Picture this: a friend undergoes testing for a serious medical condition. For fourteen days, you consume everything—articles, explainer videos, treatment headlines. The news comes back positive. Both you and your friend move forward. Yet your feed keeps surfacing the same medical content, locked into the pattern of your two-week research sprint.
Why should algorithmic systems grant more weight to what you did last month than to what you explicitly request today?
Studying news recommender systems reveals a fundamental tension. These tools do serve a purpose—they surface reporting and perspectives we might otherwise overlook. But they operate by reading behavioral traces, and a trace is not the same as intent. Finishing a video could mean you enjoyed it, that it unsettled you, or that you were fact-checking. Watch duration alone cannot reliably parse those motivations. It remains just one signal among many, and the gap between attention and intention persists.
Platforms Expand Direct User Input
Recognizing this limitation, platforms have begun offering more granular tools. On October 13, Meta unveiled Facebook controls that accept descriptions of interests and user feedback in plain language. Meta indicates that recommendations can shift within a single session, though the company notes these instructions combine with existing algorithmic signals. Instagram's Your Algorithm feature provides topic controls spanning Feed, Reels and Explore. TikTok's Manage Topics and keyword filters present alternative mechanisms for steering recommendations.
These moves raise an urgent follow-up: what occurs when a user's explicit instruction contradicts their subsequent behavior?
The Reset Problem
A broad reset function addresses a different problem altogether. Instagram's reset feature works by allowing recommendations to re-personalize based on later activity. Users can review and unfollow accounts if they wish, but resetting does not force them to sever chosen connections. For someone seeking a clean slate, the relearning process may feel refreshing. Yet someone who removes a single inferred interest might reasonably expect that particular judgment to remain gone.
Here lies a crucial distinction: the difference between teaching a profile through accumulated behavior versus correcting it through deliberate action. "I no longer want recommendations about this" should carry a specific, enforceable meaning. A system that treats such an instruction as merely another data point leaves users uncertain about when it might lose force. A person can legitimately want the system to discover new interests while simultaneously demanding that it stop reconstructing one they have explicitly rejected.
This distinction carries particular weight for news consumption. A crisis can demand our attention without dictating what we wish to understand in subsequent weeks. A reader might want to continue following a local news outlet while receiving fewer unsolicited catastrophe videos. Source, subject matter, and reporting style represent distinct dimensions; meaningful controls should let readers separate them.
What Should Correction Actually Do?
The phrase "I don't want this recommended" requires clear boundaries around scope and duration. Lowering a topic's frequency differs from pausing it for thirty days, which differs from excluding it until the user reverses the decision. TikTok explicitly states that its topic controls do not eliminate subjects entirely. Such nuances should be transparent wherever users make choices. A specified pause should not quietly expire because someone encounters one tangential article.
The inverse command—"I want to keep this"—deserves equal clarity. Modifying unwanted recommendations should maintain deliberately followed accounts unless the user requests otherwise. Suppressing recommended disaster videos need not block someone from discovering a newspaper's reporting through search. Because topics overlap and categorization will always be imperfect, interfaces should disclose how they interpreted a request and permit users to refine that interpretation.
A third element completes the picture: transparency about what comes next. What fills the void—followed sources, chosen topics, broadly popular material or some combination? Even a refreshed feed operates according to selection rules. Users should grasp those rules and possess meaningful options for what appears afterward, without needing to adjust dozens of settings.
What European Law Requires
The Digital Services Act provides a foundation. Article 27 mandates that covered platforms disclose their recommender systems' principal parameters and the mechanisms users can employ to influence them. Where multiple ordering options exist, selecting and modifying the preferred option must remain directly accessible. Article 38 requires designated very large platforms and search engines to supply at least one recommender option per system that avoids profiling. Neither provision imposes a blanket reset requirement.
European policy already extends beyond a simple toggle. The Commission's 2025 guidance on protecting minors advocates for complete, permanent feed resets paired with feedback that is swift, direct and durable. Though non-binding, this guidance shapes compliance assessments under the DSA's protection obligations. Its emphasis on durability offers a valuable principle to expand across all user groups.
The Commission's Digital Fairness Act initiative, still in development, includes examination of unfair personalization. This presents an opening to establish additional safeguards. Such measures should accompany enforcement of current obligations, with a sharp line drawn between what existing law already demands and what fresh legislation should introduce.
Three Core Requirements
Three principles should anchor any framework. First, an explicit pause or removal should persist for its stated duration unless the user actively reverses it. Second, platforms should furnish a focused mechanism to suppress an inferred interest while retaining useful preferences. Third, they should clarify how recommendations are assembled afterward. Gentler "show less" options can remain adaptive, provided their meaning stays transparent.
Earlier analysis has outlined why a simple on-off switch falls short, and Peter Chapman has investigated whether audits capture genuine changes. The logical next step involves testing corrections across time. Does a blocked topic resurface before the agreed endpoint? Does "show less" actually lower its frequency against a reasonable benchmark? Can users comprehend the replacement feed and adjust their choices again?
Such testing would evaluate whether controls fulfill their stated promises. It would not demonstrate that resets solve polarization or boost mental health—those claims demand independent proof. Better controls also cannot substitute for sound defaults for users who never access settings.
The Conversational Turn
Controls are shifting toward conversation. Meta reports that large language models now interpret what Facebook users write, and Threads permits users to address its algorithm in public posts. This lowers friction, since "keep my local newspaper, but fewer disaster videos" proves simpler to state than to configure through menus. Yet like a chatbot, a written request gets weighed against everything the system has inferred from past conduct. Meta's announcement does not clarify how long such a request remains valid or whether users can observe how it was understood. On Threads, such requests expire after one, three or seven days—suitable for a fleeting impulse but not a lasting preference. Conversational control makes these requirements more urgent, not less.
A strong recommendation system helps me locate something I seek right now. It should also honor my declaration that yesterday's curiosity no longer applies. I should be able to investigate a subject, leave it behind, and return whenever I choose.



