A 10% holdback, 15 million people: TikTok's experiment config is now evidence

Holdbacks are ordinary product practice. You cannot put a number on what a safety feature costs without a population that does not have it, and the internal document says as much: the call was "a delicate balance across safety and the ability to measure impact on DAU (daily active users) and core metrics" ^[4].

What turns a sizing decision into a subpoena is the denominator. The safeguard, built to break up echo chambers of harmful content, was off for 10% of US users ^[2], roughly 15 million people at the time ^[3], which implies a US base of about 150 million accounts ^[1]. Because the arm was defined as a share of everyone and assignment was random ^[6], something close to one in ten American users, minors included, sat in the group without the safeguard ^[2].

The 2023 document was handed over in litigation against the largest social media companies and placed under a court-ordered seal before Bloomberg Businessweek reported it ^[5]. That is the part worth reading twice: the artifact now quoted in a four-page Senate letter ^[1] is an experiment design record, carrying its own explanation that "TikTok's filter bubble prevention strategies did not take effect on this user by design" ^[7]. Blackburn and Blumenthal, who called the decision depraved, note that Congress had raised concerns about the algorithm since October 2021 ^[1][10], about three months before the January 2022 assignment date in the document ^[3].

TikTok did not respond to a request for comment on the letter; for the Businessweek story a spokesperson said the company is deeply committed to user safety and invests in Trust and Safety detection and enforcement ^[9].

The FDA is meanwhile asking whether generative AI health products should pass competency-based tests on the final, user-facing product before marketing, with rigor scaled to the product's risk profile ^[11], and floats real-world evidence collection afterwards ^[12]. Different industry, opposite order: the record gets assembled before release rather than unsealed in discovery.