Imagine a friend is being tested for a serious illness. For two weeks you read everything about it, watch explainer videos and open headlines promising new treatments. Then the results are reassuring. Your friend moves on. You move on. Your feed may keep bringing the illness back, because for two weeks you behaved like someone who wanted to know about it.

Why should what we did last month carry more authority than what we deliberately ask for today?

I research news recommender systems, and I want to be fair to them. They help us find reporting and ideas we would otherwise miss. But they interpret traces of our behavior, and a trace is not an explanation. You can finish a video because you enjoy it, because it frightens you, or because you are checking whether it is true. Watch time alone cannot reliably distinguish those reasons. It is one signal among many, yet the distinction between attention and intention remains.

The platforms are giving people more direct ways to respond. On October 13, Meta announced Facebook controls that accept descriptions of interests and feedback in users’ own words. Meta says recommendations can change within the same session, while explaining that those instructions are combined with existing signals. Instagram’s Your Algorithm offers topic controls across Feed, Reels and Explore. TikTok’s Manage Topics and keyword filters offer different ways to adjust recommendations.

Course correcting your recommendations

These developments make the next question more pressing: what happens when an instruction conflicts with subsequent behavior?

A general reset answers a different question. Instagram’s reset announcement explains that recommendations begin personalizing again through later interactions. Reviewing and unfollowing accounts is optional; resetting does not require abandoning chosen connections. For someone wanting a fresh start, relearning may be welcome. But a person who removes one inferred interest may expect that particular judgment to stay removed.

That is the useful distinction between training a profile through behavior and correcting it deliberately. “I no longer want recommendations about this” should have a defined effect. A system that treats the instruction merely as another signal leaves the user unsure when it might lose its force. Someone can reasonably want the system to learn new interests while asking it to stop rebuilding one they have explicitly rejected.

The distinction matters for news. A crisis can command our attention without defining what we want to understand next month. A reader may want to keep following a local newspaper while receiving fewer unsolicited disaster videos. The source, the subject, and the style of coverage are different things; a useful control should help the reader distinguish them.

What do users want?

So what should correction allow? “I don’t want this recommended” needs a clear scope and duration. Reducing a topic’s frequency, pausing it for a month and excluding it until further notice are different requests. TikTok explicitly says its topic controls do not eliminate subjects entirely. Such distinctions should be apparent wherever people make a choice. A specified pause should not silently expire because someone reads one related article.

“I want to keep this” should be equally clear. Changing unsolicited recommendations should preserve deliberately followed accounts unless the person asks otherwise. Pausing recommended disaster videos need not prevent someone finding a newspaper’s coverage through search. Topics overlap and classification will be imperfect, so the interface should explain how it interpreted a request and let the user correct that interpretation.

“Tell me what happens now” completes the request. What will fill the space: followed sources, selected topics, broadly popular content or some mixture? A fresh feed still has rules for choosing what appears. People should understand those rules and have meaningful options for what comes next, without being required to manage every setting themselves.

Regulating for a better user experience

European law provides a starting point. Article 27 of the Digital Services Act requires covered platforms to explain their recommenders’ main parameters and available ways to influence them. Where several ordering options are available, choosing and changing the preferred option must be directly accessible. Article 38 requires designated very large platforms and search engines to offer at least one option for each recommender system that does not rely on profiling. Neither provision amounts to a general requirement for a full reset.

European policy already goes beyond a simple on-off choice. The Commission’s 2025 guidelines for protecting minors recommend complete, permanent feed resets and feedback with swift, direct and lasting effects. The guidance is voluntary and informs assessments of compliance with the DSA’s protection duties. Its attention to persistence is a useful principle to develop for all users.

The Commission’s Digital Fairness Act initiative remains under preparation, including work on unfair personalization. It offers an opportunity to specify additional protections. That should complement enforcement of existing duties, with a clear distinction between what the law already requires and what new legislation should add.

I would make three requirements central. An explicit pause or removal should remain effective for its stated duration unless the user deliberately changes it. Platforms should offer a targeted way to suspend an inferred interest while preserving useful preferences. And they should explain how recommendations are selected afterward. Softer “show less” preferences can remain adaptive, provided their meaning is clear.

Previous arguments in Tech Policy Press have explained why an on-off switch is insufficient, and Peter Chapman has examined whether audits test actual changes. The next step is to test corrections over time. Does an excluded topic return before the agreed endpoint? Does “show less” reduce its frequency against a suitable baseline? Can people understand the replacement feed and change their choices again?

Those tests would assess whether controls honor their promises. They would not establish that resets cure polarization or improve mental health; those claims need separate evidence. Better controls also cannot replace safe defaults for people who never open the settings.

Controls are also becoming conversational: Meta says large language models now interpret what Facebook users write, and Threads lets people address its algorithm in a public post. This lowers the barrier, since “keep my local newspaper, but fewer disaster videos” is easier to say than to set in a menu. But as with a chatbot, a written request is weighed rather than obeyed, here against everything inferred from past behaviour. Meta’s announcement does not say how long such a request counts or whether users can see how it was understood. On Threads, such requests expire after one, three or seven days, which suits a passing mood better than a settled decision. Conversational control makes these requirements more pressing, not less.

A good recommendation system helps me find something I want now. It should also accommodate my saying that yesterday’s interest no longer applies. I should be able to explore a subject, leave it behind, and return on my own terms.