Think about the last ten things you asked an AI chatbot. Odds are the list includes something you'd never type into a search engine: a health worry described in detail, a salary negotiation you're rehearsing, a client's contract you wanted summarized, a fight with your spouse you needed to talk through. That context is exactly what makes AI useful — and exactly what makes the question "where does this text go?" worth ten minutes of your attention. The answers vary more between providers than most people assume, and several of them changed within the last year.
The Default Is Training
Here is the single most important fact in this article: on the consumer tiers of every major AI chatbot — including the paid personal plans — your conversations are used to train future models by default. Paying $20 a month does not change this. The business and API tiers work the other way: no training by default, contractually. The industry has quietly sorted its users into two classes — companies, whose data is protected because they'd sue, and individuals, whose data is the product unless they find the setting.
"Used for training" means your text is retained, may be reviewed by human annotators, and patterns from it can be absorbed into future models. Providers filter obvious personal data, but the filtering is imperfect — which is why the practical rule isn't "trust the filter," it's "control what goes in."
Provider by Provider — What Happens to Your Text
Policies as of July 2026. These change — a settings check twice a year is reasonable hygiene.
Training default: ON, even on paid Plus. Opt out under Settings → Data Controls → "Improve the model for everyone." Temporary Chats are excluded from training but still retained up to 30 days. The API and Business/Enterprise tiers do not train on inputs by default [1][2].
Training default: ON — a notable shift. In August 2025 Anthropic changed its consumer terms: conversations train models by default unless you opt out, with retention up to five years for opted-in users [3]. Previously Claude was the "doesn't train on you" option; that's no longer true on consumer plans without the opt-out. Claude for Work and the API remain no-training by default.
Training default: ON, with human review. The opt-out is disabling "Gemini Apps Activity" — which also deletes your chat history, a design choice that makes privacy cost you functionality [2]. Gemini through a paid Workspace business account is treated like Gmail/Drive data: private, no training.
The most aggressive of the four. Public posts and interactions train Meta AI by default, and as of December 2025, your AI chat conversations also feed ad targeting — with no opt-out for that use [4]. Treat Meta AI as the marketing tool it is: nothing you wouldn't put in a public post.
The Never-Paste List
Regardless of provider and regardless of settings, some categories don't belong in a consumer AI tool. Retention happens even where training doesn't, retained data can be breached or subpoenaed, and your opt-out doesn't rewrite the past.
- Credentials and keys — passwords, API keys, recovery codes. A pasted secret should be considered compromised the moment it leaves your machine.
- Other people's personal data — client lists, patient details, employee records. It isn't yours to share, and in many jurisdictions sharing it this way is a legal violation, not just a bad habit.
- Financial account numbers — cards, bank accounts, tax IDs. Yours or anyone's.
- Confidential business material — unreleased plans, proprietary code, anything under NDA. Several companies have learned this one via headline.
- Anything you'd be harmed by seeing quoted back — the practical catch-all. If a data breach printing this text with your name attached would hurt you, keep it out.
Paying for the personal tier buys you capability, not privacy. Only the business contract buys privacy.
— The two-class rule of AI dataThe Private Alternatives
You don't have to choose between using AI and controlling your data — you just have to choose deliberately.
- Flip the opt-outs today. Five minutes: ChatGPT → Data Controls; Claude → Privacy settings; Gemini → Apps Activity. This is the highest-value privacy action available and most users never take it.
- Use API-backed tools for sensitive work. The APIs of all major providers don't train on inputs by default — one reason work tools built on APIs are structurally more private than pasting into the consumer chat window.
- Business tiers if you handle others' data. ChatGPT Team, Claude for Work, Gemini for Workspace — contractual no-training. If clients' information touches your AI use, this is the cost of doing business properly.
- Local models for the truly sensitive. Open models running on your own hardware never send text anywhere. We covered the full setup in our self-hosted AI guide — it's more attainable than most people think.
We wrote recently about how data fragments assemble themselves — how an IP address here and a purchase there become a detailed portrait. AI conversations are the richest fragment yet added to that pile, because they contain what you meant, not just what you clicked. The providers aren't villains; they're companies doing what their terms of service say they can. The gap is that almost nobody reads those terms — and the defaults are set for the company's benefit, not yours.
So: flip the opt-outs tonight, keep the never-paste list, and route genuinely sensitive work through APIs, business tiers, or local models. AI is worth using. It's just worth using with your eyes open.