Does your AI assistant train on your business chats? It depends on the login, not the brand
Does your AI assistant train on your business chats? For most small businesses the honest answer is "some of them, yes", and the deciding factor is not which brand you picked but which login your staff are typing into. A Hacker News thread on 4 November drew more than 700 votes by pointing out that a paid personal Gemini subscription does nothing to stop Google using saved chats to improve its models. Google's own Gemini Apps Privacy Hub, updated in October, confirms it: with the default setting on, your activity is used to provide, develop and improve Google services, including training its generative AI models.
The thread was angry at Google, but the useful lesson is wider than one vendor. ChatGPT, Gemini and Claude all run the same split: consumer accounts are where training happens, business accounts are where it doesn't. Paying for a consumer plan buys a better model and higher limits. It does not buy a different data contract.
Does your AI assistant train on your business chats?
It depends on the account type, and the three big vendors are more alike than their reputations suggest.
- Gemini. On a personal Google account, "Keep Activity" is on by default for adults. While it is on, chats and uploads can be used to train Google's models, a sample is read by human reviewers (including contractors, with the chat disconnected from your account first), and a reviewed chat can be kept for up to three years. A Google AI Pro or Ultra subscription on a personal account gets no separate rule. A Google Workspace account does: the Workspace privacy hub states customer data, prompts included, is not used to train models without the customer's permission, and is not human-reviewed outside your domain.
- ChatGPT. OpenAI's data controls page says "Improve the model for everyone" is on by default for Free, Plus and Pro personal accounts. For ChatGPT Business, Enterprise, Edu and the API it is off by default.
- Claude. Anthropic took a different route in August 2025: Free, Pro and Max users had to actively choose, by 8 October 2025, whether their chats and Claude Code sessions could be used for training. Say yes and retention stretches to five years; say no and it stays at 30 days. Team, Enterprise and the API are not used for training by default.
Two exceptions apply on every tier of every vendor. Pressing thumbs up or thumbs down hands that conversation over as feedback, which all three may use for training and keep for years. And chats flagged for safety review are looked at regardless of your setting. The second one is reasonable; the first one is a button your staff press without thinking.
How do you stop ChatGPT, Gemini or Claude training on your data?
Three options, in order of how well they actually work for a business.
- Use a business account. This is the only answer that does not depend on every employee remembering a toggle. Workspace, ChatGPT Business or Team, Claude Team: in each case training is off at the account level and an administrator can see who is in the tenant. It is also the only route that gives you a contract rather than a setting.
- Flip the toggle. On a personal account you can switch training off: ChatGPT under Settings then Data controls, Claude under Privacy Settings, Gemini under Activity. The trade-offs differ. ChatGPT and Claude keep your history. Gemini does not: turning off Keep Activity means future chats vanish after 72 hours, and most connected apps, Gmail and Drive among them, stop working. That is the complaint at the heart of the Hacker News thread, and it is accurate. For a Google shop the toggle is not a real option; a Workspace account is.
- Use the disposable mode. Temporary Chat in ChatGPT, temporary chats in Gemini and incognito chats in Claude are all excluded from training. Fine for a one-off question, useless as a policy, because nobody uses them for the Friday-afternoon paste of a client's spreadsheet.
What we would do in a Melbourne business
The exposure we see most often is not a company that chose the wrong vendor. It is a company with a Workspace or Microsoft 365 tenant, properly contracted, while half the team quietly uses a personal ChatGPT Plus paid on their own card because it was faster to set up. Every chat in those personal accounts sits under the consumer rules above, and the business has no visibility of any of it.
So start with an inventory, not a policy. Ask who is using what, on which login, and whether the subscription shows up on a company invoice. Then make the business tier the easy path: one admin-created seat per person usually costs less than the personal plans you are already reimbursing. Switch off the rating buttons where the admin console allows it. And treat the Office of the Australian Information Commissioner's guidance on commercial AI products as the floor: its best-practice line is not to enter personal information, particularly sensitive information, into publicly available generative AI tools at all, and to check the terms and settings of any tool before customer data goes in.
Training is not the only question, either. "Not used for training" says nothing about how long a vendor keeps your prompts, where they are stored, or who can be compelled to hand them over. Those answers also come from the account type, and they are the second conversation to have once the first is sorted. If you want help working out which of your tools are on which contract, that audit is the opening session of our AI training for teams, and it usually turns up the same gaps as getting your business data organised in the first place.
Source: Google — Gemini Apps Privacy Hub