TikTok Gave Users a Placebo Feed Reset, New York Alleges

Every A/B test has a control group. The question New York is putting to TikTok is what it means when the control group is made of people who went into their settings, found a tool meant to get unwanted content out of their feed, pressed the button, and got nothing.

That is the allegation in a newly unsealed version of New York's lawsuit against the company, reported by Reuters on October 7. According to the filing, TikTok ran what it describes as "ghost" or placebo experiments in 2023 on thousands of users, including teens and children, who tried to turn on Algo Refresh. Those users believed the feature was on. Their feeds did not change. Users given the working version saw their recommendations reset. The lawsuit says the purpose of the comparison was to measure time spent on the app and the effect on ad revenue.

TikTok's answer is that this is how software gets built. "We continuously make improvements to our platform," a spokesperson said, adding that the company regularly tests new products and features to see how they work in the real world. TikTok says the allegations mischaracterize how Algo Refresh works, and it has broadly denied New York's claims in court filings.

What the button was supposed to do

TikTok announced the feature in March 2023, after testing it earlier that year. TechCrunch described it at launch as a way to make the For You feed show content as if the user had just signed up, with the recommendation system then learning from new interactions. It lives under Settings, then Content preferences, then "Refresh your For You feed," and TikTok warned users that the action could not be undone.

In practice, it was pitched as an exit. A feed that has learned to serve sadness, extreme diets or anything else a user wants out of has one switch for starting over. TikTok said at the time, according to TechCrunch, that its systems were already "looking for repetition among themes like sadness or extreme diets."

The lawsuit makes two claims about that switch. First, that for the placebo group it did nothing. Second, that even for everyone else, TikTok misdescribed it: the filing says the company told users Algo Refresh would reset their recommendation algorithm when the effect was only temporary.

The objection came from inside

The most damaging detail is not the test design. It is that someone at TikTok said, in writing, that the test design was a problem. According to the Reuters account of the complaint, a product manager on the Digital Well Being team called the placebo tests on Algo Refresh problematic and warned that users might find them misleading. In an internal chat, the manager wrote that the feature was about transparency and control over the algorithm, and that "a placebo group completely conflicts with that."

That framing is the heart of New York's case. A placebo arm makes sense when you are testing whether a ranking tweak changes engagement and the user never knows a tweak exists. It is a different thing when the feature is a control the user deliberately chose, and its entire value rests on doing what it says. A user who presses a reset button and keeps seeing the same content is not a neutral data point. They may reasonably conclude that the reset does not work, or that the content is simply what TikTok is, and stop trying.

The complaint also claims TikTok told users that safety was a top priority and that the app was appropriate for kids and teens while knowingly blocking some of them from safety tools, including people actively trying to avoid harmful content.

Not the first experiment of its kind

The Algo Refresh allegations land on top of an earlier, more serious one. In August, Bloomberg Businessweek reported on a confidential 2023 TikTok document describing an experiment in which the company switched off an algorithmic safeguard designed to break up echo chambers of harmful content for a control group of 10% of US users. Fortune put that group at about 15 million people.

One of them, according to the document, was Chase Nasca, a 16-year-old from Bayport, New York, randomly selected on January 25, 2022. He died by suicide within a month. The document says his account was fed thousands of videos about sadness, hopelessness, loneliness and suicide, and that "TikTok's filter bubble prevention strategies did not take effect on this user by design." It described the decision as "a delicate balance across safety and the ability to measure impact on DAU" and core metrics.

New York's lawsuit now references that case too, saying internal reviews found the boy had been shown an "onslaught" of suicide and self-harm content. Senators Marsha Blackburn and Richard Blumenthal called the earlier experiment "depraved" in an August letter to TikTok's leadership and asked whether the company had ever held back a safety feature because of its effect on engagement or ad revenue. TikTok told Businessweek it is deeply committed to the safety of its users, especially teens.

The two experiments run in mirror image. In one, TikTok allegedly removed a protection users did not know they had. In the other, it allegedly withheld a protection users explicitly asked for. In both, according to the filings, the metric on the other side of the scale was usage.

Where the case goes

New York Attorney General Letitia James sued TikTok in 2024. The suit is one of more than two dozen state cases accusing the company of designing its app to addict children and misleading the public about safety. Many of the documents in those cases have been sealed or heavily redacted, which is why details like these have arrived in drips. The test claims appeared in an updated complaint filed under seal at the end of August, which a judge later ordered unsealed.

The legal backdrop is shifting fast. In September, TikTok settled the first state case scheduled for trial, agreeing to pay Alabama up to $300 million and to change how its platform works for teens there.

These remain allegations, and TikTok will argue that randomized testing is standard practice across the industry, which it is. The harder question for the company is narrower. Testing whether a feature changes behavior is routine. Letting a user believe a feature is on when it is not is a representation, and representations are what consumer-protection law is built to police.