TikTok Withheld Safer Algorithm From 15M Users, Bloomberg Says

A confidential 2021 TikTok document, reported today by Bloomberg, says the company built a safer version of its recommendation algorithm and then chose not to roll it out to everyone. Instead, about 10% of US users, roughly 15 million people at the time, were kept on the old version as a control group so the company could see whether the change would reduce the app's stickiness.

The reason that framing matters is the framing itself. This is not the usual story of a platform being slow to notice a harm and then reluctant to fix it. According to Bloomberg's reading of the document, the fix existed, was built, and was reportedly kept from many users "by design" so the two arms could be compared on engagement. In an A/B test, the group that does not get the treatment is the group carrying whatever the treatment was designed to prevent.

What lifts this out of an engineering-ethics anecdote is who Bloomberg says was in the control arm. The document reportedly includes Chase Nasca, the 16-year-old whose parents later sued TikTok after his suicide on February 18, 2022. Their complaint alleged the app served him more than 1,000 videos related to suicide, hopelessness and self-harm on his For You page even though he never searched for those terms. The New York Supreme Court dismissed the case (Nasca v. Bytedance) on April 14, 2025, on the ground that the claims targeted TikTok's exercise of a publisher's traditional editorial functions. An internal document that shows a safer model was built and withheld is exactly the kind of evidence that changes the intent picture in the next round of similar suits.

The honest caveat is that Bloomberg's account rests on a single confidential document; take the specifics as reported, not as a settled record. The reporting does not tell us what the safer algorithm actually changed, how much engagement it cost, when or whether it was eventually rolled out to everyone, or who inside the company signed off on running the experiment that way.

The forward-looking piece is that this is now a template question. Every recommender-driven platform runs A/B tests on ranking changes, safety features included, and every one of those tests has a control group by definition. Plaintiffs, regulators and auditors will read this and ask the same thing of Meta, YouTube, Snap and TikTok's new US owner: show us the safety experiments where the less-safe arm was left running, and tell us who was in it.