Skip to content
Change language: EN
My learningAdd to Chrome
BlogLocal AI

The airplane-mode test: the only privacy proof for AI you can run yourself

Every AI tool says it respects your privacy. One test settles it: switch off the internet and see if it still works. What offline AI really requires, how a workflow can fail the test, and how to run it in five minutes.

Dannick K.
October 13, 2026 · 3 min read
PresentBoard
The Summarize this page workflow running offline: run, read the page, summarise on the device, show the summary.
Share
On this page
  1. One test settles it
  2. What has to be on your device
  3. Watch a workflow pass
  4. Run the test yourself
  5. Three ways to fail without noticing
  6. Try it

Every AI tool says it respects your privacy. Few of those claims can be checked by the person reading them. You cannot audit a company’s servers, read its logs or verify what it keeps.

There is one test anyone can run: switch off the internet and see whether the AI still works. If it does, your data cannot be leaving, because there is nowhere for it to go. This post explains what has to be true to pass, how a workflow can fail without warning, and how to run the test in five minutes.

One test settles it

The test

If it works offline, your data stayed home.

It stayed home, because there was no road out.With no network interface up, no request can carry your data anywhere.

Kinetic line
Text version
  1. The test: If it works offline, your data stayed home.

Most AI features work by sending your text to a model on someone else’s server and waiting for the answer. That round trip is the reason they need the internet, and the reason your text leaves your machine. When the model runs inside your browser instead, the trip disappears:

Your laptop
The internet
Your browserthe page you are reading
The AI modelusually on a company server
the page text
a summary
Needs internetonly to installYour text leavesnoWorks on a planeyes

Move the model into your browser. The trip disappears.Weights cached locally; inference runs in the tab. Network only for the first download.

Round trip
Text version
  1. Your laptop: Your browser · The internet: The AI model
  2. Your browser → The AI model: the page text
  3. The AI model → Your browser: a summary
  4. Needs internet: every time · Your text leaves: yes
  5. The AI model → Your laptop · Needs internet: only to install · Your text leaves: no · Works on a plane: yes

What has to be on your device

Passing the test is not a setting you switch on. It is what happens when every layer of the work is already on your machine:

  1. Every other stepNo cloud service anywhere in the chain.
  2. The pageAlready open: its text is read where it is.
  3. The engineRuns the model inside the tab, on your hardware.
  4. Model filesDownloaded once, kept in your browser's storage.

It is four layers, and every one must be local.Cached weights, a local runtime (WebGPU or WebAssembly), local input, and no remote nodes.

Layer peel
Text version
  1. Every other step / The page / The engine / Model files
  2. Every other step: No cloud service anywhere in the chain. · The page: Already open: its text is read where it is. · The engine: Runs the model inside the tab, on your hardware. · Model files: Downloaded once, kept in your browser's storage.

The first three are about the AI. The fourth is the one people forget. An automation is a chain of steps, and it is only as offline as its most connected step. A single “save to Google Sheets” at the end means the run needs the internet, however private the AI part was.

Watch a workflow pass

Here is a workflow that passes. It reads the page you have open, summarises it with a model running in your browser, and shows the result. No step needs the network once the model is installed.

Summarize This Page

TriggerRun now
PageRead the current page
AISummarize it
PageShow the summary
you pressed Run
page text read
120-word summary, written locally
summary shown in the panel
Example figures

Now run it with Wi-Fi off. Same result.One run, four steps, no network calls. Example outputs.

Live flow
Text version
  1. summarize-this-page.awf
  2. you pressed Run
  3. page text read
  4. 120-word summary, written locally
  5. summary shown in the panel

You press Run

Nothing is scheduled on a server. The workflow starts in your browser.

It reads the page

The text comes from the tab you already have open, not from a download.

A local model writes the summary

The model was installed earlier and runs on your hardware.

The result appears

It shows up in the panel, on the same machine where it was written.

Then you turn the network off

Run it again with no connection. If you get a summary, the privacy claim is a fact you checked.

Run the test yourself

It takes five minutes, and you only need to do it once per workflow:

Before take-off

The airplane-mode test

  • Install the model while you are online (done)your test
  • Press Test so it loads once (done)your test
  • Open the page you want to process (done)your test
  • Switch on airplane mode or turn Wi-Fi off (done)your test
  • Run the workflow and read the result (done)your test

Run it. A result means nothing left your device.Any step that fails is a step that needed the network.

Checklist
Text version
  1. The airplane-mode test
  2. Install the model while you are online
  3. Press Test so it loads once
  4. Open the page you want to process
  5. Switch on airplane mode or turn Wi-Fi off
  6. Run the workflow and read the result

If a step fails, it just told you something useful: that step needs the internet. Fix it or accept it, but now you know.

Three ways to fail without noticing

Myth

If the AI step is offline, my whole workflow is offline.

Fact

Only the steps that stay on the device are. One cloud step at the end, and the run needs the internet.

The fact. Check every step, not just the AI.Integration nodes and cloud models call out; on-device models and page reads do not.

Myth vs fact
Text version
  1. If the AI step is offline, my whole workflow is offline.
  2. Only the steps that stay on the device are. One cloud step at the end, and the run needs the internet.

The other two are quieter.

The browser throws the model away. Browsers manage storage for you, and when space runs low they may clear data, large models included. An offline demo that worked on Monday can need a fresh download on Friday. On AWFlow’s Local AI page, turn on Keep models installed: when the browser agrees, the storage header shows Protected.

The Local AI page: storage used by on-device models split by engine (WebLLM, TF.js, Transformers), a task list on the left (Chat, Embeddings, vision and audio tasks) and model cards on the right, such as Llama 3.2 1B Instruct with its size, variants and install state.The Local AI page: storage used by on-device models split by engine (WebLLM, TF.js, Transformers), a task list on the left (Chat, Embeddings, vision and audio tasks) and model cards on the right, such as Llama 3.2 1B Instruct with its size, variants and install state.

Turn on Keep models installed so they survive a full disk.Requests persistent storage; the header shows Protected when granted.

Guided screenshot
Text version
  1. 1. The storage header: Space used by your models, and Keep models installed.

Browsing for models goes online. Searching the model list sends your search text to Hugging Face, where model files come from. If you want the Local AI page itself to stay offline, turn off Discover models on Hugging Face in Settings › Local models.

Installing a workflow from the Marketplace also needs an account and a connection. Do it before you leave.

Try it

Install a chat model from Local AI, open this workflow, then run it once online and once with Wi-Fi off.

Before your next flight

Summarise a page in airplane mode

Open it in AWFlow, pick an installed model, switch your Wi-Fi off, and run it on any article.

Open in AWFlow (opens in a new tab)docs.awflow.io/examples/summarize-this-page.awf
Scan to open

Call to action. Viewers scan the code, readers click the link.The link opens the Summarize This Page docs example in AWFlow's import page.

Try it
Text version
  1. Summarise a page in airplane mode · Open it in AWFlow, pick an installed model, switch your Wi-Fi off, and run it on any article. · https://docs.awflow.io/examples/summarize-this-page.awf

What we covered

  1. Offline is the privacy proof
  2. Every layer must be local
  3. Test it: cut the network and run

Three. Five minutes, once per workflow.Install online, test, cut the network, run.

Recap
Text version
  1. What we covered
  2. Offline is the privacy proof
  3. Every layer must be local
  4. Test it: cut the network and run
Written by
Dannick K.
All articles

Keep reading

Back to the blog →
Local AI

Running LLMs inside a Chrome extension: WebLLM, transformers.js and the MV3 traps

What it took to run AI models inside a Manifest V3 extension: three runtimes, two Web Workers, a WebGPU probe with a WASM fallback, and the CSP rule that silently blocks your runtime.

5 min readOctober 15, 2026
Local AI

Tool calling for small on-device LLMs that don't support tool calling

Models running in the browser give you text in and text out, so agent frameworks refuse them. How we added prompt-based tool calling with a model-agnostic parser: one format asked for, many accepted, and three rules that keep it safe.

8 min readOctober 15, 2026
Local AI

We rebuilt RAG in a browser tab. Our first bug: every vector was empty.

Our knowledge bases embed, store and search your documents inside the browser. Before rebuilding them, an audit found that no knowledge base had ever had a single searchable chunk. Ten RAG lessons from that rebuild, written so you can check your own pipeline.

7 min readOctober 15, 2026

Automate the web you're on. Free, in your browser.

Add to Chrome — free