Your 'Not Interested' Button Is Lying to Your Users

You ship preference controls so people can steer their own feed. A block button. A "see less" slider. A settings page where they tell you what they don't want. You assume those work. New reporting says they mostly don't, and for some users that gap is doing real harm. Let me catch you up on what changed and what it means for the controls you own.

The button that grades one video, not a topic

A Mashable reporter hammered TikTok's "Not Interested" button for years, thinking it was a filter against GLP-1 and weight-loss content. It isn't. As she found out the hard way, hitting "Not Interested" on one GLP-1 creator just adjusts how the algorithm scores that single video. Block that page and it serves you the next one. The button never learns you want the whole category gone.

A Northeastern study backs this up. Researchers ran hundreds of bot accounts and called "Not Interested" the most effective explicit signal a user has, then watched topics reappear minutes after tapping it. Stop signaling disinterest and "many find their feeds dominated by such content again."

The reporter changed every buried setting she could find, waited the 48 hours the app warned her about, and months later was still screenshotting Wegovy ads. The controls looked like agency. They delivered almost none.

Controls buried where nobody digs

The real preference tools existed. They were just hidden. Weight-management settings sat under an "Other" menu inside "manage ad topics," several taps deep in a profile. Health and wellness was marked "Interested" and weight management was "No preference," defaults she never chose and never saw.

There's a fairness gap worth noticing. Instagram came out looking better in the same test, not because its controls are smarter, but because its main feed shows you accounts you follow instead of an endless stream from strangers. As Mashable put it, staying on your main feed kept the problem manageable. Architecture did the work the settings page couldn't.

If your product hides its real controls behind vague labels and generous defaults, you have built the same trap. Users think they've spoken. Your system never heard them.

When the feed reads distress as a signal to feed it

Here's the part that should worry a product owner. The system doesn't just ignore preferences. It can actively work against the people most at risk. A 2024 University of Melbourne study found users with a history of disordered eating got served more toxic eating-disorder content than users without that history. The platform reads lingering on a video as interest.

Dr. Blair Burnette, who runs a body-image lab at Michigan State, called that finding "extremely disturbing." A person can steer more harmful content to their own feed just by pausing on a video or searching something unrelated. People searching for recovery content reported getting served more restrictive-eating content instead.

Engagement and wellbeing point in opposite directions here. As one clinician put it, when someone is stuck in a mental illness, the content that gets the most interaction is the content that reinforces it. Your optimization target rewards exactly the behavior a struggling user needs to break.

The gadget that promises to do the work for you

The same pattern runs through wellness hardware. A Lifehacker writer testing vagus-nerve devices found the phrase "nervous system regulation" had drifted far from anything clinical, sold as a single gadget or a five-minute hack. Neuroscientist Dr. Ramon Velazquez told her the strongest drivers of nervous system health are the boring fundamentals: sleep, movement, nutrition, connection. No device replaces those.

The design lesson is the same as the feed. A control that markets certainty it can't deliver is worse than no control. Her warning: if a "relaxation" device is causing pain, cramping, or twitching, that's a sign to stop, not push through. A product framed as safe and calming can teach users to ignore their own body's signals.

The deep cut

Stop assuming your preference controls work. Test them like the researchers did. Take a fresh account, set every control to its strongest "off," then behave like a real user and count how long the unwanted content stays gone. If it comes back in minutes, your control is theater, and your users think they said no when your system heard nothing.

Two cheap fixes fall out of this. First, expose your defaults and inferred assumptions where people can actually see and change them, the way TikTok lets you view what it thinks you are, just not on page nine of settings. Second, when your ranking signal conflicts with a stated preference, let the preference win. A pause on a video is not consent. Treating it as consent is how you end up feeding the users you're hurting most.

Three questions for your team

• When a user says "less of this," does our system suppress the topic, or just downrank one item? Prove it with a test account, not an assumption.
• Where do our real preference controls live, and how many taps does it take to reach them? If weight-management or sensitive settings sit under "Other," that's a decision we made and can undo this sprint.
• Which of our engagement signals can be triggered by distress rather than interest, and what happens to a vulnerable user when we optimize for those signals?