Safety
Using local AI safely
Local AI is genuinely useful. It also makes mistakes, sometimes odd ones, and it says them with a straight face. Here's what to know before you rely on it.
What these models are, and aren't
A language model is very good at producing text that sounds right. It doesn't look things up or check facts unless a project specifically connects it to your documents, and even then it can misread them. It's a helpful assistant for drafts, ideas, and summaries. It is not an expert, a search engine, or a source of truth.
Why the models in our projects are weaker than ChatGPT
ChatGPT, Claude, and Gemini run on huge models in data centers. To fit on a home computer, the open models we use are much smaller, and usually quantized (compressed) to save memory. That trade-off is what lets them run on your laptop, but it also means they know less, follow instructions less reliably, and make mistakes more often.
"Hallucination", in plain terms
Sometimes a model invents things: a fact, a quote, a statistic, a website, a medication dose. It isn't lying; it's producing words that fit the pattern. The result can be confidently wrong, strange, or occasionally dangerous. Smaller models do this more.
Never rely on it for these
- Medical or health decisions
- Legal questions
- Money, tax, or investment decisions
- Anything involving physical safety: electrical work, chemicals, driving, emergencies
Ask a qualified person. In an emergency, call emergency services.
Check anything important
Treat what it gives you as a first draft. Verify names, numbers, dates, quotes, and instructions against a trusted source before you act on them or share them.
Private, but not automatically safe
Running AI locally means your prompts and files stay on your computer, which is a real strength. It isn't a guarantee:
- Download models and apps only from their official sources.
- Plugins, extensions, and add-ons can send data elsewhere. Check what they do.
- Anything that can read your files or run commands can make changes you didn't intend.
About the agent path
When you paste one of our agent prompts into Claude Code, Codex, or OpenCode, you're giving an AI tool permission to run commands on your computer. These tools usually ask before each step. Read what it plans to do, approve only what makes sense to you, and stop it if something looks wrong. Most of these assistants are cloud services, so what you share with them follows their privacy terms, not ours.
Our testing has limits too
Published results describe what happened on our machines, with specific versions, on a specific date. Your computer, software, and data are different, so your results may be too.
No warranty
Everything on this site is provided as-is, without warranty of any kind. You use these projects and the AI they set up at your own risk, and you're responsible for checking what it produces before relying on it.