Does Claude Train on Your Chats? One Switch Decides

Short answer: on Claude Free, Pro and Max, your chats are used to train future models if the “Help improve our AI models” setting is on for your account, and it is one of the first things a new account is asked about. That same switch also decides how long Anthropic keeps your conversations: off means the standard 30-day window, on means Anthropic may hold your data in a de-identified format for up to 5 years in its model training pipelines. Everything below comes from Anthropic’s and OpenAI’s own help pages, Anthropic’s announcement and the press coverage of the change, re-checked on 3 October 2026.

The short answer, by plan

If you moved to Claude because it felt like the careful option, the honest news is that the consumer plans work much the way ChatGPT’s personal plans do. Anthropic’s August 2025 update to the Consumer Terms was explicit about who it covers: Claude Free, Pro and Max, including Claude Code when you run it from one of those accounts. It equally explicitly does not cover the commercial side, which is Claude for Work, Claude for Education, Claude for Government and API use including Amazon Bedrock and Google Cloud Vertex AI.

So the question “does Claude train on my conversations” has no single answer. It has an account type and a toggle.

The default flipped

Before that update, ordinary consumer chats were not feeding model training. Feedback you deliberately submitted on a response, and conversations flagged for safety, were already the exception, and still are. The update changed the shape of the deal: Anthropic asked every existing user to make a choice by 8 October 2025, and made the same choice part of signup for new accounts. Saying yes puts your chats and coding sessions into training data. Saying no keeps the older arrangement.

The part worth knowing is the framing. TechCrunch, reporting on the rollout in August 2025, described the pop-up that existing users met: “Updates to Consumer Terms and Policies” in large text, a prominent black Accept button, and below it, in smaller print, a toggle for training permission already set to on. The Verge, cited in the same piece, made the obvious point about that layout, which is that people click Accept without noticing what the toggle underneath it does. Nothing happened without a click, and the toggle was there on screen. But a toggle that arrives pre-flipped is a default in every sense that matters, and it is easy to accept one in half a second without ever reading the word “training”.

That is why this article exists. If you clicked Accept on that screen without reading it, you may have said yes without ever deciding to.

Retention: the 5-year window

This part is worth looking at separately from training.

Anthropic’s privacy centre states that if you allow your chats or coding sessions to be used to improve Claude, it may retain that data in a de-identified format for up to 5 years in its model training pipelines. Decline, and the standard 30-day retention continues. Same product, same price, two retention regimes, decided by one switch.

Two clarifications worth getting right:

• The 5 years applies going forward, not backwards. Anthropic says the extended window covers new or resumed conversations after the setting is turned on. Old chats you never touched again are not swept in.
• Deleting a chat still works. Deleted conversations leave your history immediately and Anthropic’s back-end storage within 30 days, and are not used for future training. What deletion cannot do is pull your words back out of a training run that already started.

There are separate clocks that no toggle touches, and you should know they exist: conversations flagged for a usage policy violation can be kept for up to 2 years, with the safety scores kept up to 7 years, and any feedback you submit on a response is held for up to 5 years. Those clocks are why “I deleted it” is never the whole story, with Anthropic or with OpenAI. The same principle we cover in what you should never paste into ChatGPT applies here without a single change.

Plan by plan

Verdicts here are about how exposed the plan is to model training, not about whether Claude is good. They mean what they mean everywhere on Skuto: hexagon ⬣ Risk, triangle ▲ Caution, circle ● Safe.

Paying more does not buy you out. Max sits under exactly the same consumer terms as Free. The line that matters is consumer versus commercial, not cheap versus expensive, which is the same shape of answer we reached in the ChatGPT document upload question.

Note the Claude Code row properly. It carries the same verdict as Free, Pro and Max because it is the same account and the same switch. Run it from a Pro or Max login with the setting on, and your repositories, your stack traces and whatever your codebase contains fall under the consumer terms, with the same retention. Check the setting before you open your employer’s private code in it. What does not train: the Claude Code documentation says Anthropic does not train its models on code or prompts sent under commercial terms, such as Team and Enterprise plans and the API, unless the customer has chosen to share that data. For the tool itself, see our take on whether Claude Code earns its keep.

How to turn it off

Sixty seconds, on the web:

1. Click your name to open the menu.
2. Open Settings, then Privacy. The direct address is claude.ai/settings/data-privacy-controls.
3. Switch off Help improve our AI models.

On mobile the path is identical: your name, Settings, Privacy, toggle.

Three honest caveats. First, off means off from now on. Anthropic is clear that your data stays in training runs that have already begun and in models already trained. Second, safety-flagged conversations can still be used for trust and safety work regardless of the setting. Third, menus move. If the label has changed by the time you read this, look for anything with “improve” or “model training” in it, and tell us so we can fix the page.

There is also a per-conversation option. Incognito chats are never used for model improvement regardless of your account setting, which makes them the right home for a one-off you would rather not file anywhere. It is the same instinct as a temporary chat in ChatGPT.

Where Claude is better, and where it is not

Better: the choice is a genuine, reversible switch on a settings page you can find, and the commercial tiers are cleanly walled off in writing rather than in marketing copy. The privacy centre documents the retention clocks in plain numbers, including the unflattering ones. Not every vendor puts its retention clocks on one page.

Not better: the consumer default is not more protective than OpenAI’s. OpenAI may train on what you type in ChatGPT’s personal plans unless you turn “Improve the model for everyone” off, and OpenAI’s business tiers are excluded by default too. On data retention after deletion the two sit level, at roughly 30 days to clear back-end storage either side.

The 5-year figure needs care, because it is easy to swing the wrong way with it. It looks like proof that Claude holds your words longer than ChatGPT does. It is not. It is a ceiling Anthropic publishes and OpenAI does not. On the switch itself, Anthropic actually promises more: turn training off and it stops using your older stored chats in future training runs too, while OpenAI’s pages only promise that new conversations are left out, and put no clock on data it has already used. Both retention pages also keep the same kind of escape hatch: OpenAI exempts deleted chats “already de-identified and disassociated from your account” from its 30-day deletion, and Anthropic says de-identified data may be kept longer for research. So the honest comparison is a disclosed five years against an unstated number. Anthropic told you where the wall is. That is not the same as the wall being further away.

So the reputation is half earned, though not for the reason the number suggests. Anthropic gives you a clear switch, and it is frank about the clocks running behind it. What it does not give you is a consumer default that protects you more than the alternative does. If you switched to Claude for privacy and never opened Settings, you did not buy protection. You bought a company that is willing to tell you what it kept.

The habit that outlasts any setting

Every toggle on this page is a promise from a company about data it already has. The only privacy that no policy change can revoke is the personal data that never left your machine.

That is what the paste checker is for. Drop in the text before it goes to Claude and it flags names, card numbers and IBANs, API keys, lines like “password: …” and other things you probably did not mean to hand over. It runs entirely in the page you already have open, so checking never creates a second copy anywhere, including with us.

Turn the switch off. Then get in the habit of not needing it to be off.

Keep reading

GuidesCan IT See Your ChatGPT on a Work Laptop? Even When You're Logged OutOn a company laptop, IT can see ChatGPT use even when you're logged out, on a personal account or in incognito, sometimes uploads too. What each layer sees.GuidesCan Your Employer See Your ChatGPT? What's Visible and What Isn'tPersonal account on your own device and network: no. Company workspace: admins may be able to read and export your chats. A work laptop can be monitored anyway.GuidesIs It Safe to Upload Confidential Documents to ChatGPT?Not as-is on a personal plan: an uploaded file can be used for training unless you turn that off. Business and Enterprise don't train on uploads by default.