# Is it safe to use AI chatbots with personal data? What the apps keep

> We would treat an AI chatbot as fine for everyday questions and off limits for anything confidential. Before you paste anything personal, check three things: whether your chats can train future models, how long deleted chats stay, and whether the account is personal or from work. Never paste passwords, card numbers, health records or confidential files into a consumer chat.

Mindket editors · Published 2026-11-04 · Updated 2026-11-05 · https://mindket.com/learn/ai-basics/ai-privacy

## In short

- Consumer chat apps can use your chats to improve models unless you turn the setting off or pick an incognito or temporary mode, and each vendor words its promise differently.
- Turning training off does not delete anything. Data already used in training stays, and deleted chats leave the back end later, not at once.
- Work and team plans follow separate terms, and an administrator may still be able to search the chats on a work account.
- Some chats are reviewed by people or flagged by safety systems, so keep passwords, card numbers, health records and confidential files out of them.

## Before you start

You need to know which assistant you use and whether you signed in with a personal or a work account. Keep its settings open in a second tab. Every fact below was read on the vendors' own privacy pages on 4 November 2026, and these pages change, so look for the update date on yours.

## Audit your chat app in six steps

1. Note the account type. Anthropic's [privacy page](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training) covers Free, Pro and Max accounts and Claude Code used on them, and it sends commercial products such as Claude for Work and the API to a separate article.
2. Find the training setting. On Claude, open Settings, then Privacy, then the toggle labelled Help Improve our AI models, as Anthropic's [steps page](https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings) shows. In Gemini Apps the control is Keep Activity.
3. Decide and switch it. Turning the Claude setting off keeps new chats and coding sessions out of future training. Data already used in training that has started, or in models already trained, is not removed.
4. For a one-off sensitive question, open a temporary or incognito chat. In Claude, [incognito chats](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training) are not used to improve the model even with the setting on, and Google's temporary chats are not used to train its models.
5. Set a retention you can live with. Google's [Privacy Hub](https://support.google.com/gemini/answer/13594961) says Gemini Apps auto-delete defaults to 18 months and can be changed to 3 or 36 months, or set to indefinite.
6. Delete chats you no longer need. A deleted Claude conversation leaves your chat history immediately and is removed from Anthropic's back-end storage within 30 days.

## What the vendors' pages say they keep

The table puts five setups side by side. The ChatGPT app is missing on purpose. OpenAI's help center, where the app's data controls are described, did not open for us that day, so only OpenAI's [developer docs](https://developers.openai.com/api/docs/guides/your-data) on API data appear here.

| App and plan | Used to train models? | How long data stays | Source |
| --- | --- | --- | --- |
| Claude, personal plans | Only if you allow it, and never for incognito chats | Deleted chats leave the back end within 30 days. With the setting on, de-identified chats may be kept up to 5 years | Anthropic Privacy Center |
| Claude, work and API plans | Not by default, except feedback you submit | Set in a separate article for organizations | Anthropic Privacy Center |
| Gemini Apps | With Keep Activity on, yes. With it off, or in a temporary chat, no, unless you send feedback | Auto-delete after 18 months by default. Temporary chats and Keep Activity off chats stay 72 hours | Google Gemini Privacy Hub |
| Microsoft Copilot at work | No for prompts, responses and data read through Microsoft Graph | Kept with your organization's Microsoft 365 content, with retention set by admins | Microsoft Learn |
| OpenAI API, not the ChatGPT app | Not by default since March 1, 2023 | Abuse monitoring logs for up to 30 days by default | OpenAI developer docs |

## How long chats stay

Four figures sum up the retention pages. Anthropic's [retention page](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data), updated on 1 July 2026, also says that when automated systems flag a chat for a usage policy violation, its inputs and outputs are kept up to 2 years. Anthropic's [training page](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training) adds that feedback, such as a thumbs up or down, keeps the related conversation for up to 5 years and is de-linked from your user ID before Anthropic uses it.

So a switched-off training setting is no promise that nobody ever reads a chat. Chats flagged by safety classifiers may still be used for trust and safety work on Anthropic's side. A subset of Gemini chats goes to human reviewers at Google, and the reviewed chats are kept up to three years, disconnected from your account.

**Four retention figures, as the vendors state them**

- **30 days** for Anthropic to remove a deleted Claude chat from its back-end storage (Source: [Anthropic Privacy Center: How long do you store my data?](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data), read 2026-11-04)
- **72 hours** that Google keeps temporary Gemini chats and chats made with Keep Activity off (Source: [Google Gemini Apps Privacy Hub](https://support.google.com/gemini/answer/13594961), read 2026-11-04)
- **18 months** default auto-delete period for Gemini Apps activity (Source: [Google Gemini Apps Privacy Hub](https://support.google.com/gemini/answer/13594961), read 2026-11-04)
- **5 years** longest Anthropic keeps de-identified chats in training pipelines when the setting is on (Source: [Anthropic Privacy Center: How long do you store my data?](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data), read 2026-11-04)

## Work and team plans are a different deal

On the commercial side, Anthropic says it will not use inputs or outputs from its commercial products for training by default, with an exception for feedback you submit. Its [commercial privacy article](https://privacy.claude.com/en/articles/7996868) was updated on 18 August 2026, and Claude's [plan page](https://claude.com/pricing) lists model training as off by default on Team.

Microsoft's [privacy page](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) says Copilot prompts, responses and data read through Microsoft Graph are not used to train foundation models. Its [overview](https://learn.microsoft.com/en-us/copilot/plan/copilot-overview) adds that, for people signed in with a Microsoft Entra work account, prompts and responses get the same contractual protection as email in Exchange and files in SharePoint.

That protects your data from training, and it does not hide a chat from your employer. Admins can view the stored interaction data with Content search or Microsoft Purview and can set retention policies for it. Copilot also only surfaces files that you already have permission to open, so messy sharing rules at work become a privacy problem too.

## What never to paste

Anthropic's [data privacy FAQ](https://support.claude.com/en/articles/8325621-claude-data-privacy-faq) asks for caution with four kinds of content. They are financial information such as Social Security numbers, credit card numbers and bank account details, health records, passwords and private login credentials, and confidential business or personal documents. Microsoft's guidance for teens adds that personal details about other people count too.

> Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services
>
> Google, Gemini Apps Privacy Hub, [Google Gemini Apps Privacy Hub](https://support.google.com/gemini/answer/13594961) (read 2026-11-04)

- Passwords, login details and recovery codes.
- Card numbers, bank details and government ID numbers.
- Health records and medical details, yours or anyone else's.
- Contracts, source code and customer lists that belong to an employer, unless your employer has approved the tool.
- A real name with an address and phone number. Replace the details with placeholders such as NAME and ADDRESS and add them back yourself.

## Ethan Mollick on why the training switch is not a privacy switch

Ethan Mollick, a Wharton professor who studies AI, [separated two questions on X on 5 September 2026](https://x.com/emollick/status/2096333750594343208). Keeping your data out of training is a setting, and he names OpenAI and Claude as places where you can turn it off. Privacy, he wrote, is a separate question.

The pages in this lesson show why the two come apart. With training off, a deleted chat still sits in back-end storage for a while, a chat flagged by safety systems is kept longer, and Gemini chats a reviewer has read stay for years. So switch training off as step 3 says, then decide what to paste by the list above rather than by the switch.

> you can make sure your data is not used for training by turning that off in OpenAI and Claude.
>
> Ethan Mollick, Professor at Wharton who studies AI, [X](https://x.com/emollick/status/2096333750594343208) (read 2026-11-05)

[![Ethan Mollick (@emollick) on X, 5 September 2026, in a reply to his own post: you can make sure your data is not used for training by turning that off in OpenAI and Claude.](https://mindket.com/blog/shots/x-emollick-data-not-used-for-training.webp?v=15cd774e2c0a)](https://x.com/emollick/status/2096333750594343208)

Ethan Mollick on X, 5 September 2026

> A screenshot of X's own embed of the post, taken on 5 November 2026, with the time in UTC. Select it to open the post on X.

## Common mistakes

- Reading "training off" as "deleted". The setting only changes what happens to future training, and past training stays.
- Assuming a temporary chat leaves no trace. Google keeps temporary chats for a short period to respond to you and to protect Google, its users and the public.
- Using a personal account for work files. Personal and work plans follow different terms, and only the work one has administrator controls.
- Trusting a privacy page you read last year. Anthropic's retention page was updated on 1 July 2026 and Google's Privacy Hub on 26 October 2026.
- Forgetting connectors. Per Anthropic's training page, training data does not include raw content from connectors such as Google Drive, MCP servers included, though text copied directly into a chat can be included. Our post on whether [MCP servers are safe](https://mindket.com/blog/are-mcp-servers-safe) covers the other side of that link.

## Try it yourself

- Open the privacy settings of the assistant you use most and write down two lines: whether training is on or off, and how long chats stay.
- Scroll through your last ten chats and mark anything you would not want a reviewer to read. Next time, swap those details for placeholders.
- Find the update date on your vendor's privacy page and put a reminder in your calendar to read it again in three months.

## Next course: Learn with AI

This course covered how models work, how to choose one, how to check an answer and what to share. The next course, [Learn with AI](https://mindket.com/learn/learn-with-ai/ai-study-mode), puts a model to work as a study partner: study modes, tutor prompts, quizzes and language practice.

If a claim in a reply worries you, [the hallucinations lesson](https://mindket.com/learn/ai-basics/ai-hallucinations) has the checking steps.

## Check yourself

1. You turn off the model training setting in Claude. What happens to data already used in a training run that has started?

   A. It stays in the models already trained
   B. It is deleted at once
   C. It is returned to you
   D. It moves to your incognito chats

   Answer: A. Anthropic says data already included in training that has started, or in models already trained, is not removed.

2. Which kind of chat does Anthropic say is not used to improve Claude even when the training setting is on?

   A. A chat with a file attached
   B. A lengthy chat
   C. A chat started on a phone
   D. An incognito chat

   Answer: D. Anthropic says incognito chats are not used to improve Claude, even if model improvement is enabled in your privacy settings.

3. You have one question that involves sensitive details. What is the best habit?

   A. Paste everything and delete the chat tomorrow
   B. Use a temporary or incognito chat and leave the personal details out
   C. Assume a well-known app is safe
   D. Use a friend's account

   Answer: B. Temporary and incognito chats are not used for training, and details you never type cannot be kept or reviewed.

4. Which of these does Anthropic's FAQ list among things to avoid sharing in a consumer chat?

   A. Recipes
   B. Public news articles
   C. Passwords and private login credentials
   D. Questions about the weather

   Answer: C. The FAQ names financial details, health records, passwords and confidential documents.

## Frequently asked questions

### Is it safe to use ChatGPT with personal data?

This lesson cannot answer for the app, so it does not guess. Open the privacy or data settings in your copy of the app, find whether chats are used for training and how long deleted chats stay, then compare them with the table above. Only the API figures are documented here.

### Can the company read my chats?

A few people can, under conditions each vendor sets. Anthropic limits access for training use to a small number of personnel working on data without your user ID, and Google describes human reviewers who see a subset of chats.

### Do temporary or incognito chats keep nothing?

No. Google's page gives temporary chats a 72-hour life, so closing the window does not erase them at once, while Anthropic's incognito wording is about model improvement. Read your own app's page for what it keeps.

### Does turning off training delete my old chats?

No. Anthropic says turning the setting off excludes your previous and new chats from future training runs. Data already in a training run that has started, or in a model already trained, is not removed. Deleting chats is a separate step.

### Are work plans safer than personal ones?

Safer from training, yes. Safer from your employer, no: admins can search stored Copilot chats, and on a work account the retention choices sit with them rather than with you.

### Can I share a work document with a chatbot?

Ask your IT team first. Anthropic's FAQ lists confidential business documents among the things to be careful with, and a no from IT means pasting a made-up sample of the same shape instead.

## Sources

- [Anthropic Privacy Center: Is my data used for model training?](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training) (read 2026-11-04)
- [Anthropic Privacy Center: How long do you store my data?](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data) (read 2026-11-05)
- [Anthropic Privacy Center: How do I change my model improvement privacy settings?](https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings) (read 2026-11-04)
- [Anthropic Privacy Center: Is my data used for model training? (commercial products)](https://privacy.claude.com/en/articles/7996868) (read 2026-11-04)
- [Anthropic Help Center: Claude data privacy FAQ](https://support.claude.com/en/articles/8325621-claude-data-privacy-faq) (read 2026-11-04)
- [Anthropic: Claude plans](https://claude.com/pricing) (read 2026-11-04)
- [Google Gemini Apps Privacy Hub](https://support.google.com/gemini/answer/13594961) (read 2026-11-05)
- [OpenAI: Your data (API)](https://developers.openai.com/api/docs/guides/your-data) (read 2026-11-04)
- [Microsoft Learn: Data, privacy, and security for Microsoft Copilot](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) (read 2026-11-04)
- [Microsoft Learn: What is Microsoft Copilot?](https://learn.microsoft.com/en-us/copilot/plan/copilot-overview) (read 2026-11-04)
- [Microsoft Learn: Application card, Microsoft Copilot (for organizations)](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-application-card) (read 2026-11-05)
- [Ethan Mollick on X, 5 September 2026](https://x.com/emollick/status/2096333750594343208) (read 2026-11-05)

[Start prompt engineering](https://mindket.com/learn/prompt-engineering/prompt-engineering-basics)