It's gone eight, the last client left forty minutes ago, and the owner of the salon (I'm inventing her, but you know her, or you are her) is still at the reception desk with a laptop, typing into ChatGPT. She wants help wording a message to a supplier who keeps sending the wrong shade of dye. She types the supplier's name, an order number, and the reason she had to reschedule a client that afternoon. Ordinary Tuesday-night stuff.
What she doesn't know, because almost nobody tells you this when you sign up, is that on the free plan she's using, that whole conversation can become raw material for training a future version of the model. Not because anyone is spying on her salon. Because that's the setting it ships with, and she never went looking for the other one.
By the end of this, you'll know exactly which setting that is, on whichever AI tool you actually use for work, and why "personal" and "business" plans aren't only a price difference.
What "training" actually means for your words
When people worry about AI "training on their data," what happens is usually less dramatic than it sounds, and more permanent. Nobody pulls out your exact sentence and reads it back to a stranger. Your conversation gets folded into an enormous pile of other conversations, and the model shifts, very slightly, based on patterns across that pile: which phrasings worked, which got corrected, what people tend to ask and how. Your supplier problem doesn't come out the other side recognisable to anyone.
But the sentence you typed doesn't un-happen either. Once a conversation is used in a training run, there's no pulling it back out. And on more than one of the big tools, before anything reaches "training," a real person may read a sample of the conversation to judge whether the model's answer was any good. That's a detail most quick-start guides skip. A recent beginner's guide to ChatGPT I checked this week explains that there's a toggle to turn training off, which is honest as far as it goes. It just never mentions that the toggle only exists because the default is on, or that the answer changes completely once you're paying for a business account instead of a personal one.
Check your tier, not your fear
Here's the pattern, and it holds across the three tools most small businesses actually reach for. On ChatGPT's personal plans (Free, Plus, Pro), conversations are used to improve the model by default, unless you go into Settings and turn off "Improve the model for everyone." On ChatGPT Business, Enterprise, and the API, that's off from the start; there's nothing to find because there's nothing switched on.
Claude runs the same shape. Free, Pro, and Max accounts default to "Help improve Claude" being on. You turn it off in Privacy Settings, and once you do, it also shortens how long your conversations are kept. Claude for Work and the API are excluded automatically, same as ChatGPT's business tiers.
Gemini, on a personal Google account, keeps your conversations and has some of them reviewed by people to judge response quality, unless you switch off "Keep Activity" for Gemini Apps in your Google Activity Controls. If you use Gemini through a company Google Workspace account, check with whoever administers it rather than assuming. Business setups vary more here than with the other two.
None of this is hidden exactly. It's published, it's legal, and as one Twitch executive put it bluntly this week, announcing a nearly identical default for streamers' videos: "If this was opt-in, nobody would opt in. That's honestly the answer." Defaults aren't a trick. They're just how you get the number of people flipping a switch from a handful up to nearly everyone. It works precisely because most of us never open the settings menu at all.
What flipping the switch does and doesn't do
Turning training off is worth five minutes, and it's not a perfect shield. Anthropic is upfront that a conversation flagged by their safety systems can still be kept and used regardless of your setting. You're not told when this happens, and you can't ask. Anything already folded into a training run before you switched it off stays there; the toggle only protects what you type next.
So here's the honest limit, and it matters more than the toggle itself: a setting is not the same thing as a promise. If a detail would embarrass you or a client if it turned up somewhere unexpected (a real name, a health note, a number that identifies someone), that detail doesn't belong in a personal AI chat at all, business tier or not. The toggle changes the odds. It doesn't change what's actually wise to type.
Tonight's five minutes
Open whichever AI tool you actually use for the business (not the one you meant to switch to) and find the training setting: Data Controls in ChatGPT, Privacy Settings in Claude, Activity Controls in Google. Turn it off if you're on a personal plan. If you're already on a business or paid team account, confirm it's excluded rather than assuming. It usually is, but "usually" is exactly the word worth five minutes of checking.
And keep the client's actual name out of the chat either way. That's the same instinct behind when you're legally on the hook for what an AI does without you checking it: the tool's settings are real, but you're still the one who decides what goes in.
Clara F.