TL;DR:
The heavy computing behind most assistants runs on powerful remote computers, so your question has to travel there and back.
Offline, you may still get typing, saved files and some small on-device features, but not the full answer.
The trip also explains the privacy question: whatever you send is what leaves your machine.
You are on a train, the wifi drops, and you ask your assistant to explain the Pythagorean theorem. Nothing happens, or it says it cannot connect. That seems odd: the laptop is right there, it has a fast processor, and the theorem is not new. Why does the answer depend on a signal?
The short answer
A large language model is a huge pile of numbers, billions of them, and producing each word of an answer means running a lot of arithmetic over those numbers. That is far more work than an ordinary laptop can do quickly, and the numbers alone can take tens of gigabytes of storage. So most services keep the model on powerful computers in data centers and send each request over the internet.
Your device is more like a window and a microphone than a brain. It collects what you asked, ships it out, and displays what comes back.
What happens step by step
Capture. Your device records your voice or reads your typing. If it is speech, it may be turned into text first, on the device or on a server.
Send. The request is packaged and sent over the network to the service's computers.
Compute. Those computers run the model. The answer is built word by word, not all at once.
Return. The words (or an audio version) stream back to you as they are made, so you can start reading before the end.
Display or speak. Your device shows or plays the result.
Every one of the middle steps needs a connection. Break the link at step 2 and nothing reaches the computers that can answer.
Walk-through: an assistant answering on a laptop
Here is a simple timing story with invented but plausible numbers, to show where the waiting goes. Suppose you ask a question and the answer is 120 words long.
Stage | Example time | Where it happens |
|---|---|---|
Your question is captured | 0.5 s | Your laptop |
Request travels to the data center | 0.05 s | Network |
The model starts and writes the first words | 0.8 s | Remote computers |
The rest of the answer streams back | 3 s | Remote computers and network |
Display | instant | Your laptop |
Add them: 0.5 + 0.05 + 0.8 + 3 = 4.35 s until the full answer is on screen, though you can read from about 1.35 s onward because the answer streams. The key point is that almost all the time is the remote computing, not your laptop. A faster laptop would barely change it.
Now imagine the same model running on your laptop. If your machine could do the work only one tenth as fast as the data center, the 3 s stretch becomes 3 × 10 = 30 s. That is why even well-built local assistants tend to use much smaller models: they trade quality for speed.
Where the computation happens, and why it matters
There are three places the work can happen, and a given assistant often mixes them.
On your device. Fast to respond, works offline, private by default, but limited to smaller models and features.
On a remote server. Strong models and updates without installing anything, but needs a connection and sends your request off the machine.
Split. The device handles easy parts such as detecting a key press or showing the answer, and the server handles the thinking.
Which is better depends on the job. A spell checker is small and fits on a phone. A tutor that must explain any topic in any subject needs a much bigger model, which is why those usually call out to a server.
What it can't do offline
Answer open ended questions at full quality.
Look up anything recent, since a stored model only knows what it was trained on.
Use services that live on the internet, such as your email or a shared document.
What often still works: reading your own saved files, basic typing and local features built into your operating system. Some devices include small on-device models for narrow tasks such as dictation or photo sorting, though what is available varies a lot.
Slow is not the same as offline: A weak connection can look like a broken assistant. If answers arrive in bursts or time out, test with a web page before blaming the app.
Privacy and control
The network trip is exactly why privacy settings matter. Whatever you send leaves your device, even if only for a moment, and you should know whether it is stored afterwards. Read the product's privacy page and look for plain statements about what is kept and for how long.
A useful habit is to control when the assistant listens and looks, because those are the moments data is created. DuckyHelper, for example, only listens or looks while the pill is open and for that answer, with no listening or watching when idle. Whatever tool you use, check that it makes the same kind of promise.
Try it yourself
Run a two minute test. Ask your assistant a question with the internet on, note how it responds, then turn wifi off and ask again. See what error appears, and whether anything still works, such as opening a saved file. You will learn quickly where your tool does its thinking.
If you are curious how files get read once the connection is up, see dragging files in for help. To think about what is safe to send, read data safety basics for AI tools at work.
Frequently asked questions
Will assistants ever work fully offline?
Small ones already do for narrow jobs. Whether a full tutor can run on an everyday laptop depends on hardware improving and models getting smaller, and nobody can promise a date.
Does a faster internet make answers faster?
Only a little. Most of the delay is the remote computing, so past a basic connection a faster line barely helps.
Is a connected assistant always sending my data?
Not necessarily. It sends what you ask and what it needs for that request. Check whether it works only when you trigger it.
Get started
