Published on 2026-07-25
The recommendation rabbit hole: how far can 'up next' actually take you?
You clicked one video an hour ago. Now you're somewhere strange. Here is what research says about the slope, and how to step off it.
You clicked one video about, say, sourdough bread. An hour later you're deep into a stranger's theory about grain shortages, and you genuinely cannot reconstruct the path that got you there. Nobody dragged you. You clicked 'next' every time, and every time it felt like an easy, reasonable choice. That gap between where you started and where you landed has a name now: the rabbit hole. What research actually finds inside it is stranger than the scary version everyone repeats.
The slope isn't random, it's tuned
Autoplay and 'up next' panels aren't guessing. They're built to maximise one thing: how long you keep watching. A recommendation that holds you for another eight minutes scores better than one you'd click away from in ten seconds, regardless of which one taught you more or left you better off. So the system nudges toward whatever kept people similar to you glued last time, and content that's a little more intense, a little more extreme, a little more 'you won't believe this' tends to win that particular contest more often than the calm, ordinary version of the same topic. Nobody programmed it to prefer intensity for its own sake. It just learned, from millions of sessions, that intensity is what tends to keep the next click coming.
How often it actually happens
In 2021, the Mozilla Foundation ran the largest independent look at this to date, using a browser tool called RegretsReporter that 37,380 volunteers installed to flag videos they wished they'd never seen. Across those reports, spanning 91 countries between July 2020 and May 2021, one figure stood out: recommended videos were 40 percent more likely to be flagged as regrettable than videos people had deliberately searched for. That matters because recommendations aren't a side feature. By the platform's own numbers, they drive around 70 percent of all watch time, an estimated 700 million hours every day. When the thing doing 70 percent of the steering is also the thing most likely to land you somewhere you regret, the slope is not an urban legend.
- open your watch history and scan it for the moment things turned
- use 'not interested' or 'don't recommend this channel' instead of just scrolling past
- clear or pause your watch history occasionally to give the model less to work with
- start the next session from a search, not from whatever autoplay queues up
But the drop is usually gentler than the myth
Here's the honest complication. A 2025 study in the Proceedings of the National Academy of Sciences, run by researchers at the University of Pennsylvania's Computational Social Science Lab with collaborators from Harvard, MIT, Princeton and Washington University, tested this directly instead of just assuming it. Across four experiments with roughly 9,000 participants, people were randomly given either a slanted or a balanced recommendation feed on a YouTube-like platform and watched for about 23 minutes before answering questions about their views. In most of the experiments, the recommendations barely moved anyone's opinions at all. One exception showed up: viewers already leaning a certain way drifted a little further in that direction, not the reverse for everyone else. The researchers' own conclusion was that the rabbit hole, over a short session, was not the extremizing machine it's often described as.
Put the two findings side by side and the picture gets more useful, not more confusing: the recommendation system really is tuned toward whatever holds attention, and that tuning really does surface more regrettable content than searching does. But most single sessions don't drag most people to a wildly different place. What seems to matter more is repetition over weeks and months, and how much a person already leaned that way before they ever clicked play. A single evening rarely rewires you. A year of evenings, all nudged the same direction, is a different question, and it's the one worth paying attention to.
The system isn't steering you off a cliff in one session. It's nudging you, gently, every single day, toward whatever kept people like you watching the longest.
Taking the wheel back is mostly mundane
None of this requires deleting an app or swearing off video forever. It requires noticing that 'up next' is a suggestion tuned for a business goal, not a curriculum tuned for you, and treating it accordingly. Glance at where a long session actually ended up compared to where it started. If the gap surprises you, that's useful information, not a reason to panic. The slope is real. It's also one you can step off, on purpose, any time you choose to click 'search' instead of letting the next video simply start. The bread recipe was never the problem. Letting the app pick everything after it, session after session, without ever checking where the picks are trending, is the part worth changing.
Sources
- YouTube Regrets: Findings (Mozilla Foundation) (mozillafoundation.org)
- Study challenges YouTube's rabbit hole effect on polarization (Computational Social Science Lab, University of Pennsylvania) (css.seas.upenn.edu)