What your video feed learns when you keep watching

Imagine spending an evening watching videos about wedding cakes because you are helping a friend. The next day your feed still looks like you are planning a wedding. Nothing mysterious has to happen: the system has evidence that cake videos held your attention, while the reason for that attention was never supplied. This is an illustrative scenario, not a test of either platform.

Recommendation systems make that inference repeatedly. TikTok’s current documentation names watching in full, skipping, liking, sharing and commenting among the signals that can influence its For You feed. It also uses information about the content and the user’s settings or circumstances. For most users, it says interactions, which can include watch time, generally receive greater weight than other factors.[1]

That is evidence about behaviour, not a complete account of what someone values. Watching a video can mean enjoying it, checking whether a claim is credible or trying to understand something a relative sent. In our wedding example, the system sees the viewing activity more directly than the temporary purpose behind it.

Different screens can answer different prediction questions

YouTube says it compares a person’s viewing habits with those of similar viewers. Its listed signals include watch and search history, subscriptions, likes, dislikes, explicit negative feedback and satisfaction surveys. The signals also have different roles across the product: watch history is especially important on Home, while the video currently playing is the main signal for what appears next.[2]

This helps explain why a single session and a longer viewing history can pull in different directions. A person who normally watches gardening may still receive a cake-related next video during the wedding research. The product is choosing something relevant to a particular moment, as well as drawing on past activity. That is our interpretation of the documented distinction, rather than a claim about any fixed ranking formula.[2]

Time spent is a clue, but satisfaction is another target

YouTube’s published account of its recommendation system describes an effort to move beyond clicks and raw watch time. In that 2021 explanation, satisfaction surveys asked viewers to rate videos; ratings of four or five on a five-point scale counted towards a measure of valued watch time. Models also predicted responses for videos that a viewer had not rated, with predictions checked against held-out survey responses.[3]

The interesting idea is that staying and feeling the time was worthwhile are different outcomes. A platform can train models to estimate both. YouTube’s current help page still lists satisfaction surveys as a recommendation signal; it does not publish one universal set of weights that lets a viewer calculate why every particular video appeared.[2]

A feed also helps create the evidence it learns from

There is a feedback problem beneath personalisation. A recommender first chooses what someone can easily see. Their reaction then becomes data that may influence later recommendations. Items never shown have fewer opportunities to generate a response. A system that treats the resulting activity as if people had chosen freely from the entire catalogue can learn partly from its own earlier choices.[4]

Research by Allison Chaney, Brandon Stewart and Barbara Engelhardt examined this mechanism in simulations, published at RecSys 2018. In their simulated environments, repeatedly learning from recommendation-influenced behaviour could make consumption more similar across users without a matching improvement in user value. These were controlled simulations, not an audit demonstrating that every TikTok or YouTube feed narrows, or that a specific topic inevitably leads to another.[4]

In the cake example, more cake videos create more opportunities to watch them. The system’s evidence is partly shaped by what it put within reach. A practical question follows: which action communicates the correction you actually want?[4]

Match the control to the unwanted signal

On YouTube, the menu beside a recommendation provides Not interested, and eligible menus also offer Don’t recommend channel. YouTube’s help page additionally suggests pausing watch or search history for one-off research that you do not want to influence recommendations. These controls address different information: rejecting a suggestion and changing the history available to the system are separate actions.[5]

For the iPhone and iPad route documented by YouTube, open your profile picture, then Settings and Manage all history. An X beside a video removes that item. Take care with broader deletion: YouTube says deleting watch history for a selected period also deletes search history for that period. Choosing a single mistaken entry is therefore a materially different action from clearing a whole day.[6]

TikTok’s help instructions place Manage topics under Profile, Menu, Settings and privacy, then Content preferences. Move a topic’s slider and save. The change concerns recommendations in For You, not the Following feed, profile or inbox, and may take time to affect recommendations. TikTok’s announcement explains that the sliders change topic frequency rather than block a category completely.[8][7]

For a particular TikTok recommendation, the current For You help page describes Share followed by Why this post. The explanation may point to activity such as watching similar posts or following the creator. That is useful context, although a displayed reason is not a full reconstruction of the model’s calculation. The same help page also describes Not interested as a way to receive fewer similar posts.[9]

A useful goal is to make the feed better represent what you want to spend time on now. That can involve supplying a direct preference, removing a misleading event or deliberately looking for something outside the current stream. None requires treating the feed as a verdict on your identity. It is a changing set of predictions, informed by signals you can partly shape.

Report an error or suggest a correction