How to keep your private data away from AI servers (and still use ChatGPT)

If you’ve ever pasted a tax question, a bank statement, or a chunk of work code into ChatGPT, you’ve probably had that little voice in your head: “Should I really be putting this here?” You’re not wrong to worry. By default, ChatGPT sends your data to OpenAI’s servers and keeps it there for a while.

The good news: you can keep the convenience of the ChatGPT desktop app while swapping in a “local brain” that runs entirely on your own machine. That’s a big win for privacy—especially when we’re talking about sensitive financial and work information.

Why this matters for your money

From a pure privacy standpoint, it’s unsettling that so much of our digital life flows through a few big AI providers. From a money standpoint, it’s worse. We routinely feed these tools:

• Tax returns and pay stubs
• Bank and credit card statements
• Loan documents and social security numbers
• Investment summaries and brokerage exports
• Work documents, proprietary code, and client data

If that information leaks or gets misused, the fallout isn’t just embarrassing—it’s financial. Think identity theft, drained accounts, or exposure of sensitive business data. This is the same reason I’m picky about how I send payments and why I’ve stopped mailing paper checks whenever I can avoid it: the more copies of your information that exist, the more chances something can go wrong.

What actually happens to your ChatGPT data

When you use the ChatGPT Mac app, it follows the same rules as the web version. That means:

• Your chats live in the cloud. Every prompt, every response, every file you upload, and every screenshot you attach gets stored on OpenAI’s servers and tied to your account.

• By default, your data can be used to train models. Unless you go into settings and explicitly opt out, your content helps improve future models.

• Even if you opt out, retention doesn’t disappear. Opting out stops training use, but OpenAI can still keep your chats for around 30 days (sometimes longer) for safety and abuse monitoring, even after you delete them from your chat history.

So if you paste a PDF of your tax return or your company’s internal docs into ChatGPT, you’re not just “talking to an AI.” You’re sending that file to someone else’s computer, where it sticks around for a while.

Two extra features that quietly expand what gets uploaded

The Mac app adds a couple of powerful—but risky—features that can silently pull in more of your data than you realize.

1. “Work with apps”

When you enable “work with apps,” ChatGPT can read from other open apps on your Mac. It can grab roughly 200 lines of content from:

• Your IDE or code editor
• Your terminal
• Notes apps and text editors
• Possibly anything else that’s open and readable

If you have a database password in a terminal window, a secret API key in a config file, or proprietary code in Xcode or VS Code, that content can be bundled into your request and sent to OpenAI’s servers. It feels magical when it “just knows” what’s on your screen—but that magic is powered by more of your local activity being uploaded.

2. “Computer history”

OpenAI also rolled out a feature called “computer history.” If you opt in, it records your clicks, typing, and app switching over time to build a richer memory of what you’re doing.

The worrying part isn’t just what gets sent to the cloud. The memory files themselves sit unencrypted on your Mac. Any other app running as the same user can read them. That means your activity history is exposed on your own machine before it even reaches OpenAI’s servers.

Again, this is convenient, but it widens the attack surface: more data stored in more places, with more ways for it to leak.

What local AI models do differently

Local models flip this model on its head. Instead of sending your data out to a remote server, you download a model to your own computer and run it there. One popular way to do this is with an open-source app called Ollama.

Here’s what changes when you use a local model through Ollama:

• Your prompts and responses never leave your device. The “thinking” happens on your own CPU/GPU. There’s no outbound network traffic for the actual chat.

• Ollama doesn’t collect or store your prompts. Their privacy policy is explicit: for local models, they don’t collect, store, or transmit your inputs. They don’t see your data at all.

• You can literally unplug the internet and keep going. If you turn off Wi‑Fi or pull the network cable and the model still works, that’s the sign it’s truly local.

This doesn’t mean your computer suddenly becomes invincible—you still need basic security hygiene—but it does remove one giant risk: constantly shipping your sensitive information to a third party.

How to set up Ollama on your computer

Ollama works on macOS, Windows, and Linux, but the direct integration with the ChatGPT desktop app is Mac-only for now. Even so, the basic setup is very similar across platforms.

Here’s the high-level process:

1. Download Ollama. Go to the Ollama website and click the download button for your operating system.

2. Install the app. On Mac, drag Ollama into your Applications folder. On Windows and Linux, follow the standard installer steps.

3. Launch Ollama. Open the app so it can start the local server it uses to run models.

4. Pull a model. Ollama lets you download different models by name. A good starting point is something like “Ornith 1.59B” (you can copy the exact model name from their website under the Models section). It’s powerful but still small enough that most Macs, including cheaper ones, can run it.

Once the model is downloaded, you’re ready to connect it to your tools.

Connecting local models to the ChatGPT Mac app

If you’re on a Mac and already using the ChatGPT desktop app, you can plug your local Ollama models right into that interface.

Here’s how the integration works at a high level:

1. Open the Ollama desktop app.

2. Go to the “Apps” section in the sidebar.

3. Find ChatGPT and flip the integration on. You can also toggle this from the menu bar if that’s easier.

4. Restart the ChatGPT app. The integration needs a restart to take effect.

Now comes the slightly awkward part: picking which models show up.

Adding your local models to ChatGPT’s picker

Inside the Ollama settings for ChatGPT, you’ll see a section where you can choose which models appear in the ChatGPT model picker. Unlike some other integrations that use a dropdown, this one uses a text-like block where you list models.

A few quirks to know:

• There are already five models selected by default.
• You can’t add a sixth model without removing one first.
• That means you have to delete one of the existing entries, then add your local model by name.

It’s a little clunky, but once you’ve added your preferred local models, you’re done. Restart the ChatGPT desktop app one more time, and your local models will appear right alongside the OpenAI ones in the usual model picker.

From there:

• The UI looks the same.
• Keyboard shortcuts work the same.
• You can switch between local and cloud models mid-conversation if you want.

The only real difference is whose “brain” is answering you—OpenAI’s servers or your own hardware.

A big warning about mixing local and cloud models

This part is critical if you care about privacy.

If you share a private document or paste sensitive information into a chat while using a local model, that data stays on your machine. But if you later switch that same conversation over to an OpenAI model, everything in that chat history—including the earlier private content—can get sent to OpenAI’s servers.

To stay safe:

• Use separate chats for sensitive and non-sensitive work.
• If you’ve shared anything private and want to switch models, start a new chat before changing from local to cloud.

This is the same mindset I use with financial apps: I don’t mix “play” accounts and real money accounts in ways that could accidentally cross wires. Keep clean boundaries.

What this setup does—and doesn’t—protect

Running local models is a big improvement, but it’s not magic. Here’s what it actually does for you.

What it helps with

• Keeps prompts and responses on your device. No automatic upload of your sensitive data to an AI provider when you’re using a local model.

• Reduces the risk of third-party leaks. If OpenAI or another provider has a data issue, your local-only conversations aren’t part of that pool.

• Gives you a clear “offline test.” If it works with Wi‑Fi off, you know it’s not silently streaming your content out.

What it doesn’t solve

• It’s per-user, not company-wide. This setup protects your own machine. It doesn’t automatically secure your whole team or organization. Everyone would need their own setup.

• It doesn’t fix a compromised computer. If your device already has malware or someone else has access, local models won’t save you. Basic digital hygiene still matters.

• It doesn’t remove all cloud use. You might still choose to use OpenAI or other cloud models for certain tasks. The key is being intentional about when you do that.

How local models actually feel to use

There used to be a huge quality gap between cutting-edge cloud models and what you could run locally. That gap is much smaller now than most people think.

In practice:

• First load can be slower. When a model first spins up, your computer has to load it into RAM. That can take a bit, especially on lower-end machines.

• After that, speed is often perfectly fine. Once loaded, responses can feel as fast—or sometimes faster—than cloud models, because there’s no network latency.

• Quality is “good enough” for a lot of work. For drafting emails, summarizing documents, brainstorming ideas, or doing basic code help, local models are often more than capable.

Personally, I like a hybrid approach:

• Use cloud models for public, low-risk tasks like general research, casual brainstorming, or rewriting something that doesn’t contain private details.

• Use local models for anything that would hurt if it leaked: legal documents, medical details, proprietary work, or sensitive financial records.

When to absolutely avoid cloud AI with your data

There are some categories of information I strongly recommend keeping off cloud AI tools unless you’re 100% sure about the provider’s policies and your own comfort level:

• Full tax returns, W‑2s, 1099s, or other documents with SSNs
• Bank and credit card numbers, routing numbers, or full statements with account details
• Loan applications, mortgage documents, or anything with deep personal profiles
• Medical records or detailed health histories
• Proprietary code, client contracts, or internal company strategy docs

These are exactly the kinds of details that can fuel identity theft, account takeover, or business damage if they end up in the wrong hands. If you’re curious about how these risks show up in the real world, it’s the same general category of concern I wrote about in my piece on what license plate cameras mean for your privacy and your wallet: more data, more tracking, more potential misuse.

Cloud AI isn’t evil—but you should be picky

I’m not in the “never use cloud AI” camp. These tools are incredibly useful, and for a lot of everyday tasks, they’re perfectly fine. But I do think we should be more deliberate about what we feed them—especially when money, identity, or work secrets are involved.

Local models give you a third option between “use nothing” and “send everything to OpenAI”:

• Keep the shortcuts, integrations, and comfy UI you like.
• Swap the underlying brain to a local model for sensitive work.
• Use cloud models only when the risk is low and the benefit is high.

That’s a much healthier balance between convenience and safety.

Practical next steps

If you’ve been pasting private financial or work data into ChatGPT, here’s what I’d do next:

1. Review your ChatGPT settings. Turn off training where possible and think twice about enabling features like “work with apps” or “computer history.”

2. Install a local model tool like Ollama. Set up at least one capable local model you can switch to for sensitive tasks.

3. Separate your chats by risk level. Use different conversations (and ideally different models) for high-risk vs low-risk content.

4. Adopt a simple rule of thumb. If it’s the kind of information you’d guard on paper—tax returns, bank data, legal docs—keep it on your machine and use local AI.

Protecting your financial life isn’t just about budgets and investments. It’s also about where your data lives, who can see it, and how many copies of it exist. Local AI is one of the few tools that actually lets you pull some of that control back.