Skip to content
How to Secure AI
← Resources

Reference

Which AI tools train on your data

A maintained comparison of what the major assistants do with what you type — by tier, with the date each claim was last checked against the vendor's own terms.

By Mark Jones, IT Consulting SpecialistPublished July 16, 2026Sources last verified July 16, 2026

Every major AI assistant now sells at least two versions of itself: a personal tier and a business tier. They look nearly identical and behave very differently in the one respect that matters to a business — what the vendor is permitted to do with what you type.

This page tracks that, by tier, against the vendors’ own current terms. Each claim carries the date it was last checked.

How to read this

Three columns do most of the work:

Trains on your input by default. Whether, absent any action by you, the vendor may use your conversations to improve its models. The words by default matter — an opt-out that exists but that nobody has toggled provides no protection in practice.

You can turn it off. Whether a control exists, and whether an ordinary user would ever find it.

Retention. How long the vendor holds your content, and whether you can set that yourself. Retention is the sleeper issue: a vendor that doesn’t train on your data but keeps it for three years still has your data for three years, and that’s what shows up in diligence.

What this page deliberately does not do is tell you a vendor is “safe.” Safe depends on your contracts, your regulator, and your data. It tells you what the terms say. What that means for you is a separate conversation.

The comparison

Tool and tier Trains by default Can you turn it off Retention Checked
ChatGPT — Free / Plus Yes Yes, under Data Controls 1 Until you delete it 2026-07-16
ChatGPT — Business / Enterprise 2 No Opt-in only Admin sets it 2026-07-16
OpenAI API No Opt-in only 30 days, or none on request 2026-07-16
Claude — Free / Pro / Max Yes 3 Yes, under Privacy 30 days off, 5 years on 2026-07-16
Claude — Team / Enterprise No Opt-in only 4 Indefinite by default 2026-07-16
Anthropic API No Opt-in only 30 days 2026-07-16
Gemini — Free / consumer Yes Yes, in Gemini Apps Activity 18 months 5 2026-07-16
Gemini — Workspace No Needs your permission 90 days to indefinite 2026-07-16
Microsoft Copilot — Consumer Yes 6 Yes, under Privacy 18 months 2026-07-16
Microsoft Copilot — M365 / commercial No Contractual Your tenant, admin sets it 2026-07-16

The notes are the interesting part

1. Rating a reply overrides your opt-out. On ChatGPT’s consumer tiers, giving a response a thumbs-up or thumbs-down sends that entire conversation to OpenAI for training — even with “Improve the model for everyone” switched off. The opt-out you set months ago does not survive a single idle thumbs-up.

2. “ChatGPT Team” is no longer a thing. It’s now called ChatGPT Business. If you’re comparing plans against an article written before the rename, you’re reading about a product that doesn’t exist under that name.

3. Claude’s consumer default is genuinely unclear, and I’m not going to pretend otherwise. The Consumer Terms say Anthropic may train on your materials “unless you opt out of training through your account settings” — that’s opt-out language, and it’s the operative contract. Anthropic’s own Privacy Center describes the same control in opt-in terms. Since 8 October 2025 existing users have had to make an active choice, and new users choose at signup, so in practice most people have answered the question one way or the other. What I could not establish from Anthropic’s own pages is which way the toggle ships for a brand-new account. If that matters to you, open Settings → Privacy and look, rather than trusting anyone’s summary — including this one.

4. On Claude’s business tiers, feedback is the leak. Anthropic doesn’t train on Team or Enterprise content, but a user submitting thumbs-up/down feedback can hand over that content anyway. Owners can shut this off entirely: Organization settings → Data and Privacy → “Rate chats”.

5. Gemini’s human-review copies outlive your delete button. Conversations sampled for human review are kept up to three years and are not removed when you delete your activity or turn “Keep Activity” off. Deleting your history deletes your copy, not theirs. This is stated plainly in Google’s own documentation and is, to my mind, the single most under-appreciated line on this page.

6. Consumer Copilot’s training default doesn’t apply to everyone. Microsoft excludes users signed in with a work (Entra ID) account, users who aren’t signed in at all, under-18s, and users in Brazil, China (excluding Hong Kong), Israel, Nigeria, South Korea and Vietnam. Everyone else is trained on unless they opt out, under Profile → Privacy → “Training on conversation activity”.

Sources

Every claim above was read from the vendor’s own current pages on 16 July 2026. Where a vendor’s terms and its help pages disagree, that disagreement is noted rather than resolved.

OpenAI

Anthropic

Google

Microsoft

The pattern underneath the table

Once the cells are filled, the same shape shows up almost everywhere, and it’s worth stating plainly because it’s more durable than any individual row:

Consumer tiers tend toward training-by-default with an opt-out. You are, in a real sense, paying with your data — even on the paid personal tiers, which surprises people who assume that paying changed the arrangement.

Business and API tiers tend toward not training by default, contractually. This is the vendors’ central promise to enterprise buyers and the thing they’re least likely to quietly reverse, because it’s what the contract is for.

The gap between those two is a per-seat fee. Usually modest. The same model, the same answers, a different legal posture and an admin console.

Which brings you back to the only question that actually matters, and it isn’t which vendor is best:

Which tier is your team logged into right now?

If the answer is “I don’t know” — and it usually is — then the table above is interesting but premature. Start with what your team is already doing, then come back here.

Every claim on this page is checked against the vendor’s own current terms, with the date shown. Vendors change these quietly. If you find something out of date, tell me and I’ll fix it.