How the fox knows what you're doing

What it looks at

Three things, and only while a sitting runs

The app in front, the title of its window, and — in a browser — the address of the tab. That is the whole of it.

For most of the day that is already enough. Xcode is work. Slack is work. Netflix, TikTok, a game client: not work, and the fox's marks start the moment you land there. Nothing is inferred and nothing is scored — a short list of places nobody opens mid-sitting by accident, and everything unrecognised is work. A focus companion that scolds you for opening a PDF reader it hasn't heard of is one you turn off within a day.

Asleep

You're working, or somewhere the fox has no opinion about. The clock runs.

Awake

You've drifted. Three minutes of marks over its head, the clock holds, and it forgives completely when you come back.

YouTube

A lecture from a rabbit hole

Half the lectures a student watches are on YouTube, and so is every time sink ever made. The address decides nothing, so three readers take a turn.

1. The page's own tags. YouTube publishes a category for every video — Education, Gaming, Music — along with the uploader's keywords, the channel and the length. That is the uploader's own claim, filed with the platform, and it beats anything a title can be read to mean. When it speaks, it's final.
2. Apple Intelligence, on your Mac. For the videos the tags leave open — "People & Blogs" says nothing — the model on this Mac reads the title, the channel and the tags and calls it: study, fun, or unsure. It is allowed to say it doesn't know, and told which way the mistakes cost.
3. The words in the title. While the first two are still thinking, and on a Mac with neither, a hundred-line word list makes the same call you'd make from a tab strip: "lecture", "강의", "proof" one way; "gameplay", "vlog", "먹방" the other. A title carrying both is a game wearing a study word.

Shorts are the one exception. Nobody has ever studied from a feed that scrolls itself, and the address says so outright — no reader is asked.

Google

A search for work from a search for the score

Google is the other address that decides nothing. The same page serves "kubernetes rolling update", "photosynthesis definition", "elden ring boss order" and "lakers score". The words you typed are the whole evidence — and they're already in the URL, so there is nothing to fetch.

The fox reads them the same two ways, and names one of five things: work, study, games, sports or shopping. Work and study keep the clock running. Games, sports and shopping wake it.

1. The word lists first. A query that names a game, a league, a team, a coupon, a theorem or an error message is placed on the spot, no model needed. Games and sports are checked first, because they name themselves loudly — "minecraft tutorial" is Minecraft.
2. Apple Intelligence for the rest. "Who won last night", "how to beat the water temple", "why do leaves change colour" — the model on this Mac is asked what the search was for, and may answer unsure.

Only a results page counts. Google Docs, Gmail, Maps and the home page are never looked at: a Doc is work, and the fox has no business reading it. The Shopping tab is the one results page that needs no reading at all — a page of prices is shopping whatever you typed to get there, and the fox says so without asking the model.

Being wrong

Three marks over a cartoon fox

Every reader above is a heuristic, and every heuristic is sometimes wrong. What none of them is allowed to be is wrong silently.

When the fox wakes, the bar says why — "you're on YouTube", "you're on Google (sports)" — and the clock holds rather than counting the minutes as focus. A mistaken mark costs you three punctuation marks and is cleared the instant you switch back. The opposite mistake — a fox that sleeps happily through an hour of Shorts — would cost the whole feature. So where the readers are unsure, YouTube leans toward drift and Google leans toward work, because that is where each one's bad guess is cheapest.

What leaves your Mac

Almost nothing

Apple Intelligence runs on the Mac. The window title never leaves it. The search you typed is read out of a URL that's already on your screen. Nothing about any of this is sent to AsX — there is no AsX server in the picture.

There is exactly one request, and it goes to a site you're already on: to read a YouTube video's own tags, AsX fetches the same watch page your browser just loaded. One page per video, no account, no cookies, nothing identifying who's asking, and the answer is cached only until you quit. Offline, or behind a consent wall, the page simply isn't read and the fox judges by window title alone.

• No model on this Mac? The word lists answer alone. Nothing is sent to a model anywhere else, ever — the judge is local or it is nothing.
• Apple Intelligence off? Same. Turn it on in System Settings and the fox starts using it at the next sitting, no restart.
• Nothing is stored. The verdicts live in memory and die with the app. The day's ledger records where the time went, not what you searched for.

The full picture is on the privacy page. What happens once the fox has decided — three marks, and then it comes forward — is the second article.