You said one episode. It is now past one in the morning and you are four deep. The usual explanation involves a psychology experiment from 1927 — and that experiment has since failed to replicate. The real answer is better, and slightly more uncomfortable.
Let us start with the part that is not in dispute: somebody chose the number of seconds you get to decide.
When the credits roll, Netflix shows you a countdown. At the end of it the next episode begins on its own. Today that countdown is five seconds. It has not always been.
When the feature launched in August 2012 it was fifteen seconds. The engineer who built it later described the test: they tried five, ten and fifteen. Ten won. Five was, in his words, too fast — jarring. Fifteen gave people too much room.
Netflix now runs five. The developer’s own reading of why the losing number eventually won: viewers have been conditioned to expect autoplay. The design did not just get optimised once. It was re-optimised against an audience it had already changed.
He also said something worth sitting with. Autoplay produced, in his words, by far the biggest increase in hours watched of any feature they ever tested.
Autoplay arrives switched on. A button to turn it off did not exist until February 2020, after years of complaints.
That matters more than the countdown. The best-known demonstration of default power comes from organ donation: countries where you are enrolled unless you opt out have consent rates roughly double those where you must opt in. Same people, same values, different starting position on the form.
An opt-out interface reverses the cost of doing nothing. Continuing is free. Stopping is the thing that takes effort. Every piece of friction between you and the next episode has been located and removed, because features that did not lift hours watched were deleted.
And here is the number that should be better known. In 2024 researchers at the University of Chicago ran an actual randomised experiment — seventy-six Netflix users, split in two, monitored for up to seventeen days. Turning autoplay off cut average daily watching by twenty-one minutes. Average session length fell by seventeen and a half minutes. The gap between one episode and the next grew by twenty-four seconds — the pause the countdown had been compressing.
Seventy-six people is a small study, and it was US-only. It is also the best causal evidence that exists on this question, because the one organisation with the real data has never published it.
Almost every article on this subject tells you about Bluma Zeigarnik. A psychologist in a café notices that waiters remember unpaid orders perfectly and forget them the moment the bill is settled. She turns it into an experiment. Unfinished things stay open in the mind; finished ones get filed away.
Three problems.
The café was in Berlin, not Vienna — Zeigarnik was a graduate student under Kurt Lewin at the University of Berlin. More awkwardly, the words waiter, café and restaurant appear nowhere in her 1927 paper. The anecdote is attributed to Lewin, told in inconsistent versions, and is best treated as a framing story rather than an observation anyone recorded.
And the effect itself did not survive. A 2025 meta-analysis at the University of Bern pooled the replication attempts and found that once you exclude Zeigarnik’s own study, interrupted tasks are recalled about 49% of the time — which is to say, no advantage at all. The average effect size was 0.15. Their conclusion was blunt: the Zeigarnik effect lacks universal validity.
In the same Berlin lab, a year later, Maria Ovsiankina ran a different experiment. She interrupted people mid-task, then gave them free time and deliberately did not tell them to go back.
Across twenty-one published studies, people resume an interrupted task about 67% of the time, against a coin-flip baseline. That one held. Lewin wrote that all the later work in his laboratory was built on it.
So the famous sibling — the memory effect — collapsed, and the forgotten one — the behavioural effect — stood. Which is fortunate, because resumption is what a cliffhanger actually exploits. Nobody is claiming you remember the unfinished episode better. The claim is that you go back to it. And that is the claim with evidence behind it.
One detail from that literature is worth keeping: paying people to resume made them less likely to. The pull to finish is intrinsic, and money displaces it.
It is tempting to treat all of this as a Silicon Valley invention. It is not.
Dickens industrialised the serialised cliffhanger with The Pickwick Papers, published in nineteen monthly instalments from 1836, each ending in a way designed to manufacture demand for the next. The word itself comes from Thomas Hardy’s A Pair of Blue Eyes, serialised in 1872–73, in which a character is left literally hanging off a cliff face, eye to eye with a fossil.
The hook was written before the story roughly a hundred and ninety years ago. What changed is not the hook. It is that the gap after the hook used to be a month, and is now five seconds.
Two things are genuinely established here.
The first is narrative transportation: when you are absorbed in a story, you stop evaluating it as an argument. This has been measured across 132 effect sizes from 76 studies. It is the best-supported psychological claim on this page.
The second is the parasocial bond, a term coined in 1956 to describe intimacy at a distance. After the Friends finale, researchers surveyed 279 people and found measurable distress, strongest in those who had reported the strongest bond. Calling the end of a series a small breakup is about the right size of claim.
But the popular version overreaches. It is often said that the brain files fictional people exactly where it files friends. The sharpest study on precisely that question found the opposite: there is a boundary inside the medial prefrontal cortex between real and fictional others, and fictional others activate it less. What makes the finding interesting is the exception — that boundary was measurably blurred in lonelier people.
There is also brain-imaging work showing viewers’ activity moving in step while watching the same film. It is real, and it was five subjects in 2004. Worth mentioning; not worth leaning on.
In April 2017 Reed Hastings, then running Netflix, said this on an earnings call: “when you watch a show from Netflix and you get addicted to it, you stay up late at night. We’re competing with sleep, on the margin. And so, it’s a very large pool of time.”
The phrase that usually gets dropped is on the margin, and leaving it out is what makes the quote look like a confession rather than a description of a market. It is worth quoting whole. It is still a remarkable thing to say out loud.
Does binge-watching actually cost you sleep? The honest answer is: probably, and the evidence is weaker than you would like. A study of 423 young adults linked binge viewing specifically — not ordinary viewing — to poorer sleep quality and more insomnia symptoms, with pre-sleep arousal as the mechanism. But it was a cross-sectional survey, as was every study in the 2022 meta-analysis on binge-watching and mental health. Those found small-to-moderate correlations with stress, anxiety and loneliness, and cannot tell you the direction. People who sleep badly may simply watch more.
Turn autoplay off. This is the only intervention here with a randomised controlled trial behind it: twenty-one fewer minutes a day, seventeen fewer per session. One decision, made once, in a settings menu. Nothing else on this list is close.
Decide in advance what the trigger is. If-then plans — “when the credits roll, I stand up” — have been tested across 94 independent studies with a medium-to-large effect on actually doing the thing. Far better than resolving to watch less.
Be sceptical of screen-time apps. The systematic review of digital self-control tools concludes there is still limited understanding of whether they work, and notes the obvious conflict: they are asking the platform to restrain itself. This is the weakest item on the list and usually the first one recommended.
One tactic gets repeated a lot — stop in the middle of an episode rather than at the end, so there is no cliffhanger to answer. It is reasonable, and there is no study of it. It also runs into the one effect that did replicate: interrupted things get resumed 67% of the time. You may be trading an easier exit tonight for a stronger pull tomorrow.
None of this is an accident, and it is not only essayists who have noticed. In 2019 a bipartisan bill in the US Senate proposed barring large platforms from interfaces that intentionally impair user autonomy, and named autoplay specifically.
The countdown is five seconds. The switch is in the menu. That is the whole fight.
Short detours through the parts of being human that nobody explains.
Subscribe on YouTube