Every few episodes, Netflix pauses whatever you only half-remember starting and asks the one direct question in its entire interface: Are you still watching? Not what you want to watch. Not what kind of evening you were hoping for, or whether you would rather feel informed or pleasantly numb. Just whether a body remains on the couch — because if nobody clicks, the counting stops. It is a polite little question, and also the whole philosophy of the modern feed compressed into four words: the system is exquisitely attentive to what you do, and completely incurious about what you want.
None of this is an oversight. TikTok’s own documentation explains that its ranking is built from behaviour — what you like, share, comment on and, weighted most heavily, what you watch to the end; finishing a longer video counts as a “strong indicator of interest.” Your device and account settings receive less weight, the company says, “since users don’t actively express these as preferences.” Read that sentence twice. The one place where you can set something deliberately — the one interface where you get to tell the machine anything — is discounted precisely because it is stated rather than enacted. YouTube’s published recommendation research ranks videos by expected watch time rather than clicks, and Netflix has credited its recommender with roughly four-fifths of hours streamed, valuing it at more than a billion dollars a year in retained subscribers. This is the revealed-preference machine: never ask, always infer; treat every tap as a tiny referendum and every completed video as a landslide.
The two of you
Economists have a tidy name for something everybody knows in their body. There are revealed preferences — what your behaviour demonstrates — and reflective preferences, what you would endorse after a moment’s thought. The first self opens the app to check one thing; the second surfaces forty minutes later having watched a stranger restore a rusty knife, a man rank every airline meal he has ever eaten, and three video essays on why a film you never saw is a betrayal. Nobody deceived you. You were served. Every tap was honoured: the impulsive self ordered for the table all evening, and the kitchen never once asked whether anyone was hungry.
The strongest evidence that the two selves genuinely disagree is experimental. In 2018, the economists Hunt Allcott, Luca Braghieri, Sarah Eichmeyer and Matthew Gentzkow paid thousands of Facebook users to deactivate their accounts for the four weeks before the US midterm elections. On average, the quitters reported higher subjective well-being, knew less about the news and — after the payments ended — kept using Facebook far less than the control group. In the authors’ careful phrasing, deactivation “reduced post-experiment valuations of Facebook, suggesting that traditional metrics may overstate consumer surplus.” Translated from the econometric: once people stepped off, they found that the thing they had been doing for hours a day was worth less to them than their own behaviour had advertised. If time spent were satisfaction, quitting would have hurt. It did the opposite.
The gap can now be drawn as a diagram. In a 2022 paper, Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan model a platform that is genuinely benevolent — it wants to maximise its users’ utility — but can observe only engagement. They show that when users have inconsistent preferences, which is to say when users are people, there are directions the feed can drift in which engagement rises while utility falls. The system gets better at its objective and worse at its purpose, with no bug, no rogue engineer and no villain. The flaw is not in the code but in the compression: attention is a lossy encoding of approval, and we have rebuilt the media diet of a civilisation on the artifacts of that encoding.
An honest version of this argument has to concede something. When researchers re-ran the feeds of a sample of Facebook and Instagram users during the 2020 US election in simple reverse-chronological order — no ranking, just time — those users spent substantially less time on the platforms and engaged less. Minute to minute, revealed preference is real: the ranked feed wins the hand’s vote, over and over. (The same study found no significant effect on political polarisation, which suggests the stakes here are stranger and more banal than the usual brainwashing story.) The problem was never that the algorithm shows you things nobody wants. It is that it reliably shows you things only one of you wants — the self who votes every second — while the other, who votes rarely and late, was never given a ballot.
The metric with a payroll
What a system measures, the world eventually manufactures. Creators study the objective the way students study an examiner: the hook inside two seconds, the retention edit, the thumbnail arms race, the apology video delivered in the same cadence as the unboxing video, because both answer to the same curve. The extreme case is a matter of court record. In September 2016, Facebook disclosed that it had overstated average viewing time for video for two years — by 60 to 80 percent, the Wall Street Journal reported at the time. In that interval, publishers including Mic and Mashable had laid off writers and staffed up video teams to chase the numbers; the “pivot to video” was, in effect, a measurement error with a payroll. Advertisers sued, and Facebook settled the class action for $40 million. A feed does not merely filter culture. It commissions culture, in bulk, and pays on delivery of the metric rather than the meaning.
The missing text box
Against this machinery of inference, the stated-preference toolkit is almost comically thin: on TikTok it is a long-press and a menu item reading “Not Interested” — a veto over individual videos, not a brief about your life. Nowhere in any major feed is there a field that asks what you would like more of in your head, or what tonight’s internet should leave in you. A recent preprint surveying decades of recommender research — not yet peer-reviewed — concludes that the field still treats behaviour as ground truth and evaluates its systems on next-click prediction, which mostly replays the past back at us. The omission is not mysterious. The Allcott experiment implies that a feed serving the reflective self would be used less and valued more, and used less is the one outcome the engagement objective cannot be asked to produce. The gap between your two selves is not a market failure. It is the market.
So the next time the screen dims and asks, with infinite patience, Are you still watching?, notice what the question takes for granted: that your continued presence is the unit of account, and that your satisfaction — unmeasured and unasked — can be safely ignored. The algorithm is not reading your mind. It was never asked to. It is reading your hand, and it takes the silence of everyone else you might have been as consent.