Ask a room of ten people whether they’d paste a client contract into an AI chatbot and half will shrug. Ask the same room about Claude and you’ll usually get a pause. Anthropic has spent years building a reputation as the careful one, and that reputation does a lot of quiet work in people’s heads. It’s worth checking whether the reality matches the vibe.
Here’s what actually happens to a Claude conversation, which switches change the picture, and where the gaps sit that no setting will close for you.
What Happens to a Claude Conversation After You Hit Send
Every message you type is stored on Anthropic’s servers. That’s true of every hosted AI assistant, and pretending otherwise would be silly. The interesting questions are how long it stays, who can look at it, and whether it ends up shaping future models.
Retention: the 30-day rule and its exceptions
On consumer accounts, deleting a conversation pulls it from your history immediately and Anthropic says it purges the data from its backend systems within 30 days. Two exceptions matter. Conversations flagged by automated safety systems can be kept for up to two years so the trust and safety team can investigate. And if you delete your whole account, the same 30-day window applies to the bulk of the data, with the same flagged-content carve-out.
Business and API customers get a similar default 30-day retention window, but with more room to negotiate. Anthropic offers zero data retention arrangements for eligible enterprise customers, which is the arrangement regulated industries tend to ask for first.
Human review is real, and it’s narrow
Anthropic employees and contracted reviewers do read some conversations. The published policy is that this happens for flagged content, abuse investigations, and a small sampled subset used to improve safety systems. If your chat is never flagged and never sampled, the odds of a person reading it are low but not zero. If you paste something you’d be uncomfortable seeing in a screenshot on a colleague’s monitor, assume there’s a nonzero chance.
The training question changed in late 2025
This is the part most people get wrong. Anthropic updated its consumer terms so that chats and code sessions can be used to improve models unless you opt out, and the choice applies going forward rather than retroactively. Opting out does not un-train a model that already learned from earlier data. Nobody can do that, for any chatbot, so treat the setting as a forward-looking brake rather than a delete button.
The toggle lives in Settings under Privacy. Find it, read the current wording, and decide. It takes about ninety seconds and most people never do it.
Settings Worth Touching Today
None of these are hidden, but they’re scattered enough that few users find all of them. Working through the list once is a decent afternoon’s investment.
- Model improvement setting. Turn it off if you’re using Claude for work involving client names, internal strategy, or unreleased code.
- Incognito chats. Rolled out in 2025, these stay out of your history and out of training, though Anthropic still holds them for roughly 30 days for safety review. Useful for sensitive one-offs, not for a whole workflow.
- Bulk deletion. Old conversations from 2024 are still sitting there. Clear out the ones you no longer need, since deleted data has a shorter life than stored data.
- Connectors and integrations. Every connected app or MCP server you’ve authorised widens the surface area. Revoke anything you set up once and forgot.
- Two-factor authentication. Obvious, still skipped by a surprising number of people, and the single biggest protection against someone else reading your chat history.
While you’re in the settings screens, it’s worth remembering that Claude’s strengths and weaknesses shape what you’ll naturally end up pasting into it. A tool that’s excellent at careful document editing invites you to hand over real documents, which is exactly the behaviour this article is asking you to think about.
Where Claude Privacy Gets Messy
Connectors hand your data to someone else
Connect Claude to Google Drive, a project management tool, or a custom MCP server, and the request passes through infrastructure that isn’t Anthropic’s. The third party’s privacy policy applies to whatever it sees. Anthropic’s policy does not follow your data into a partner’s systems, no matter how reassuring the settings page looks.
Long context encourages big pastes
Claude’s 200,000-token window is genuinely useful, and it’s also a temptation. People drop entire contracts, full customer databases in CSV form, and months of support tickets into a single conversation because the model can handle it. The privacy risk scales with what you paste, not with how the tool behaves. If you’re using that capacity for real work, it’s worth understanding what 200,000 tokens actually lets you do before you make it a habit with client material.
Support tickets and shared links
When you contact support, you’re handing over a real conversation to a real person. That’s fine and often necessary. It stops being fine when the pasted transcript contains an API key, a patient reference number, or a salary figure you didn’t notice on page four.
Business Accounts Follow a Different Rulebook
If you’re on Claude for Work, Team, or Enterprise, or you’re calling the API, the commercial terms replace the consumer ones. Anthropic states that inputs and outputs from commercial products aren’t used to train models by default. Admins get SSO, domain capture, audit logs, and the ability to set retention centrally. The zero-retention option exists here too.
This is the single biggest practical split. A team member using a personal Pro account for the same work is operating under consumer terms, with consumer training defaults, and no admin visibility. Shadow usage is a policy problem dressed up as a convenience problem. It’s the same pattern you’ll see if you compare how different assistants handle workplace data, which is the thread running through Claude vs Gemini and how each one fits real working habits.
Europe, GDPR, and the Paperwork Nobody Reads
Anthropic offers a data processing addendum and publishes a sub-processor list, which is the minimum any EU buyer will ask for. Data residency options are more limited than what the hyperscalers offer. If that’s a blocker, there’s a workaround: Claude models are available through Amazon Bedrock and Google Vertex AI, so processing can sit inside a cloud agreement and region you already have in place. Legal teams tend to find that far easier to sign than a new vendor relationship.
One detail people miss: your prompts are only part of the data footprint. Account emails, payment records, device and browser information, and usage metadata all sit under the same policy and the same retention schedules.
Habits That Beat Any Setting
Configuration only goes so far. Most real leaks happen because someone was in a hurry.
- Strip identifiers before you paste. “Acme Corp, contact Jane, £240k” becomes “a mid-sized manufacturer, procurement lead, quarter-million budget.” The answer barely changes.
- Rotate any credential that touches an AI tool. If a key went in, it’s compromised, and rotating takes two minutes.
- Keep credentials and personal data in a separate document, then paste around them rather than through them.
- Treat every conversation like an email to a vendor you don’t fully trust. That mental model gets you most of the way.
If your team is still deciding which assistant deserves the access, a broader AI chatbot comparison across the main options is a better starting point than picking the one with the nicest interface and sorting out data terms later.
The One-Question Test Before You Paste
Before you drop anything into Claude, ask a single question: would I be comfortable if this exact text appeared in a support ticket screenshot six months from now?
If the answer is yes, paste freely. Claude’s default posture is more conservative than most of its competitors, and for ordinary work, drafts, research, and code review, the risk is genuinely modest. If the answer is no, or if you hesitated for more than a second, that hesitation is the useful signal. Redact it, summarise it, or use an enterprise account with a signed agreement behind it. The tool will still help. It just won’t be holding a copy of something you’d rather it didn’t.