Why You Cannot Stop Scrolling
A slot machine holds no model of you. A feed holds nothing else. The argument is that the second one builds the first by accident
Answer to the question
Why is a social media feed so hard to stop scrolling?
Because a good but imperfect recommender produces a variable-ratio reward schedule by accident: most items are mildly worth continuing and an unpredictable few are genuinely rewarding. Ferster and Skinner found that schedule produced the highest response rates and the strongest resistance to extinction. Whether design or accident is doing the work is untested, but the schedule is there either way.
Key takeaways
- A recommendation system is a prediction engine. Its entire objective is to reduce uncertainty about what you will engage with next, and every improvement to it is an improvement in accuracy.
- A variable-ratio schedule works on the opposite principle. It pays out after an unpredictable number of responses, and the unpredictability is the active ingredient.
- Those two things produce the same result, and that is the argument: a recommender that is good but imperfect generates a variable-ratio schedule as a by-product, without anyone designing one.
- Ferster and Skinner catalogued the schedule in Schedules of Reinforcement in 1957. It produced the highest response rates and the strongest resistance to extinction of anything they tested.
- Nobody outside the platforms can test which account is doing the work, because the experiment requires degrading a live recommender for real users and measuring what happens.
The two claims are usually made in the same breath, and on the face of it they cannot both be true.
The first is that your phone is a slot machine: refresh the feed and the reward arrives on an unpredictable schedule, the arrangement that produces the most persistent behaviour psychology has ever measured. The second is that the feed knows you, that it holds a model of your attention built from years of behaviour and serves what that model predicts you will stay for.
Randomness and prediction are opposite engineering goals. One works because it cannot be anticipated. The other works because it anticipates you. A slot machine holds no model of the person pulling the handle and its payout ratio is fixed before anyone sits down. A ranking algorithm does nothing but hold a model of you, and every engineering hour spent on it goes into making the next item less of a surprise.
| Slot machine | Recommendation system | |
|---|---|---|
| Optimising for | A payout ratio fixed in advance | Predicted engagement, recalculated per person |
| Knows about you | Nothing | A behavioural model built over years |
| Improves by | Staying unpredictable | Becoming more accurate |
So the mechanisms point in opposite directions, and the resolution is that the second one manufactures the first.
A recommender is not trying to be predictable to you. It is trying to be accurate to itself, and those are different things. The model works better than chance and nowhere near certainty. What that produces at your end is a stream in which most items are mildly worth continuing and an unpredictable minority are genuinely rewarding. The ratio is set by the accuracy of the model rather than by a randomiser, but from the seat of the person holding the phone it is a variable payout arriving after an unpredictable number of swipes.
That is the thing Ferster and Skinner described, arrived at from the other direction.
Their 1957 book Schedules of Reinforcement runs to more than 700 pages of what happens to behaviour under different payout rules. The variable-ratio schedule, where reinforcement comes after an unpredictable number of responses averaging to some ratio, produced the highest and steadiest response rates of anything they ran, and behaviour trained on it was the hardest to extinguish. Stop paying on a fixed schedule and the subject notices at once. Stop paying on a variable one and it keeps going, because an absent reward is indistinguishable from an ordinary gap.
A casino has to design that schedule deliberately and tune it by hand. A recommender arrives at one as a by-product of being imperfect, and it arrives at a schedule tuned to the individual rather than to the average customer, which no casino can do.
The strongest version of the opposing case does not need any of this, and it is worth stating at full strength because it may be the right one.
On that account the distinction between designing a schedule and stumbling into one is a distinction about intent, and intent is not what does the damage. Philosophers have made the structural case: a 2024 paper in Topoi by Cristina Voinea, Lavinia Marin and Constantin Vică argues that platforms operate as attentional scaffolds built on exactly this schedule, and the argument does not depend on how the schedule got there. The machine gambling literature documents the design intent in slot machines in detail, and the family resemblance to a feed is not subtle. Patents, former-employee testimony and the platforms' own disclosed engagement metrics all point the same way.
Neither account has been tested against the other, and the reason is mundane rather than sinister. Separating them requires degrading a live recommender for a randomised subset of real users and measuring what happens to session length. The only organisations that could run that experiment are the ones with the least reason to publish the result.
The published engagement figures cannot stand in for it either. They report how much people use a product and how that changes when a model is improved, which is consistent with both accounts and discriminates between neither. A rise in session length after a ranking upgrade is what you would predict if accuracy were the active ingredient, and equally what you would predict if a better-tuned model happened to produce a better-tuned schedule. The metric everyone has is the one metric that cannot settle it.
Which leaves the slot-machine people holding the better position, and they should have the last word on it.
They would say the analogy was never about the dice. It was about the schedule. Whether the schedule was drawn up by a designer in Las Vegas or fell out of a ranking model in Menlo Park is a question for historians of technology, and it changes nothing about what the schedule does to a person holding a phone at one in the morning. The mechanism is there. Asking who put it there is a way of not looking at it.
Sources
Ferster, C. B. & Skinner, B. F. (1957). Schedules of Reinforcement. Appleton-Century-Crofts. The variable-ratio schedule produced the highest response rates and the strongest resistance to extinction of the schedules tested.
Voinea, C., Marin, L. & Vică, C. (2024). "Digital Slot Machines: Social Media Platforms as Attentional Scaffolds." Topoi. The strongest statement of the position this article sets itself against, and the one given the last word.
No primary source is cited for the claim that recommender accuracy generates a variable-ratio schedule. It is an argument from the two mechanisms, presented as such, and the article states that the experiment separating them has not been run.