In July 2021 the Wall Street Journal published an investigation of TikTok's recommendation system that solved the methodological problem which makes most claims in this area unfalsifiable.
Rather than asking users what happened to them, they built the users.
More than a hundred automated accounts were created, each assigned a set of interests unknown to the platform, each registered with a date of birth — some as thirteen to fifteen year olds. The bots watched videos, and their only behavioral signal was dwell time: they lingered on content matching their assigned interest and scrolled past the rest. They did not like, comment, follow or search.
The system identified the assigned interest within, in most cases, under two hours, and in some cases around forty minutes. It then narrowed sharply. Accounts assigned an interest in sadness were, within hours, receiving feeds composed substantially of depression content. Accounts assigned interests adjacent to disordered eating received content on extreme dieting, and the paper documented accounts registered as minors being served drug, pornographic and self-harm-adjacent material.
Two mechanisms.
Dwell time bypasses self-report. Everything a person tells a system about themselves is filtered through self-presentation. What they linger on is not. A recommendation system reading behavioral traces can identify a latent interest — including a vulnerability the person has not articulated to themselves — faster than any survey instrument, because it is not asking.
And the short-form loop is a variable-ratio schedule with an unusually fast cycle. Most videos are not rewarding; occasionally one is; the interval is unpredictable and measured in seconds. Chapter 19's schedule, at the highest frequency any medium has achieved.
What makes this evidence rather than anecdote is the design. Identical inputs, controlled variables, reproducible funnels — the bots are a laboratory instrument, and the method has since been reused by other investigations and by researchers.
TikTok responded that the experience of bots is not representative of real users and that it had introduced tools to diversify recommendations and to limit repetitive content in sensitive categories.
The general finding, and it applies to every system in this part: a recommender does not need to know anything about you that you have chosen to disclose. It needs to know what you stop on.
Cross-ref: Defense — Detecting Behavioral Profiling