: How private are consumer AI apps?
Prompts and files: Conversations are kept on the provider's servers under its consumer terms. Venice says of its own Anonymous mode: "you should assume the provider is storing your content."
Who can see them: The provider.
Reachable by subpoena: Whatever the provider keeps. The preservation order in NYT v. OpenAI made OpenAI keep consumer chats from ChatGPT Free, Plus, Pro and Team that it would otherwise have deleted. According to OpenAI, the order stopped applying on September 26, 2025, though it still holds a limited set of user data from April to September 2025 (OpenAI).
The trade-off: The least control, in exchange for the lowest cost and no setup.
Good fit for: Personal questions with nothing confidential in them.
Examples: ChatGPT Free, Plus and Pro; Venice in Anonymous mode.
How can you tell which level a tool is on?
Ask the provider four questions, and read the answers on its own pages rather than in its marketing:
- Where does the model run? On your device, on your servers, or in the provider's cloud.
- Can anyone at the provider read my data while the AI works? Look for a technical reason, such as an enclave you can check or a design with no staff access, and not only a policy.
- Is a copy kept afterwards, and for how long? Look for a number of days, and for the exceptions.
- Which features does that cover? Zero data retention is often set per feature, so check the one your team will use.
The answers place a tool on one level. Then pick the highest level your team can run, and go lower only as far as the work needs. For the words to look for, see What Is Zero Data Retention?.