What actually happens when an AI tool reads your typing
When you use most AI writing tools, something specific happens the moment you start typing.
Your keystrokes, or at minimum the text you're composing, leave your device. They travel to a server. A model on that server processes them and sends a suggestion back. You see the result in a fraction of a second and it feels instant.
But the data made a round trip. And that matters.
What the round trip includes
Different tools handle this differently. Some send only the last few words. Others send the full paragraph, or the full document context, to generate better suggestions. Some log what you type for model training. Most have privacy policies that say they won't, but the transmission happens either way.
The server receiving your text is operated by a company. That company has employees, security practices, legal obligations and data retention policies. If they're breached, your text could be in the breach. If they're acquired, your data goes with the acquisition. If a government subpoena arrives, what they hold is what's discoverable.
None of this is hypothetical. It's just how cloud software works.
The specific concern with writing tools
Writing tools are different from, say, a cloud spreadsheet. When you're composing an email or a document, the content often reflects what you're thinking, not just what you've decided to share. Drafts contain things that don't make the final version. Notes include context that was never meant to be visible.
Professionals with confidentiality obligations notice this acutely. Lawyers pause before typing client names. Doctors reconsider before describing symptoms. Therapists, HR teams, financial advisors. But this isn't only a professional concern.
If you'd feel uncomfortable with a stranger reading your draft over your shoulder, you should probably know where that draft is going.
How local AI is different
Local AI means the model runs on your device. Your text never leaves. There's no server receiving your drafts, no company that can be breached, no policy to trust.
The tradeoff, historically, was quality. On-device models were slower and worse. That's changed. Modern Macs have chips capable of running small language models at speeds fast enough for real-time autocomplete. The quality gap has narrowed considerably.
This is what Typeahead does. The model runs on your Mac. When you type, the suggestion is generated locally. Nothing is transmitted. You can disconnect from the internet entirely and it works exactly the same.
How to verify it yourself
You don't have to take any company's word for it. With a cloud AI tool, open your Mac's Activity Monitor, find the process, and watch the network column while typing. You'll see the traffic.
With a local tool, turn off your Wi-Fi and keep typing. If the suggestions still work, nothing is going to a server. That's a complete verification — no network means no transmission.
Privacy policies are documents. Offline functionality is proof.