Do ChatGPT, Gemini or Copilot train on your company's data? Plan by plan

When someone on your team pastes a contract, a customer record or a report into an AI assistant, hardly anyone asks what happens to that text afterwards. It depends above all on one thing: the plan they are using the tool on.

On the personal versions of ChatGPT, Gemini, Microsoft Copilot and Claude (the free ones and the individual paid ones too), what you type can be used to train the vendor's models unless a setting is switched off. On business plans and API access, all four vendors say they do not train on your data by default. DeepSeek trains by default and stores data in China. We checked each vendor's official pages on 9 October 2026.

The table at a glance

In ChatGPT, Gemini and Claude, rating a response with a thumbs up or down can put the conversation into training even if you switched the setting off. Checked on the official pages on 9 October 2026.

What "training on your data" means

Training a model means using text so it learns to respond better. If your conversation goes into that set, it is no longer only in your account: it becomes part of the material the vendor uses to improve its product, and some of it may be reviewed by its staff.

Training is one of four things worth checking. The other three are how long the conversation is kept, whether people read it, and whether there is a contract obliging the vendor to process that data on your company's instructions. Personal plans do not come with that contract: the vendor handles the data under its own consumer terms.

Plan by plan

ChatGPT (OpenAI)

• Personal plans (Free, Plus and Pro): OpenAI may use conversations to train its models. You switch it off under Settings > Data controls > Improve the model for everyone, or through its privacy portal. Paying for Plus or Pro does not change this: they are still personal plans.
• Temporary chats: not used for training while they remain temporary.
• Careful with ratings: if you give a response a thumbs up or down, the whole conversation may be used for training even if you switched the setting off.
• Business plans (Business, Enterprise and Edu) and API: by default OpenAI does not use what is typed or what it answers to improve its models, and it signs a Data Processing Addendum for Business, Enterprise and the API.

Gemini (Google)

• Personal Google accounts: with the Keep Activity setting, on by default for users aged 18 and over, activity is used to improve Google's services, including training its generative AI models. Some conversations are reviewed by people, and those are kept for up to three years, no longer linked to your account.
• With the setting off: new conversations are not used for training unless you send feedback, and they are kept with your account for 72 hours.
• Company Google Workspace accounts: Google says it does not use Workspace data to train its generative AI models without permission.

Microsoft Copilot

• Copilot with a personal Microsoft account: Microsoft uses conversation activity, including uploaded files and images, to train its models unless the user opts out in Copilot's privacy settings. Microsoft excludes some countries from that training, and Spain is not one of them.
• Excluded by Microsoft: work accounts (Entra ID), people using Copilot inside Microsoft 365 Personal or Family apps, people who are not signed in, and users under 18.
• Microsoft 365 Copilot with a work account: prompts, responses and company data are not used to train its foundation models.

Claude (Anthropic)

• Personal plans (Free, Pro and Max): Anthropic uses conversations to improve Claude if the user allows it through the Model Improvement privacy setting. Incognito chats are not used.
• Ratings: if you rate a response, the whole conversation is kept for up to five years and may be used for training.
• Commercial plans (Team, Enterprise and API): its commercial terms state that Anthropic may not train models on customer content, and they incorporate a Data Processing Addendum.

DeepSeek

• Its app and website: DeepSeek uses data to train and improve its models. To object you have to write to its privacy address, as there is no setting in the app.
• Where the data goes: its policy says it collects, processes and stores the data in the People's Republic of China.
• DeepSeek's open-weight models work differently: if they run on your servers or with a European provider, DeepSeek sees no data at all.

What we recommend to a company

1. No personal accounts for work involving customer, employee or contract data. Neither free nor individual paid ones.
2. If the team uses an assistant, it should be on the business plan and with a data processing agreement. The GDPR (Article 28) requires a contract with anyone processing personal data on your company's behalf.
3. On any plan, check the training setting and avoid ratings when the conversation contains sensitive data.
4. Write down what can and cannot be pasted, with examples from your sector. The real problem is usually use nobody has authorised: we covered it in shadow AI in 2026.
5. For processes with sensitive data (health, case files, payroll), an application of your own, with the data in the European Union (in Spanish) and no third-party training. We explain it in secure AI for companies (in Spanish).

The terms change: how to check them

These terms change several times a year, sometimes by a single sentence. We read them on the official pages linked at the end. There is also an independent record, Trained On, that logs each change in the fine print of the main tools with its date. Before deciding, check the current version and repeat the check every few months.

Frequently asked questions

If I pay for ChatGPT Plus, do they stop training on my data?

Not necessarily. Plus and Pro are personal plans, so training depends on the Improve the model for everyone setting. The plans that do not train by default are Business, Enterprise, Edu and the API.

Are temporary or incognito chats safe?

In ChatGPT and Claude they are not used for training while they remain temporary or incognito. Even so, the conversation passes through the vendor's servers and is kept for a while, so they are no substitute for a business plan with a contract.

Is the API safe for personal data?

OpenAI and Anthropic do not train on API data by default, and the contract includes commercial terms and a data processing agreement. What remains to be decided is where the data is processed and how long it is kept, which is set when the application is designed.

What should we do if someone has already pasted customer data into a personal account?

Switch training off on that account, delete the conversations involved and review it with the person responsible for data protection in your company, who will decide whether anything else needs doing.

What about open-weight models installed on our own servers?

If the model runs on your servers or with a provider you contract in the European Union, the model's maker does not receive your data. It is the most private option, at the cost of maintaining the infrastructure.

If you want to use AI with your company's data without depending on the terms of a personal account, tell us which process you have in mind.

This article is for information and is not legal advice. Tecnea develops private AI applications, so we have an interest in companies not working with personal accounts. We do not recommend any of these tools. Each vendor's terms were checked on their official pages on 9 October 2026 and may change.

Sources

1. OpenAI, "How your data is used to improve model performance" (accessed 9 October 2026): help.openai.com
2. OpenAI, "Enterprise privacy at OpenAI" (accessed 9 October 2026): openai.com
3. Google, "Gemini Apps Privacy Hub" (updated 24 September 2026): support.google.com
4. Google, Gemini help on Gemini Apps activity: support.google.com
5. Google, "Generative AI in Google Workspace Privacy Hub": support.google.com
6. Microsoft, "Privacy FAQ for Microsoft Copilot" (accessed 9 October 2026): support.microsoft.com
7. Microsoft, "Enterprise data protection in Microsoft 365 Copilot": learn.microsoft.com
8. Anthropic, "Is my data used for model training?" (updated 16 March 2026): privacy.claude.com
9. Anthropic, "Commercial Terms of Service" (effective 17 June 2025): anthropic.com
10. DeepSeek, "Privacy Policy" (updated 10 February 2026): cdn.deepseek.com
11. Trained On, a record of AI tools' training terms: trainedon.me
12. Regulation (EU) 2016/679 (GDPR), Article 28: eur-lex.europa.eu