Recommendation is not knowledge of you

Someone watches three videos about starting a small business. Within days almost everything suggested is about entrepreneurship. The common conclusion is that the platform has worked out who they are.

This matters because the popular claim about recommenders is stronger than the evidence supports, and a course that repeats it will be wrong in a way its audience can eventually check.

Here is what has actually been measured. Across the browsing histories of 309,813 representative Americans and 21.4 million watched-video pageviews from 2016 to 2019, far-right content averaged 0.17 per cent of total viewing in 2016, rising to 0.30 per cent by 2019. Of arrivals at that content, about 55 per cent came from external links, the platform homepage and search, against 36 per cent from a previous video. The authors state that they find no evidence that engagement with far-right content is caused by recommendations systematically (Hosseinmardi et al., PNAS 2021).

In the largest field experiments on this question, replacing an algorithmic feed with a reverse-chronological one for three months, across tens of thousands of consenting users, changed political attitudes, polarisation and knowledge by amounts that were not detectable.

So the honest position has three parts. Recommenders demonstrably concentrate what you see. They demonstrably increase time on the platform. And the studies that manipulated feed narrowing directly, over weeks, mostly did not find the attitude change everyone assumes follows from it.

For the business-video case, then, the defensible claim is not that the platform is shaping the viewer's mind. It is narrower and more useful: after four days, this person is being shown a small neighbourhood of a large subject, chosen because three early clicks resembled other people's early clicks, and the parts of the subject they are not being shown are invisible to them. Nobody withheld anything. Nothing was decided about them personally. They are simply no longer being offered the rest.

The practical consequence is about opportunity rather than belief.

Somebody researching how to start a business through a recommender will be shown the parts of that subject that hold attention: motivation, branding, product ideas, the founder's story. The parts that do not hold attention are tax registration, contracts, the failure rate in their sector, and what a poor cashflow month looks like. Those are also the parts that determine whether the business exists in two years.

The recommender did not judge that the boring parts were less important. It measured that they were watched less, which is a fact about audiences, not about business.

This is why the correction is not to distrust the platform. It is to notice that you have stopped choosing, and to introduce one source of the same subject that the recommender did not pick. One is enough. The habit is what matters, and it takes about a minute.