It’s the question I get asked more than any other

Can I safely use my data with AI?

I get asked this question probably more than any other. It shows up in the Q&A, in the hallway afterward, and in the slightly nervous email three days later from the person who didn’t want to ask in front of the room. And it almost never arrives on its own. It turns up wearing a dozen different outfits. Should I put in customer names? What about my company’s IP? I’m in mortgage (or healthcare, or wealth management), so am I even allowed to? My company gave us Copilot, is that the safe one? I use ChatGPT for everything, should I be worried? Should I connect my calendar? My email?

Sound familiar?

What I’ve come to realize is that all of those questions are really one question, and the reason people feel overwhelmed isn’t that the answer is complicated. It’s that nobody has ever given them a clear one. So let’s fix that. The good news, and I mean this, is that you can absolutely use your data with AI safely. “Safely” just depends on a few things you actually control, and once you can see them, the whole topic stops being scary.

The reassuring part, and I mean this, is that almost all of it comes down to a few things you control. So let’s start there.

You’re probably already doing this (most people are)

Let me put the scale of it on the table. In a TELUS Digital survey of 1,000 US enterprise employees, 57% admitted they’d entered confidential company information into public AI tools like ChatGPT, Gemini, and Copilot, and 68% were doing it through personal accounts rather than anything their company had set up. Which is the polite way of saying that most of the data going into AI is going in through the back door, by people who genuinely have no idea what the rules are. (I find that more reassuring than alarming, oddly. It means the problem isn’t recklessness. It’s a clarity gap, and clarity is fixable.)

So before we talk about which tool is “safe,” we need to talk about what actually happens to your words after you hit enter. Because almost nobody has been told.

What actually happens when you type something into a chat box

When you type a message into ChatGPT, Claude, Gemini, or Copilot and send it, your text leaves your device and travels (encrypted, in transit) to the company’s servers. That’s where the model lives. Your words get processed, an answer comes back, and in most cases a copy of that conversation is saved to your account as history so you can find it again later.

That much is the same everywhere. The part that varies, and the part that matters, is what happens to that saved copy next. Depending on which tool you’re using, which plan you’re on, and which settings are switched on, your conversation might be used to help train future versions of the model, it might be kept on the company’s systems for a long time, and in some cases a sample of conversations gets reviewed by actual humans to check quality and catch abuse.

One detail that surprises people. Deleting a chat usually removes it from your history right away, but there’s typically a back-end window (often around 30 days) before it’s fully purged, and crucially, anything that’s already been used to train a model can’t be pulled back out afterward. The training has happened. To my mind that’s the single most important reason to get your settings right now rather than later, because the only data you can fully control is the data you haven’t sent yet.

None of this means the tools are dangerous. It means they’re not a sealed vault, and you should know which version of them you’re standing in.

The one distinction that answers most of your questions

If you remember nothing else from this piece, remember this. There’s a world of difference between a consumer account and a business or enterprise account, and that single line settles most of the worry.

A consumer account is the personal one. You signed up with your own email, maybe you pay for a Plus or Pro or Max subscription, and there’s no contract between your company and the AI provider. On these accounts, your conversations are generally fair game for model training (sometimes by default, sometimes via a setting you’re asked to choose), and they tend to stick around longer.

A business or enterprise account is different in the way that counts. Your organization has signed a commercial agreement with the provider, and under that agreement your data isn’t used to train the models, you get admin controls and audit trails, and the company can decide how long things are kept. This is the version built for company data.

So when someone asks me “is ChatGPT safe?”, it’s surprisingly hard to give a straight answer. It’s a simple question, but a few things hide underneath it. Which version are you on? Which account are you logged into? Which settings are switched on? Answer those and you’ve basically got it.

The five-second setting to check today

On any consumer account, there’s usually a toggle deciding whether your chats get used to train the model, and turning it off is the highest-value five minutes you’ll spend this week. Menus move around, so treat these as a map rather than gospel, but as of right now:

Google, to its credit, openly tells people not to put confidential information into consumer Gemini, which tells you something.

The four tools, plainly explained

License levels, from least protected to most

People assume that paying for AI makes their data private. It’s the most common misconception I run into, and it’s worth busting head-on. The ladder runs from least protected to most protected.

• Free consumer. Trains on your data (by default or via a setting), longest retention, no contract, no oversight. Perfectly fine for low-stakes, non-confidential work. Brainstorming a blog title, summarizing a public article, that sort of thing.
• Paid consumer (Plus, Pro, Max). This is the one that surprises people. Paying for a personal plan buys you more capability, not more privacy. The data rules are essentially the same as the free tier. You’re not safer because you handed over a credit card.
• Business or Team. Now you’re under a commercial agreement. No training on your data by default, admin controls, and usually a Data Processing Addendum. This is the floor for company data, the point at which you can reasonably relax.
• Enterprise. Adds the heavier machinery. Single sign-on, audit logs, retention you control, data residency choices, and independent certifications like SOC 2 Type 2 and ISO 27001 that prove the provider’s controls have been examined by an outside auditor.
• API with zero data retention. This is for built applications rather than chat windows. In the right commercial setup, API use can be configured for very low retention, which is why it’s often the preferred route for sensitive applications. It’s what most regulated and genuinely sensitive workloads get built on, and it’s the tier we build on at Rivet Labs precisely because it keeps client data off the provider’s systems entirely.

 A consumer chatbot is a gifted assistant with a long memory and no contract. Use it accordingly.

“But what about my IP? Can someone pull it out of ChatGPT?”

This one comes up constantly, usually from content creators and from leaders guarding something proprietary, and the two cases pull in opposite directions, so let’s take them one at a time.

For content creators, my honest opinion is going to sound almost reckless. So what?

Hear me out. If you’re a speaker, an author, a coach, anyone whose ideas are meant to travel, then the entire point of your IP is that it gets out into the world. The moment your framework leaves your mouth on a podcast, lands in an article, or goes up on your website, you should simply assume AI knows about it. Not because of anything you typed into a chatbox, but because the public web is precisely what these models are trained on. Your published work is out there being scraped, quoted, and absorbed, and there’s no setting anywhere that changes that. Trying to keep public ideas away from AI is like trying to keep your billboard away from people with eyes.

So rather than fret about it, claim it. Loudly, repeatedly, everywhere. The protection for a content creator was never secrecy, it’s being the recognized source. If your ideas are so clearly yours that people can’t think them without thinking of you, AI knowing them is free distribution, not theft.

One small piece of honesty so you’re not surprised. AI won’t always attribute. It can repeat an idea, sometimes even your exact phrasing, without a citation, the same way a person can quote something clever they once heard without remembering where. Annoying, yes. But the answer to that is the same as it’s always been for anyone with ideas worth stealing, you out-publish, you out-associate, and you make your name inseparable from the concept. (Ask me how I feel watching one of my framework names show up in someone’s LinkedIn post with no credit. And then ask me whether I’d rather it never traveled at all. Easy choice.)

Now the opposite case. If you’re trying to hold something back from competitors, a not-yet-launched product, a pricing model, an unreleased strategy, the calculus flips entirely, and this is where I get firm. Genuinely confidential, unpublished IP doesn’t go into a consumer AI tool. The cautionary tale here is real. Samsung engineers reportedly pasted proprietary source code into ChatGPT to debug it, and that’s the exact scenario you’re trying to avoid, because that code was never meant to see daylight. Unpublished IP belongs under the same rule as client data and regulated information from the rest of this piece. Business or enterprise tier under a proper agreement, or it stays out.

The clean line to carry around is this. Published IP, share it and own it. Unpublished IP, protect it like the asset it is.

Your published ideas don’t need protecting from AI. They need claiming. It’s the unpublished ones that need the locked door.

A short, nerdy detour for the curious (skip if you like)

Once you understand this next bit, a lot of the fear settles down. When your conversation is used to train a model, your words aren’t being filed away in some giant searchable database, sitting there as a tidy block of text waiting for a stranger to call it up. That’s the mental picture most people have, and it’s the wrong one.

What actually happens is that your text gets broken into tokens (small fragments of words) and used to nudge the statistical patterns inside the model, the thing engineers call its weights. The model doesn’t store your sentence. It very slightly adjusts its sense of which words tend to follow which other words. Your individual contribution mostly dissolves into an ocean of patterns.

For something to come back out word for word (researchers call this memorization), it generally has to have appeared in the training data many times in the same form. Stanford researchers studying this found that a non-trivial amount of repetition is needed for verbatim memorization to happen. Something you typed once, in one conversation, is extraordinarily unlikely to ever resurface in someone else’s chat. The odds of a competitor casually stumbling onto your single prompt while noodling on their own idea are, practically speaking, somewhere near nil.

The honest footnote, because I’d rather you have the real picture than a comforting half of it. That repetition can come from the wider world, not just from your own keyboard. If a popular idea of yours gets quoted, syndicated, and re-posted across hundreds of pages, it can become memorizable that way, no chatbox required. Which loops us right back to where we started, because it’s your published footprint that AI learns most reliably, and that was always the part you wanted out there anyway.

“But I’m in a regulated industry”

This is the question I get from mortgage, healthcare, finance, and legal audiences, and the answer is yes, you can use AI, but almost never on a consumer account.

The agreements you’re required to have only exist on the business and enterprise tiers. For most industries that means a Data Processing Addendum, and for anyone touching protected health information it means a Business Associate Agreement, or BAA. OpenAI, for example, will sign a BAA for its healthcare and API products with zero data retention configured, and it won’t do so for Free, Plus, Pro, or even standard Business. The contractual protection simply doesn’t reach the consumer tiers, no matter how carefully you use them.

So if you’re in mortgage, the nonpublic personal information your clients hand you (Social Security numbers, account numbers, income documents) doesn’t belong in a personal ChatGPT window. Not once, not “just to draft something quickly.” The same logic covers protected health information in healthcare, privileged material in legal, and client holdings in wealth management.

One thing gets lost in every vendor brochure. The agreement only covers the provider’s side of the line.

Your provider’s contract protects their side of the line. Everything your team types before the data gets there is still on you.

You can have the most ironclad enterprise contract on the planet and still spring a leak the moment one well-meaning person pastes a client file into their personal account to save five minutes. The technology can be compliant while the behavior around it isn’t. That’s why the real work in regulated industries is less about choosing the tool and more about making sure your people know, without having to guess, what they can and can’t put where. (And please, check with your own compliance or legal team and your regulator’s guidance. I build this stuff for a living, which is exactly why I know I’m not the person to give you the legal version.)

Should you connect your calendar, email, and files?

You can, and honestly it’s often where AI gets genuinely useful. An assistant that can see your calendar can actually schedule things. One that can read the right folder can answer questions about your own work instead of making things up.

The simplest way to think about a connection is this. When you link a data source to an AI tool, or upload a file into it, treat it exactly as though you’d typed every word of that content into the chat box yourself. Because functionally, that’s what you’ve done. Whatever the tool actually opens and reads, the confidential clause buried in that PDF, the client details in that email thread, the numbers in that Google Doc, all of it becomes an input the moment the AI processes it. The same permissions, the same warnings, and the same training and retention rules apply. A connection doesn’t get a special pass just because you didn’t type the words by hand.

Which means this lands right back on the distinction we keep coming back to. On a consumer account, a sensitive document you let the AI read is treated like sensitive text you pasted in. On a managed work account under an agreement, that same document inherits the business-tier protections instead. The connection takes on whatever rules the account already has.

So the principle is to match the access to the sensitivity. Connecting a tool to your email or Drive is like handing a new assistant the keys to specific rooms. You wouldn’t give a brand-new temp the master key to the entire building on day one, and you don’t need to here either. On a personal account, be conservative about what you plug in. On a managed work account, those connections are usually controlled by your IT admin, which is precisely where that decision belongs.

A final reminder…

Vendor policies can (and do) change regularly. Make sure you’re regularly checking terms, conditions and policies!

• • •

The big picture

If you want the whole thing on one screen, here’s the decision in a single view. Screenshot it, pin it above your desk, send it to the person on your team who keeps asking.

• • •

Your homework

• Know which account you’re standing in. Personal, or work. This one answer resolves most of the anxiety, so check it before anything else.
• On any consumer account, go and turn off training today. It’s five seconds and it’s the difference between your words feeding a future model or not.
• Keep genuinely sensitive or regulated data out of consumer tools entirely. If it has to go into AI, it goes through a business or enterprise tier under a proper agreement, or it doesn’t go in at all.
• Find out whether your company has an AI policy. If it does, read it. If it doesn’t, that absence is the actual risk, far more than the tools themselves.
• For connectors, match the access to the sensitivity. Start narrow. You can always open more doors later.

The real gap isn’t the technology

Underneath every version of this question is the same thing. People aren’t being reckless. They’re guessing, because nobody handed them a clear answer. That’s a leadership gap, not a technology one.

The organizations that get this right give their people three things so the guessing stops. Principles, meaning what we actually believe about handling data with AI. Policies, meaning the rules in plain language with real examples, the “never paste a client’s account number into a chatbot” kind of specific. And playbooks, meaning which tool to use for which kind of work. When those three exist, something lovely happens. The anxiety drops and the safe usage goes up, at the same time.

That’s a lot of what I do with teams, turning “I’m scared I’ll do something wrong” into “I know exactly what’s fine and what isn’t.” If your people are guessing about this (and the survey says they almost certainly are), let’s have a conversation about giving them a clear answer instead.