When YouTube Wanted More Watch Time, It Got Longer, Worse Videos
Creators stretched videos and buried the payoff at the end, and the system rewarded them for it.
The Royal Institution · There is No Algorithm for Truth - with Tom ScottYouTube did not set out to reward bad videos. It set its recommendation system, the automated software that picks what to suggest next, to increase one number: watch time.
Science communicator Tom Scott, who has 1.9 million YouTube subscribers and usually gets between 250,000 and 1 million views per video, put it simply. If viewers stayed until the finish of a 20-minute upload, the software read that as success.
Then creators did the logical thing. They stretched running times and saved the payoff for the final minutes, obliging viewers to sit through everything to get the answer. The result, Scott said, was that weaker videos got suggested more often.
It is a textbook case of Goodhart's Law, which holds that, "When a metric becomes a target, it ceases to be a good measure." Once watch time became the prize, length stopped being proof that viewers loved a video, and started being proof that creators knew how to extend one.
YouTube then changed its public formula. The company said it favored strong clips that kept people using the site, even when the next watch came from a different maker or genre.
The machinery changed too. In 2017, Jim McFadden, who was then technical lead for YouTube recommendations, said the new system from Google Brain, Google's artificial intelligence research group, could generalize. As Scott described it, the old system might see that you watched one comedian and offer another clip just like that one. The new model, McFadden said, figures out "other comedians who are similar but not exactly the same," spotting adjacent ties and less obvious patterns.
The underlying test stayed simple. A clip counted as good if it drew someone onto the site, or if it stopped that person from clicking away and ending the session. The system takes in those signals and works out what will keep people watching clips and, what mattered most to Google, keeping viewers around for the ads shown between clips.
Measure viewing, and you change viewing. Creators did not get lazier. They got exactly as strategic as the scoreboard told them to be.
Questions people ask
Does YouTube still use watch time to rank videos?
Yes, but not alone. YouTube says it now weighs a mix of signals, including watch time, how often viewers choose a video when it is shown, likes and dislikes, and survey responses about satisfaction. The goal is to approximate videos people find valuable, not just videos they sit through.
What is Goodhart's Law and where did it come from?
It is the idea that once a measure becomes a goal, people will game it and it stops being useful. The British economist Charles Goodhart first described it for economic policy in the 1970s, noting that monetary targets broke down once governments tried to control them. It now shows up everywhere from school test scores to social media metrics.
Why does YouTube care so much about keeping people from leaving?
Because YouTube sells attention to advertisers. More time on the site means more ads can be shown between and around videos. A video that brings someone to YouTube, or stops them from closing the app, is therefore valuable to the business even if the viewer did not love it.