In August 2025, Anthropic told its consumer users something that quietly rewrote the deal: unless they opted out by a deadline, their Claude conversations would be used to train future models, with retention extended to five years. OpenAI and Google made comparable moves the same season. Reporting at the time noted the pattern clearly: the major providers had shifted consumer products toward training-by-default, while leaving business and API customers untouched.
That last clause is the entire story. What happens to text you paste into an AI tool depends far less on which company you chose than on which *door* you walked in through: consumer, business, or API. Most people never learn there are different doors.
Does AI Training Use the Data You Paste In?
On consumer plans of major AI tools, yes, often by default: your conversations may be retained and used to train future models unless you find the setting and opt out. On business, enterprise, and API plans, the default is the reverse: providers state they don’t train on your data. The same model, the same chat window, radically different data handling, decided by your account type.
Anthropic’s own announcement spells out the consumer side: with training enabled, retention extends to five years; opting out returns you to a 30-day retention window; deleted conversations aren’t used for future training. TechCrunch’s coverage of the change confirms the split that matters: the policy applies to Free, Pro, and Max consumer accounts, while business, government, education, and API customers are unaffected.
The Three Doors, Compared
| Door | Typical training default | Typical retention | Who it’s for |
|---|---|---|---|
| Consumer (free and personal paid plans) | Training on unless you opt out | 30 days if opted out; years if opted in | Individuals |
| Business / Enterprise | No training on your data | Contract-defined, admin-controlled | Teams under commercial terms |
| API | No training on your data | Short, sometimes days; zero-retention arrangements exist for qualifying customers | Developers and products built on the model |
Anthropic’s API data retention documentation illustrates how different the API side is: conversation content isn’t retained by default in most cases, retained data is never used for training without express permission, and zero-data-retention arrangements exist. The consumer product and the API are, for privacy purposes, almost different products.
Exact defaults and windows shift, and they shifted substantially in 2025, so treat the table as the shape of the system and verify your specific provider’s current policy page before pasting anything sensitive.
What “Deleted” Doesn’t Always Mean
Deleting a chat removes it from your history and, under normal policy, starts a clock on back-end deletion. Two caveats deserve to be widely known.
First, training is a one-way door. If a conversation was already used to train a model before you deleted it or opted out, that training isn’t reversed. Opting out stops future use; it doesn’t unlearn the past.
Second, legal process can override policy. During litigation with news publishers, OpenAI was ordered by a court to preserve consumer ChatGPT conversations, including deleted ones, a demand the company publicly fought as conflicting with its privacy commitments. The lesson isn’t about one company. It’s that any provider’s deletion promise operates inside a legal system that can suspend it.
A Practical Rule Set for What to Paste
- Assume consumer chats are the least protected tier you have access to, and check your training toggle today; on several major products the default became “on” in 2025.
- Anything you’d redact in an email to a stranger, redact before pasting: names, account numbers, credentials, health details, unreleased financials.
- Company confidential material belongs behind business or API terms, where no-training defaults and admin controls apply, not in a personal account, even a paid one.
- Client and third-party data is a legal question, not just a privacy one: pasting someone else’s personal data into a consumer tool may violate obligations you hold regardless of the provider’s policy.
- If a tool can act on outside content, remember data can leave through behavior too: hidden instructions in a document can attempt to redirect a connected tool, the risk covered in our explainer on prompt injection.
The Free-Tier Connection
None of this is hidden malice; it’s the business model of free. Consumer tiers, especially free ones, are where providers gather the conversational data that improves future models, which is a large part of what you’re trading for the price of zero. That trade can be perfectly reasonable for low-stakes use. What it shouldn’t be is invisible, and it’s one of the real costs weighed in our companion guide to free vs. paid AI tools.
For contrast, we publish exactly how AI is and isn’t used in our own work in our AI editorial policy, because the standard we apply to vendors should be one we can meet ourselves.
Frequently Asked Questions
Is it safe to paste work documents into ChatGPT or Claude? On a personal consumer account, treat it as unsafe by default for anything confidential: consumer tiers may retain and train on conversations unless you’ve opted out. Under business, enterprise, or API terms, providers state customer data isn’t used for training, which is why organizations route work use through those tiers.
If I opt out of training, is my data gone? Opting out stops future conversations from being used and typically shortens retention. It doesn’t reverse training that already happened, and deleted data can still persist briefly in back-end systems, or longer under legal holds.
Do paid personal plans protect data better than free ones? Usually not in the way people assume. On several major products, paid consumer plans fall under the same training defaults as free ones. The meaningful line is consumer versus business or API terms, not free versus paid.
Can AI companies read my conversations? Policies generally permit access in limited cases: safety review of flagged content, legal compliance, and debugging. Routine human reading of ordinary chats isn’t the model, but the possibility is why credentials and secrets should never be pasted anywhere.
Which tier should a small business use? The smallest business tier the provider offers, or the API, rather than personal accounts. The no-training default and administrative controls are the point, and the price difference is small compared to the exposure of routing company data through consumer terms.
The One Check Worth Doing Today
Open the privacy settings of whichever AI tool you use most and look for the training toggle. Whatever you find, you’ll know which side of the line you’re on, which is more than most users of these tools can say. Then match the door to the data: consumer accounts for public-shaped questions, business or API terms for anything you’re paid to keep confidential. Our ongoing coverage of these policies lives under AI tools.