What "private" really buys you

What you really do get

The promise from lesson one holds up. When the model is on your machine, your words have no reason to travel.

LM Studio says as much in its own documentation. It does not need the internet to work, and that covers chatting with models, chatting with documents, and running a local server.

There is a second promise, and it is quieter. The model does not learn from you. LM Studio puts it plainly: the model does not learn from chats. Nothing was sent anywhere, so there is nothing for anyone else to train on.

That is the real core of it. No account to lose. No service that can be switched off. Nothing of yours sitting in somebody else's logs.

Backed by LM Studio offline operation, Manage chats.

Privacy belongs to the model, not the app

This is the sentence to take away from the lesson. An app is not private on its own. The model you picked inside it is the thing that decides.

Ollama shows this well, because its FAQ answers the question twice in one breath. Asked whether it sends your prompts back to ollama.com, it says Ollama runs locally, and that it does not see your prompts or data when you run locally.

The same answer then keeps going. For cloud-hosted models it says it processes your prompts and responses to provide the service. It adds that it does not store or log that content and never trains on it.

Both halves are true. They describe two situations, and you chose which one you are in when you picked a model. Ollama's docs say cloud models are automatically offloaded to its cloud service.

There is a door you can shut. Ollama documents a local-only mode, turned on by a setting in a file called server.json, or by an environment variable. Restart it and the log records that cloud is disabled.

It also names the price. Turn that on and you lose Ollama's cloud models and its web search. A tool that tells you what a safety setting costs you is doing the honest thing.

Backed by Ollama FAQ, Ollama cloud models.

The app around the model still talks

The model is one part of what you installed. The app around it is another, and the app has its own reasons to go online.

LM Studio lists them. Searching for models makes requests to huggingface.co. Downloading a model needs a connection. So does its catalogue of models, and so does the check for app updates.

None of that sends your conversation anywhere. It does mean the phrase runs on your device describes the model, and not the whole program.

Look hardest at features bolted on top. Web search, or a hand-off to a bigger model in the cloud, both send something out. That can be a sound design, and track four is about building it on purpose. The point is to know which parts do it.

Backed by LM Studio offline operation.

Where your machine ends is a setting

There is a wall around all of this, and it is not quite where people picture it.

LM Studio can act as a server for other programs. Its docs say it can serve models on localhost or on the local network. Localhost means this machine and nothing else. The local network means the other computers in the building.

Ollama works the same way. Its FAQ says it binds to 127.0.0.1 on port 11434 by default, which is your own machine only. One environment variable changes that address, and then the machines around you can reach it. The same page explains how to let browser extensions in as well.

None of that is a flaw. It is what makes a model on your desktop useful from your laptop, or from a tool in your browser.

But notice what each one does. It moves the wall. Open one of those doors and you have not stopped running locally. You have changed who counts as local, and that is a decision somebody made in a settings file.

Backed by LM Studio offline operation, Ollama FAQ.

Your disk is not a vault

Staying on your computer is not the same as disappearing.

Chat apps save what you type. LM Studio keeps conversations as plain JSON files, and its documentation gives the exact path. On a Mac and on Linux they sit in a folder named .lmstudio in your home folder, inside a folder called conversations. On Windows the same folder sits in your user folder.

You can go and look right now. LM Studio says to right-click a chat and choose Reveal in Finder, or Show in File Explorer. It also asks you not to edit those files by hand.

Anything that can read your files can read those. Other people who use the machine. Anything you have installed. Whatever backs your home folder up.

Disk encryption answers the lost-laptop version of this. Apple says a Mac with Apple silicon or a T2 chip encrypts its data automatically, and that FileVault adds a layer needing your login password. Microsoft describes BitLocker as encryption for whole drives, aimed at theft and lost devices.

Neither one helps against a program running as you, on an unlocked machine, right now. Different risk, different tool.

Backed by Manage chats, Protect data on your Mac with FileVault, BitLocker overview.

A model file is still a download

You are putting a file of several gigabytes from the internet onto your computer. That earns the same care as any other download.

Some formats carry more risk than others. Hugging Face warns that loading a pickle file can allow dangerous attacks that run code of the attacker's choosing. Its advice is to load models only from people and groups you trust.

The safetensors format was made for this reason. Hugging Face describes it as a way to store tensors safely, as opposed to pickle.

The rule that follows is dull and it works. Get the file from the maker's own page, and prefer safetensors where you have the choice.

Backed by Pickle Scanning, Safetensors documentation.

Private does not mean allowed, or right

Two things privacy never bought you.

The first is permission. A licence travels with the file, and it still applies when you are offline. Meta's Acceptable Use Policy is part of the Llama licence. It forbids using Llama for the unlicensed practice of a profession, and it names financial, legal and medical work. Lesson seven works through this properly.

The second is accuracy. A private wrong answer is still a wrong answer. Microsoft's own card for phi-4 warns that language models can produce content that sounds reasonable but is wrong or out of date.

One more thing, said out loud. Whether running locally satisfies a rule you are under is a question about that rule. Your contract, your regulator, your employer. This course cannot answer it, and neither can a product page.

Backed by Llama 3.3 Community License Agreement, Llama 3.3 Acceptable Use Policy, Phi-4 model card.

Six checks worth doing

Each of these takes about a minute, and each one answers a different question from the others.

• Read the name of the model you are actually using. In Ollama, a name ending in -cloud means the answering happens on Ollama's computers.
• Find your app's own page about working offline. LM Studio publishes one. Whatever it lists as needing a connection is your answer.
• Open your chat folder and look at the files. Decide whether you are content for them to sit there in plain text.
• Check that whole-disk encryption is switched on. Apple calls it FileVault. Microsoft calls it BitLocker.
• If you turned on a server so another machine could use the model, write down who else can reach it.
• Read the licence for the model before you use it for paid work. Lesson seven shows you where to look.

Backed by Ollama cloud models, Protect data on your Mac with FileVault, BitLocker overview, Llama 3.3 Acceptable Use Policy.