The short answer
The best private AI for sensitive data in 2026 is Claude on Anthropic's Team or Enterprise plan, because that content is not used for training by default and Enterprise adds custom retention, audit logs and SCIM. Vibe by Mistral AI is the runner-up for EU data location, with Enterprise opted out of training. The best free option is self-hosting open-weight models such as Cohere Command A+ under Apache 2.0, so data never leaves your hardware.
| Our pick | Claude Team or Enterprise, by Anthropic |
|---|---|
| Runner-up | Vibe Enterprise, by Mistral AI (EU hosting by default) |
| Best free option | Self-hosted open weights (Cohere Command A+, GLM-5.3-Flash, DeepSeek-V4-Pro) |
| Personal plans | ChatGPT, Claude, Perplexity, Vibe and Grammarly train by default unless you opt out |
| Business plans | ChatGPT Business and Enterprise, Claude Team and Enterprise, Perplexity Enterprise: no training by default |
| Remember | Opting out is not retroactive, and not training does not mean not stored |
What "private" really means for an AI tool
Three separate questions decide whether an AI tool is safe for sensitive data, and marketing pages tend to blur them:
- Training: can the provider use your prompts and files to improve its models? On most personal plans the answer is yes by default, until you switch it off.
- Retention: how long does the provider keep your data, even when it does not train on it? "Not used for training" does not mean "deleted".
- Location and control: where is the data processed and stored, under which law, and who at your organisation can see and govern it?
Two more details catch people out. Opting out usually only works from that point on: Anthropic says switching off training does not remove data from training that has already started or from models already trained, and Perplexity says its opt-out only covers data collected after you switch it off. And thumbs up or down feedback can still be used for training on some services, including ChatGPT, even with training switched off.
Training defaults by tool and plan, November 2026
| Tool | Personal plans | Business plans | Where to switch it off |
|---|---|---|---|
| ChatGPT | Free, Plus and Pro: on by default (the pricing page also shows the opt-out on Go) | Business, Enterprise, Edu and ChatGPT for Healthcare: not used by default | Settings, Data controls, Improve the model for everyone |
| Claude | Free, Pro, Max, Ultra: used unless you opt out | Team and Enterprise: not used by default | Settings, Privacy, Help Improve our AI models |
| Gemini | Used to improve services, including training, while Keep Activity is on; some chats read by human reviewers | Workspace: no human review or training outside your domain without permission | myactivity.google.com/product/gemini, Keep Activity |
| Microsoft Copilot | Updated app (since 18 August 2026): not documented, no switch shown | Work accounts: not used to train foundation models | No switch documented in the updated app |
| Perplexity | Free, Pro, Max: on by default | Enterprise Pro and Max: never used | Account settings, Preferences, AI data retention |
| Vibe (Mistral AI) | Used by default until you opt out | Enterprise: opted out by default; Team: admins can switch it off | Mobile: Settings, Account, Data & Account Controls |
| Notion AI | Not used by default | Not used by default | No switch needed |
| Grammarly | On by default | Off by default for Enterprise, sales-bought and Education accounts | account.grammarly.com/security/privacy |
| DeepL | Free: texts may be used, no opt-out described | Paid plans: never used, deleted after translation | Use a paid plan |
| DeepSeek | Used after de-identification unless you switch off Improve the model for everyone; stored in China | No business chat plan | App settings, or email privacy@deepseek.com |
Our pick: Claude on the Team or Enterprise plan
Why: Claude, made by Anthropic, combines strong work on long, sensitive documents with clear business terms. Team (2 to 150 users) and Enterprise content is not used for training by default. Team adds SSO and admin controls. Enterprise adds SCIM, audit logs, a Compliance API, role-based access, custom data retention and a HIPAA-ready offering, which is what legal, HR, finance and health teams usually need to sign off. For one-off sensitive questions on any plan, incognito chats are not saved to history or memory and are not used for training, though they are kept for 30 days by default.
Who it suits: organisations that want a top general assistant for confidential documents and can buy a work plan.
The catch: only Enterprise sets custom retention, and Anthropic says data for Claude Enterprise and its API is stored in the US, so EU organisations with strict data location rules should check with their data protection officer. Conversations flagged by Anthropic's safety systems can still be used for trust and safety work. If you prefer OpenAI, ChatGPT Business (2 to 200 people) and Enterprise are also excluded from training by default, and new ChatGPT Enterprise and Edu workspaces can store content at rest in Europe. Compare them in ChatGPT vs Claude.
Runner-up: Vibe by Mistral AI, for EU data location
Why: Vibe, Mistral AI's assistant (formerly Le Chat), keeps data in the EU by default, and Mistral is a French company. Vibe Enterprise is opted out of training by default and can be deployed privately or on premises, with SAML SSO and audit logs. Its Work mode shows its tool calls and asks for confirmation before sensitive actions such as sending email, posting messages or changing data in connected tools.
Who it suits: European organisations for whom where the data sits matters as much as training, and teams that may later want to run models on their own infrastructure.
The catch: individual plans train on chats until you opt out, and Mistral has not confirmed in what we could check that the Team plan is excluded by default, so ask or have an admin switch training off for the organisation. Some processing uses listed subprocessors with temporary transfers, and zero data retention is not available for Vibe Work or Chat. See ChatGPT vs Vibe.
Best free option: run an open-weight model yourself
Why: the only way to be certain no provider sees your data is to run the model on hardware you control. Several strong models are free to download:
- Cohere Command A+ (released 20 May 2026) has open weights under Apache 2.0. Cohere says it runs on as few as two H100 GPUs in its compressed W4A4 form. Cohere's North platform can also be deployed self-hosted, in a VPC, on premises or air-gapped, and for private deployments Cohere says it receives no prompts or generations. See our Cohere guide.
- GLM-5.3-Flash from Z.ai is under the MIT licence; the larger GLM-5.3 uses a custom licence. Z.ai's hosted service may use data to improve models, which self-hosting avoids.
- DeepSeek-V4-Pro weights are on Hugging Face under the MIT licence, but at 1.6 trillion parameters (49 billion active) it needs data-centre hardware.
- Mistral Medium 3.5 is a 128B open-weight model from Mistral AI.
Who it suits: organisations with an IT or engineering team and data that must never leave their infrastructure.
The catch: free to download is not free to run. You pay for the hardware, and you own security, updates, access control and monitoring. Licences differ (Cohere's North Small Translate weights, for example, are for non-commercial use only), so read each one. And a self-hosted model is only as private as the server it runs on. Read what an open-source model is and open-source ChatGPT alternatives.
How to choose a private AI setup: three questions
- What is the most sensitive thing anyone will paste in? Plan for that, not the average. Client secrets, health data and personnel files need a work plan with a data processing agreement or self-hosting.
- Does data location matter legally? If it must stay in the EU, look at Vibe, at OpenAI's EU data residency for new Enterprise workspaces, or at self-hosting, and ask each provider in writing.
- Who will run it? No IT team: buy a business plan. An IT team with GPU budget: self-hosting becomes realistic.
Workflow: setting up AI for sensitive work, end to end
- Classify your data. Agree four levels: public, internal, confidential and regulated, with examples for each.
- Map levels to tools. Public and internal: any approved tool with training switched off. Confidential: business plans only. Regulated: Enterprise with the right agreement, or self-hosted.
- Buy and configure. Set up SSO, check training is off by default, set retention where the plan allows it and restrict connectors to what people need.
- Sign the paperwork. Get the provider's data processing agreement and, for health data, any HIPAA arrangement the plan offers.
- Switch off personal-plan training. Tell staff how to do it in the tools they use privately, and that work data stays out of personal accounts.
- Train people. A one-page rule sheet with the data levels, the approved tools and a redaction habit for anything borderline.
- Review quarterly. Defaults change: Microsoft's updated Copilot app and GitHub's April 2026 change to training on Copilot interaction data are recent examples. Recheck each tool's privacy page every quarter.
Role-specific advice is in AI for HR, AI for legal documents and AI for finance.
6 prompts for working privately with AI
Use these to set up your rules and to keep sensitive details out of chats. Run the redaction prompt in a self-hosted model or on a work plan, never on a personal account.
Prompts to copy
Write an AI data policy
Draft a one-page AI use policy for a [company type] of [size] in [country]. Define four data levels (public, internal, confidential, regulated) with three examples each from our work: [describe]. For each level, say which of these tools may be used: [list approved tools and plans]. Plain English, no legal jargon. List points our legal team must check.
You get: A first-draft policy staff can actually follow
Redact before you ask
Replace every personal or confidential detail in this text with a labelled placeholder: names as [PERSON 1], companies as [COMPANY 1], amounts as [AMOUNT 1], plus emails, phone numbers, addresses and ID numbers. Keep the meaning intact and output a key I can store separately. Text: [paste]
You get: A safe-to-share version and a separate key
Check a vendor's privacy terms
Read this AI vendor's privacy policy and terms: [paste or attach]. Answer in a table, quoting the exact wording for each: is our data used for training by default, can we opt out, how long is data kept, where is it stored, which subprocessors are named, and what changes when we buy a business plan. Write 'not stated' where the documents are silent.
You get: A quoted privacy summary for procurement
Prepare vendor questions
We are evaluating [tool] for [use case] with [type of data]. Write 12 precise questions for the vendor about training, retention, data location, access by their staff, subprocessors, security certifications, incident notification and deletion on exit. Group them by topic.
You get: A vendor questionnaire your team can send
Outline a data protection impact assessment
Outline a data protection impact assessment for using [tool] to [purpose] with [data types]. Sections: description of processing, necessity, risks to individuals, measures, residual risk. Fill in what you can from this information: [paste], and mark every gap for our data protection officer.
You get: A DPIA skeleton with gaps flagged
Compare self-hosting with a business plan
Compare two options for [team size] people using AI on [data type]: a business plan of [tool] or self-hosting [open-weight model]. Cover data control, set-up effort, ongoing maintenance, skills needed, and the risks of each. Do not estimate prices; list the cost items we must get quotes for.
You get: A decision table without invented numbers