Whether an AI tool trains on what your team types depends less on the brand than on the plan. As of October 2026, OpenAI says it may train on content from individual ChatGPT accounts unless the user opts out, but does not train on business-plan or API data by default. Anthropic asks individual Claude users to choose, while its business plans and API fall under commercial terms. Google says Workspace does not use customer data, including prompts, to train models without the customer's permission. The practical question is which plan each person is actually using.
The same AI brand can treat your data very differently depending on the plan. The question for a small business is not "is ChatGPT safe?" but "which account is each person on, and what do that plan's terms say?"
OpenAI (ChatGPT). For services for individuals, OpenAI says it "may use your content to train our models". Users can opt out in the privacy portal or data controls, and Temporary Chat is not used for training. For business plans and the API, OpenAI does not train on inputs or outputs by default. (OpenAI)
Anthropic (Claude). Since the August 28, 2025 consumer terms update, Claude Free, Pro and Max users choose whether their chats help improve Claude. If they allow it, retention is five years; otherwise it stays at 30 days. Claude for Work (Team, Enterprise), the API and Claude for Education are excluded. (Anthropic)
Google (Gemini in Workspace). Google says Workspace "does not use customer data for training models without customer's prior permission or instruction". Gemini follows existing permissions: if a user can't open a file or email, Gemini won't retrieve it. (Google Workspace Privacy Hub)
For each AI tool: the plan or account type, who uses it, what client data goes in, and the date you last checked the vendor's terms.
The free AI tool list collects this in about 10 minutes, and the AI Use Register Kit turns it into a register and a usage policy.
This brief describes vendor statements. It is not legal advice and does not say whether any tool is appropriate for your data.
More research