AI and confidential data: privacy, ownership, and the rules
With AI and confidential data, the account matters more than the tool: on personal plans, ChatGPT, Gemini, and Perplexity train on your conversations by default and people may review them, while business plans do not train by default and come with a data processing agreement. On top of that come privacy laws, professional rules, copyright, and, for European clients, the EU AI Act.
What happens to the data you give an AI assistant
Before anyone on your team pastes a client email into a chatbot, answer four questions: does the company train on it, who can read it, what does the law require, and who owns what comes out. They are the core of using AI at work in a small business, and the companies’ own help pages give part of the answer: Google, for one, asks Gemini users not to enter confidential information that they would not want a human reviewer to see.
The answers depend mostly on the type of account. Personal accounts, free or paid, are built for individuals: training is usually on by default, and the company may review conversations for safety and quality. Business plans are built for organizations: no training on your content by default, a contract that makes the AI company your service provider, and an administrator who can see what the team does.
Which account for which data
| Data | Personal account, training off | Business plan with a data agreement |
|---|---|---|
| Public information, general questions | Fine | Fine |
| Your own drafts, without client details | Fine | Fine |
| Client documents, contracts, internal figures | Avoid | Fine if your professional rules and your client allow it |
| Personal data about clients or staff | Avoid | Only with a lawful basis, an updated privacy notice, and care with transfers |
| Health data, passwords, bank details, ID numbers | Never | Only in a tool specifically approved for it |
This is our recommendation, based on the companies’ terms and regulators’ guidance, not a legal rule. Professionals with confidentiality duties, such as lawyers and accountants, should also follow their regulator.
Four facts that surprise most people
- Turning off training does not cover everything. OpenAI says a thumbs up or down can make a whole conversation available for training, even with training off.
- Temporary chats are still kept for a while: up to 30 days at OpenAI and Anthropic, 72 hours at Google, and 24 hours at Perplexity.
- Deleted is not always gone for your employer. Google Vault and Microsoft Purview can retain conversations that users have deleted, depending on the plan.
- Paying is not the same as protecting. The UK’s Solicitors Regulation Authority warns that both free and paid AI tools may pose risks to client confidentiality.
What changed in 2026
Advice written a year ago is already out of date in several places:
- UK: the Data (Use and Access) Act 2025 rewrote the rules on automated decisions and international transfers, mostly from February 5, 2026, and since September 30, 2026, UK GDPR refers to the Information Commission, the successor to the ICO.
- EU: Regulation (EU) 2026/1744, in force since July 27, 2026, softened the AI literacy rule and moved the main high-risk rules to December 2, 2027, while the transparency rules have applied since August 2, 2026.
- Australia: from December 10, 2026, covered businesses must describe in their privacy policy the computer programs that make or help make significant decisions about people.
- Canada: the federal privacy reform, bill C-36, was introduced on June 15, 2026 and remains a bill.
- US: the Supreme Court declined on March 2, 2026 to review the case that denied copyright to a purely AI-generated work.
- Microsoft: its new Copilot app, launched on August 18, 2026, stopped using personal conversations to train foundation models.
Training on your conversations
ChatGPT, Gemini, and Perplexity train on personal accounts by default; Claude asks you to choose; Microsoft’s new Copilot app stopped training foundation models on personal conversations in August 2026. Our page on whether the main assistants train on your data gives the switch for each.
Client and confidential information
What the companies advise, what regulators of lawyers and accountants require, and a traffic-light rule for your team: see putting client information into AI.
What an employer can see
Business plans let administrators retrieve conversations, sometimes even deleted ones, and a personal account opened with a work email can end up linked to the employer. See work and personal AI accounts.
Privacy law by country
UK GDPR, the EU GDPR, PIPEDA and Quebec’s Law 25, Australia’s Privacy Act, and US state laws all ask the same five questions in different words. See the privacy rules for AI tools in each country.
A policy for your team
A one-page policy with approved tools, data rules, and a review date, ready to copy: see the AI use policy template.
Ownership and disclosure
Purely AI-generated material is not protected by copyright in the US, while the UK still protects computer-generated works for now. See who owns AI-generated content, and when you must say you used AI.
The EU AI Act
If your work reaches people in the EU, the AI Act applies even if you are based elsewhere, with light duties for businesses that only use AI tools. See the AI Act for small businesses.
Once the rules are set
With the right account and a page of rules, the next step is using AI for everyday tasks: emails, summaries, spreadsheets, and checking the answers. The same rules apply to recording and summarizing meetings and to automation tools for a small team, which reach further into your data. To compare assistants and plans first, see comparing ChatGPT, Claude, and Gemini and paying for an AI plan.