Every consumer AI tool has a thumbs up and a thumbs down button next to the answer. Most people do not press them. The ones who do, press them quickly. Tap, dismiss, carry on.

I want to spend a few minutes on what that button actually does, because the answer is not what most users I speak to assume. And I want to do it in the same week that Google updated the Gemini Apps privacy terms to spell out, in plain English, that some of your saved chats are reviewed by human beings and kept on file for up to three years.

This is a reader-service piece. There is no grand argument. There is just a list of things you ought to know about your AI tools, and a short set of practical actions to take this week.

The Gemini change, in plain English

Google’s Gemini Apps Privacy Hub was updated in May 2026 with the following, which I will lift word for word so there is no ambiguity:

“Humans review some saved chats to improve Google AI. To stop this for future chats, turn off Gemini Apps Activity. If this setting is on, don’t enter info you wouldn’t want reviewed or used.”

The page goes on to spell out the mechanics. A subset of saved chats is reviewed by Google’s trained service providers. The reviewers look at how Gemini responded. If the response could have been better, they show Google’s AI what a better response would look like. Suggestions from reviewers are then used to train future models.

If you do nothing, that is the default.

Two things are worth knowing on top of that.

First, chats reviewed by human reviewers, along with related data such as your language, device type, location info or feedback, are not deleted when you delete your activity. They are retained for up to three years.

Second, your Google Workspace content (the things in Gmail, Drive and the rest of the Workspace suite) is not reviewed or used to improve the Gemini app. That is the consumer-business split Google has been gradually formalising for a year, and it matters. Personal Gemini is one product. Workspace Gemini is another.

How do you switch it off for the consumer side? Go to myactivity.google.com/product/gemini and toggle Gemini Apps Activity off. Even then, Google retains chats for 72 hours to respond to you and to help keep Gemini safe. And the sad thing is that by turning off activity in Gemini, none of your chats will be saved for you to have access to them either. What a facepalm.

ChatGPT, in plain English

OpenAI’s help centre on how your data is used is clear that personal-tier ChatGPT chats can be used to improve OpenAI’s models, and that you can switch this off in your settings under Data Controls by turning off “Improve the model for everyone”.

What is less clear, until you read it carefully, is what happens when you press the thumbs button.

When you tap thumbs up or thumbs down, you are submitting feedback. That feedback is, by design, attached to the conversation around it, because the conversation is what makes the feedback meaningful. OpenAI’s own documentation and developer community guidance make clear that if you choose to provide that feedback, the associated content can be used to improve the system, even if you have separately opted out of training in your settings.

In other words, the thumbs button is the explicit opt-in route that survives your opt-out everywhere else.

If you are using ChatGPT for work in any sensitive area, the practical implication is that you should treat that thumbs button as a publishing action, not a private gesture. If you would not be comfortable with a trained reviewer at OpenAI looking at the whole conversation, do not press it.

Claude, in plain English

Anthropic’s Privacy Center and Help Center page on user feedback settings are, in fairness to Anthropic, quite clearly written. The relevant points are:

• On consumer plans (Free and Pro), if you press the thumbs up or thumbs down button on a Claude response, the entire related conversation is stored in Anthropic’s secure back-end for up to five years.
• That feedback-linked data can be used to improve Claude models. This is the explicit opt-in route, the same as ChatGPT.
• On Claude for Work, Claude Enterprise, Claude for Education and Claude Gov, your data is not used to train Anthropic’s models. That is covered under the commercial terms and is not subject to opt-in or opt-out toggles.
• On Team and Enterprise plans, an owner can switch off the ability for members of their organisation to submit feedback at all. The setting is called Rate chats, and it sits under Organization settings → Data and Privacy.

That last point is the one I want every IT lead, operations lead and compliance lead to write down. If your business is on a Claude Team or Enterprise plan, the default for the Rate chats toggle is probably not where you would set it if you thought about it deliberately. Go and look. Decide. Document.

The pattern across all three

If you read the three policies side by side, the pattern is clear.

The default position on consumer AI is now that human review is part of how the product improves, and your chats are part of the training pipeline unless you have actively moved them out of it.

The way out of the training pipeline is different on each platform, but it always exists. The way back in, even after you have opted out, is the same on each platform: it is the thumbs button.

This is not a scandal. It is not even particularly surprising. It is how machine learning gets better. Human reviewers grading model outputs, paired with paying customers willing to provide signal, is the dominant loop by which the major AI products have improved over the last three years. The companies are now being more explicit about it because regulators and journalists are pushing them to be.

The reason it matters for you is that the relationship between “free or low-cost AI” and “the data feeding the next model” is now explicit. And the burden of managing that relationship sits squarely with the user.

What to do this week

If you take five minutes from this piece, take it to do these four things.

1. Open Gemini’s privacy settings. Go to myactivity.google.com/product/gemini, and decide whether you want Gemini Apps Activity on or off. Either choice is a valid choice. The wrong move is to leave it on the default without thinking about it.

2. Open ChatGPT’s Data Controls. In Settings → Data Controls, look at “Improve the model for everyone”. If you are using ChatGPT for client work, professional research or anything sensitive, you almost certainly want that off. Then re-train your reflex on the thumbs buttons. Treat them as publish, not vote.

3. Open Claude’s settings. If you are on a consumer plan, the same logic applies. If you are an owner on a Team or Enterprise plan, go to Organization settings → Data and Privacy and look at the Rate chats toggle. Decide. Tell your people what the decision is.

4. Use business tools for business things. Microsoft 365 Copilot, Claude for Work, Google Workspace Gemini, the OpenAI Business tier, and Perplexity Enterprise all sit on commercial terms that do not feed training. The price for that protection is usually a paid licence and a slightly less generous chat window. That is a fair trade. If you are doing client work, regulated work, or anything where the data has a value to someone other than you, that is the trade you should be making.

The honest takeaway

The AI you use is improving partly because people like you, often without realising it, are tagging examples for the people training it.

That is not, by itself, a bad thing. It is the engine that has made these tools as good as they are. But the burden of deciding what you are willing to contribute has now formally landed on the user. And the easiest way to contribute more than you meant to is by tapping a thumbs button on a conversation you would not have wanted to read out loud in a room.

The good news is that all three platforms now give you the controls. They are not always in the place you would expect, but they exist, and they are not hidden.

Go and use them. Be deliberate. And the next time you reach for the thumbs button, take an extra second to ask whether you really want the conversation around it to leave your screen.

That extra second is what AI hygiene actually looks like in practice.

I discussed the rate button and the Gemini privacy update on Chapter 14 of Prompt Fiction, our podcast on what is actually happening in AI.

Scott Quilter FBCS | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD