Talk to AI Without Feeding It

People tell chatbots things they wouldn't tell their doctor: symptoms, contracts, relationship problems, source code, the draft resignation letter. On a cloud service, all of it lands on someone else's servers. It can be kept, reviewed by staff, used to train future models, exposed in a breach, or handed over under legal order. And because you're usually logged in, it's attached to your name.

You have three ways to cut that exposure, from strongest to weakest: run the model yourself, put a buffer between you and the provider, or tighten the settings on the service you already use.

1. Run it locally

A model running on your own computer never sends your prompts anywhere. No account, no logs, works offline. Open-weight models have become good enough that for summarising, drafting, rewriting and coding help, a laptop is often all you need.

What your hardware can handle. Model size is counted in parameters:

Choose quantized versions (compressed weights, usually in the GGUF format). They run far better on consumer hardware for a small loss in quality. A modern GPU, or Apple Silicon's unified memory, makes everything faster.

The easy way: LM Studio

LM Studio is a desktop app for macOS (Apple Silicon), Windows and Linux that handles the whole thing with a normal interface (docs):

1. Download and install it.
2. Open the model search, pick a model that fits your RAM (the app flags what your machine can run), and download it.
3. Load it and start chatting. Once the model is downloaded, LM Studio runs entirely offline.

It can also chat with your documents locally, so contracts and notes never leave your disk. And it can act as a local server with an OpenAI-compatible API, so other apps and scripts can use your local model instead of a cloud one:

# download a model (this one needs ~16 GB RAM) and chat from the terminal
lms get openai/gpt-oss-20b
lms chat

# serve it locally (default port 1234)
lms server start
curl http://localhost:1234/v1/models

LM Studio's app isn't fully open source; the open-source tools below are, if that matters to you.

Open-source alternatives

• Ollama — command line, macOS / Windows / Linux. ollama run llama3.2 downloads and starts a model from its vetted library.
• Llamafile — Mozilla-backed; a model and its runtime in a single executable file. Nothing to install.
• Kobold.cpp — for tinkerers: fine-grained controls, plus local image generation and speech-to-text.

For the most transparency, pick truly open models whose training data is published, such as those from Ai2.

Download models safely

A model file is code you're running. Prefer the built-in libraries of Ollama or LM Studio. When downloading from Hugging Face directly, look for a verified organization, a clear model card, real download numbers, and check the file's SHA256 against the one listed on the page:

shasum -a 256 model.gguf          # macOS / Linux
certutil -hashfile model.gguf SHA256   # Windows

2. Put a buffer in between

When you need a frontier model your laptop can't run, avoid handing your identity over with every prompt.

• Anonymised chat. Duck.ai from DuckDuckGo offers several well-known models with no account. It relays requests so the model provider doesn't see your IP address.
• Logged-out use. Some chatbots work without signing in. Prompts are still sent and may be stored, but they aren't tied to an account history. Use a private window and don't volunteer your name.
• Third-party wrappers. Apps like T3 Chat put many providers' models behind one interface and subscription, and API routers like OpenRouter do the same for developers. The model providers then see the wrapper as the customer rather than you. But the wrapper now sees everything, so read its privacy policy, and remember your prompts still reach the underlying provider.

None of these make cloud AI private. They reduce how much of you travels with each prompt.

3. Lock down the account you use

If you stay on a mainstream service:

• Opt out of training. ChatGPT, Claude and Gemini each have a setting controlling whether your chats are used to improve models. Find it in privacy or data controls and switch it off.
• Use temporary or incognito chats for anything sensitive, and set history to auto-delete where offered.
• Delete old conversations you don't need. Stored chats are breach material.
• Don't connect everything. Every integration with your email, drive or calendar is more data the assistant, and its provider, can reach.

What never goes in a cloud prompt

Passwords, API keys and recovery codes. ID and bank numbers. Health records, legal documents, and anything confidential from work. Other people's personal details, since they never agreed to share them. If you must work with a document like that, strip the names and numbers first, or keep it local.

The most private prompt is the one that never left your machine. The second most private is the one with nothing of you in it.