Why Does Rating an AI Answer Override Your Privacy Settings?
Rating an AI answer with the thumbs-up 👍 or thumbs-down 👎 button sends that conversation to the company that runs the AI. On a personal ChatGPT, Claude, or Gemini account, this happens even with the training setting turned off. OpenAI, Anthropic, and Google all say so on their own privacy pages.
You probably think you have opted out
Most business owners who take AI seriously have done the sensible thing. They pay for a plan, open the settings, and turn off the option that lets the company train on their conversations. In ChatGPT, it is called "Improve the model for everyone". In Claude, it is "Help improve Claude". In the Gemini app, it is "Keep Activity".
The thumbs buttons sit outside that setting. Every provider treats a rating as a separate, deliberate act. You have chosen to report something, so they collect the material they need to look at it. That material is the conversation, and in one case, it is a whole day of conversations. The switch you turned off does not apply.
In the shadow AI post, the problem was team members using AI tools that the business never approved. This one catches the people who did everything right and still handed over their data one click at a time.
What happens when you click thumbs up or thumbs down?
Each company describes it slightly differently, and the wording matters because these are the companies' own words.
ChatGPT. OpenAI's page on how your data is used to improve model performance explains the opt-out, then adds one line. "If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models." One click on one answer, and the whole chat is in.
Claude. Anthropic's privacy centre lists the two situations in which your data is used for training, even with the training setting turned off. One is a safety review. The other is feedback. In Anthropic's words, "we will store the entire related conversation, including any content, custom styles, conversation preferences, or model settings, in our secured back-end for up to 5 years." Anthropic removes the link to your account first. It still contains whatever you typed.
Gemini, on a personal account. Google's Gemini Apps Privacy Notice is the one that should stop you. With Keep Activity turned off, submitting feedback still collects your feedback, any files, and "associated conversations". Google defines those as "the last 24 hours of your chats" and anything uploaded in them. That material is "reviewed by specially trained teams" and "retained for up to 3 years, disconnected from your Google Account". Rate one answer at 4pm, and you may have sent everything since 4pm yesterday.
Why do the companies do this?
Because feedback is the most useful training data they have. A conversation where a person has said "this answer was wrong" tells the model something a million unrated chats cannot. The companies are open about the purpose. Anthropic says it may use feedback to "train our AI models". Google says the data helps "maintain and improve Gemini Apps". Nobody is hiding it. It is in the privacy notice, but most people never read that.
The rating helps them, but doesn't do much for you. Your next answer doesn’t get better just because you clicked thumbs down on the last one. The model does not learn from you in the moment. The only thing that changes is that a copy of your conversation now sits in a review queue for years.
Is a business plan any different?
Yes, and this is where the business gap opens up. Plenty of businesses with five to fifty team members have people on ChatGPT Plus or Claude Pro, paid on their own cards or expensed. That is not the business plan. Those are personal accounts, and everything we discussed above applies to them.
The business plans have different terms. OpenAI's enterprise privacy page says ChatGPT Business and Enterprise data is not used for training by default. Then it names the exception: "If you have explicitly opted in to share your data with us (for example, through our opt-in feedback mechanisms)". The thumbs buttons are the opt-in. Anthropic's commercial page says the same feedback rules apply to Claude Team and Enterprise. The difference is that an owner can switch off the rating buttons for the whole organisation under "Rate chats" in the data and privacy settings.
Gemini in Google Workspace is different. Google's privacy hub for generative AI in Workspace says feedback "is not used to train any of the generative AI models" behind Workspace. Google keeps it for up to 18 months to fix bugs, and only sees your prompt and the answer if you tick a box to share them.
What should you do about it?
Never rate a chat. That applies on personal and business accounts alike. The rating does nothing for you. On a personal account, it sends the conversation, or a day of conversations, to a review team for years.
Find out which accounts your team is actually using. A ChatGPT Plus subscription on a team member's own card is a personal account with personal-account terms, even if the business pays. If several people are on personal plans, the business plan is often cheaper per head. It also comes with the data terms you assumed you already had.
Turn the buttons off where you can. On Claude Team and Enterprise, the owner can disable "Rate chats" for everyone. ChatGPT Business has no switch that OpenAI mentions, so there it has to be a rule your team knows. Gemini in Google Workspace does not need one because Google does not use its feedback to train the AI.
Put it in your AI policy. Our shadow AI post suggests a one-page AI policy? Add one line: do not use the rating buttons. It takes ten seconds to explain and it closes the gap that the privacy setting leaves open.
What a thumbs up or thumbs down sends, plan by plan
| What Rating an AI Answer Sends, Plan by Plan | ||
| Plan | What a thumbs up or down sends | How long it is kept, and who can switch it off |
|---|---|---|
| ChatGPT (Free, Plus, Pro) | The entire conversation, which may be used to train OpenAI's models even with the training setting off. | No retention period stated. No admin switch, because it is a personal account. |
| Claude (Free, Pro, Max) | The entire related conversation, including files, custom styles, and settings. May be used for training. Incognito chats excluded. | Up to 5 years, with the link to your account removed. No admin switch. |
| Gemini (personal Google account) | Your feedback, any files, and the last 24 hours of your chats and uploads, even with Keep Activity off. Reviewed by trained teams. | Up to 3 years, disconnected from your Google Account. No admin switch. |
| ChatGPT Business (and Enterprise) | No training by default. OpenAI names its opt-in feedback mechanisms as the exception, so a rating can share the data. | No retention period stated for feedback. OpenAI's page mentions no switch, so it needs a team policy. |
| Claude Team (and Enterprise) | The same as personal Claude: the entire related conversation, which may be used for training. | Up to 5 years. An owner can turn off Rate chats for the whole organisation. |
| Gemini in Google Workspace | Your rating and an issue category. Your prompt and the answer only if you tick the box. Not used to train the models. | Up to 18 months. Kimbley IT manages the admin settings for you. |
Frequently asked questions
-
Only for ordinary chats. OpenAI, Anthropic, and Google each say that when you rate an answer, the conversation is collected anyway. On a personal Gemini account, it is the last 24 hours of chats. The training setting and the rating buttons are separate, and only one of them is under the switch.
-
It is safer, though the rules differ. Gemini in Google Workspace does not use feedback to train its models, but it keeps feedback for up to 18 months. Claude Team lets an owner disable ratings. ChatGPT Business treats feedback as the one opt-in exception to its no-training default. The simplest policy is the same across all plans: do not rate.
-
The entire related conversation, including any files, custom styles, and settings, is stored for up to five years. Anthropic removes the link to your account before it is used and does not combine it with your other chats. Incognito chats are excluded.
-
No. Google's Workspace privacy hub says feedback is not used to train the models behind Workspace's AI features. It is kept for up to 18 months and used to fix bugs and analyse problems. Your prompt and the answer are only included if you tick the box to share them.
-
Ideally no. Personal accounts, whether paid or free, operate under consumer terms. A business plan gives you the no-training default, admin controls, and, in some cases, the ability to turn off feedback. If your team is on Google Workspace, Gemini is already there on business terms, and nobody needs a separate plan.
Book a call
If you are not sure which AI accounts your team is using, book a video call with Kimbley IT using the form below. You will get a straight answer on what your team has today and what to change.
