The teardown

What the popular AI tools are actually allowed to do with your data.

Not the marketing. The terms. We read what ChatGPT, Gemini, Copilot, Claude, Perplexity, Grok, Meta AI and DeepSeek permit themselves to do with what you type, on the tiers most professionals actually use.

Read from each tool's published consumer terms. Reviewed July 2026.

On the consumer tiers that a solo GP, a small firm or a sole practitioner actually uses, the mainstream AI tools generally store what you type, many train on it by default, some let staff read it, and several sit under legal regimes outside the UK. "The business tier is safe" is true and beside the point, because most professionals are not on the business tier. If you would not post it publicly, most of these tools are not where it belongs.

Consumer tiers, read from the terms.

The rubric is the same for each tool: does it train on your inputs by default, how long does it keep them, can a human read them, and what is the catch. Figures are for the personal or free tiers unless noted.

ToolTrains on your input by default?RetentionHuman reviewThe catch
ChatGPTYes, opt-out buried in Settings, Data ControlsSaved until you delete, then up to 30 daysYes, staff may reviewAPI and business tiers are not trained on; consumers are. A 2025 US court order (the NYT case) forced retention of even deleted chats, since largely lifted.
Google GeminiYes, unless turned offApps Activity 18 months by default; reviewed chats kept up to 3 years even after you deleteYes, a subset goes to human reviewersPaid API and Vertex are exempt; the consumer app is not.
Microsoft CopilotYes on consumer, on conversation activity unless opted outPer policyPer policyCommercial (work account) and the Microsoft 365 apps are covered by enterprise data protection; standalone consumer Copilot is not.
Anthropic ClaudeYes, unless you decline; the Aug 2025 setting defaults to on30 days if you decline training, 5 years if you allow itFor safety classificationConsumer plans only (Free, Pro, Max); Work, Gov, Edu and API run under separate commercial terms and are exempt.
PerplexityYes, with an opt-out toggle in the productKept until you delete, then open-endedPer policyIn 2026 Perplexity dropped its 30-day deletion promise and its change-notification promise from the policy text; the in-product opt-out toggle remains. Enterprise and the Sonar API are zero-retention.
xAI GrokYes, on your prompts, its responses, and your public posts (non-EU)Retained for security regardless of your settingxAI can read chats; not end-to-end encryptedOpt-outs are split across several settings and are not retroactive.
Meta AIYes, on your Meta AI chats and public postsPer policyPer policyThere is no simple in-app opt-out and the WhatsApp toggle was removed; UK and EU users can object under GDPR via a form. WhatsApp messages stay end-to-end encrypted, but Meta AI interactions do not.
DeepSeekYesStored indefinitely on servers in ChinaPer policySubject to China's 2017 National Intelligence Law, which can compel access. Banned for government use in several countries. A Jan 2025 breach exposed over a million records.

Methodology: each verdict is read from the tool's own current published consumer terms and from public reporting, on the tier a typical solo professional uses. Terms change often, so we date-stamp this page and you should confirm any single cell against the live policy, linked in the sources below, before you rely on it.

In 2026 the protections got weaker, not stronger.

The usual advice is "just opt out." The problem is that the opt-outs are quietly being weakened across several providers at once:

The direction of travel is less user protection, not more. A tool you were told to "just opt out of" last year may quietly train on you this year. An opt-out you have to keep watching is not confidentiality.

Your duty does not have an opt-out.

If you are a solicitor, a doctor, an accountant or in HR, the confidential material you handle is not yours to feed into a system that stores or trains on it. Every UK regulator has said so, and a UK tribunal has ruled that putting confidential documents into a public AI tool can waive privilege and trigger a duty to notify the regulator and the ICO. The tools in the table above are, on their consumer tiers, mostly the wrong place for that work. That is the entire reason Hush exists.

The rubric, and the tool that passes it.

A tool is safe for confidential professional work when the answer to all four is right:

Hush runs on hardware we own and keeps no conversations or documents. What you type is processed, then gone, so there is nothing to leak, sell, subpoena or quietly start training on. It is the confidentiality a Microsoft lawyer told the French Senate, under oath, he could not promise.

Use AI on the work you cannot put into ChatGPT.

Try it on a real letter or a real document, the kind you would never paste into the tools above. No card, no account needed.

Start Free →

Prefer to check one tool at a time? Use the interactive AI training checker. Or see the side-by-side comparisons, or read our own security page.

Sources and verification. Read each tool's own policy and confirm before relying on this page. OpenAI ChatGPT · Google Gemini · Microsoft Copilot · Anthropic Claude · Perplexity · xAI Grok · Meta AI · DeepSeek (IAPP). Reviewed July 2026.

This page summarises third-party terms as published and reported at the review date, for general information. It is not legal advice, and the tools' terms change. Verify any point against the primary source before you act on it.

Compiled by the founder, Dr W.J Carter.