A thumbs-down on an AI answer shouldn't double as permission to train on the conversation. I might be telling the machine its arithmetic's dreadful, not offering it the customer file as a leaving present.
Yet AI feedback can do rather more than register disappointment. Depending on the product and account, submitting it can make the associated conversation available for model training, even when the normal setting says otherwise. Training means using examples to improve the system's future behaviour. That's a different transaction from correcting one bad answer.
I think the distinction belongs beside the button, in language a busy person can understand. A small control shouldn't carry a large surprise.
Your complaint can become their raw material
On ChatGPT's individual plans, Free, Plus and Pro, you can opt out of having new conversations used to improve the models. But if you then submit feedback on a response, the entire conversation associated with it may still be used for training.
Not just the sentence that annoyed you. The conversation.
That matters because a business question can contain much more than the question. Imagine asking an assistant to explain why a customer proposal doesn't add up. The thread could include pricing assumptions, the customer's objections and several attempts to rewrite the offer. That's a hypothetical example, but it shows why the boundary matters: the disappointing answer and the commercially useful context arrive in the same parcel.
The chatbot gets a poor performance review and the training department gets the stationery cupboard. Efficient, if you're the training department.
Anthropic's commercial-product policy has its own version of this distinction. It doesn't use inputs and outputs from those products to train its models by default. Explicit feedback or bug submissions can allow chats and coding sessions to be used for training.
Its thumbs-up and thumbs-down feedback policy also says the entire related conversation, including content and settings, is stored for up to five years. It de-links feedback from user and customer IDs before use and doesn't combine it with your other conversations.
Those are real limits, and they matter. They don't turn the submission into a one-bit vote about whether the answer was any good.
Five years is quite a commitment for a relationship that began with “this spreadsheet looks wrong”.
The account matters more than the logo
This isn't a claim that every feedback button works the same way. The differences are the point.
For ChatGPT Business, Enterprise and Edu, ordinary response feedback doesn't itself permit training on the associated conversation. OpenAI's policy says using those prompts, responses or files for model improvement requires a separate agreement with the organisation.
So “we use ChatGPT” doesn't answer the useful question. Which account? Under which terms? With which sharing controls?
Nor does paying automatically settle it. ChatGPT Plus and Pro sit in the individual-plan category for this feedback rule. The relevant distinction isn't simply free versus paid.
I don't think this makes one supplier virtuous and another villainous. It makes the blanket reassurance “your data isn't used for training” an incomplete purchasing answer. The exception can be the bit an employee encounters during an entirely ordinary afternoon.
Call me cynical, but the model supplier benefits when a correction also supplies a useful example. That's an understandable commercial incentive. It isn't automatically the customer's incentive. A business wants the bad answer fixed; the supplier may also gain material for improving the product it sells.
I'm quite happy for companies to ask for that exchange. I object to treating it as too minor to deserve proper attention.
Keep the complaint, shrink the parcel
My practical preference is to separate reporting a problem from sharing the original business conversation wherever possible. A made-up example can demonstrate a broken calculation without carrying the customer's proposal along for the ride. If the fault depends on the original material, then someone needs to decide whether sharing that material is acceptable.
That's a judgement about information leaving the business, not a test of whether staff are sufficiently enthusiastic about AI.
There are controls worth examining before turning it into a training lecture. Anthropic lets Team and Enterprise owners disable members' thumbs-up and thumbs-down submissions through its Rate chats setting. OpenAI also lets workspace admins enable or disable response feedback. The consequences differ, so the settings deserve a look rather than a universal ban.
None of this means feedback is bad. Better examples can help improve software. But useful feedback and permission to reuse a conversation aren't the same thing, and founders shouldn't let a cheerful little icon blur them together.
I'd rather judge an AI supplier by how clearly it handles these mundane boundaries than by how warmly it promises to understand my business.
A thumbs-up is a review, not a deed of gift.
Would clearer sharing rules make you switch AI suppliers?