The Take — TikTok's real offence is a button that lies

The European Commission indicted an objective function on Friday. The part of the case that will still matter in two years is smaller and harder to argue with: every control TikTok built to override that objective function was tested, and every one of them lost.

That is the offence I would write the decision around. Not "engagement optimisation is illegal" — a claim Brussels cannot measure and TikTok can comply with on paper — but "you shipped a control that does not control, and your own data showed it." One of those is a precedent every feed in Europe can be held to. The other is a decade of litigation about what the word "disable" means.

The facts are on the table in this morning's brief: the Commission's preliminary findings on both risk assessment and mitigation, a potential fine of up to 6 percent of worldwide annual turnover that nobody outside ByteDance can size because neither company publishes global revenue, and a third set of TikTok findings this year after addictive design in February and minors in July.

Now the sentences that decide the case. Diversification measures "do not sufficiently reduce how often similar content is recommended." The feed-refresh control helps "only briefly," and tapping "not interested" stops working after a while, officials said at Friday's briefing, as The Record reported from Brussels. The Commission's own wording is the damning one: the ranking "prioritises showing users content based on viewing time, over their choice to see less of certain content."

Read those three together and you have something narrower and more damning than a philosophical objection to watch time. You have a product that offers the user a lever, watches the user pull it, and re-learns its way back to where it was. The user is not choosing badly. The user is being outvoted by their own thumb.

The evidence is observable, which is the whole point

The detail I keep coming back to is procedural. The Commission spent two and a half years building what its officials call "amongst the world's most advanced investigatory infrastructures," analysed roughly 20,000 videos with psychologists, psychiatrists and system-design specialists — and then said the testing "didn't really need access to the exact algorithm. We had enough information without it."

That is the difference between a finding that enforces and one that declaims. You cannot audit a loss function from the outside. You can audit a promise: here is the control, here is what the user asked for, here is what the feed did for the next thirty days. Input, output, distribution. Any regulator, researcher or journalist with a test account can run that experiment — which is why this file may become a rule rather than a press release.

The rest of the remedy list — "rely more on what users explicitly say," "disable viewing time as a recommendation signal" — reads as directions, not standards: no number, no test a compliance team can fail. Our brief flagged that on Friday, and it is the weakest joint in the case.

The strongest case for the other side

The steelman here is genuinely good, and part of it I accept.

Watch time is a better measure of what people want than what they say they want. Almost nobody presses "not interested"; almost everybody keeps watching. Obey the button and you degrade the feed for the many to satisfy the few, and the few are not a stable set — they are whoever is having a bad week. Stated preferences are also notoriously aspirational. A feed built strictly from what users declare they want becomes a museum of their intentions, and people do not open TikTok to visit a museum.

Second, "the button stops working after a while" may be a description of personalisation rather than a defect in it. Any model that learns from behaviour will drift back toward behaviour: that is what learning is. Demanding that a suppression survive forever is demanding the system stop updating for that user — a much bigger product change than it sounds, and one the DSA never says in words.

Third, and most awkward for the Commission: observational evidence of harm across a thousand legal videos is not a controlled trial. The causal chain — repetition, then sadness, then disorder — rests on internal documents, expert interviews and a literature review. Respectable, and still not proof that a different ranking would have produced a different life.

Why the take still holds

Because the duty in the DSA is about mitigation of a risk you have already identified — not about designing a perfect feed, and not about proving causation to a scientist's standard. The Record's account is explicit that TikTok had flagged some of these issues in its own risk assessments. Once you know, "we shipped a button" stops being an answer and becomes the thing being examined. The question is no longer whether the loop exists; it is whether the mitigation measurably reduced it. On the Commission's evidence, it did not.

And the fix I am arguing for is not exotic engineering. Suppression that decays over days is a design choice; suppression that holds until the user lifts it is also a design choice. Hard negatives, session-level constraints, a rule that survives the next retrain — every recommender team knows how to build these, because they build them for spam and brand safety. What they cost is engagement, concentrated in exactly the cohort where the loop is doing its work. That cost is the business model, and it is fair to put it on the table. It is not fair to rewrite the objective function, because that remedy dies in court and takes the enforceable one down with it.

There is an uglier reason to prefer the narrow theory. In August, US senators pressed TikTok over an internal experiment that withheld a mental-health safeguard from roughly 10 percent of its US users, one of whom was 16-year-old Chase Nasca, which we covered at the time. That does not prove a feed caused a death, and nobody serious should present it that way. It proves the company could measure the difference between a loop and a broken loop at population scale, and measured it by withholding the fix. "Make your controls real and show us the numbers" asks for that experiment run in reverse. "Stop weighting watch time" asks for a philosophical argument Brussels will lose.

What would change my mind

Four things, in order of weight.

If the eventual non-compliance decision makes "disable viewing time as a recommendation signal" the binding obligation rather than one option among four, I will say the Commission overreached, and I expect the measure to be annulled or to become unenforceable — at which point the good half of the case goes with it. Precision about which remedy is mandatory is the whole ballgame.

If TikTok publishes ninety-day data showing that "not interested" suppression persists for users who were not re-exposed to the topic, and the Commission's field test only measured a short window, the factual core of my argument collapses. I would still think the controls are hard to find and weakly explained — but "weakly" is not "falsely."

If a controlled experiment shows that users handed a constrained feed for a month prefer it less even after being shown what the loop was doing to them, then this is a story about preferences rather than manipulation, and the case is paternalism with a press office. I do not expect that result.

And if February's addictive-design findings are still unresolved when the KIDS Act reaches the statute book, the binding constraint is the Commission's throughput, not TikTok's design. Nine months and counting is not a good sign.

Every recommender relearns. That is not the scandal. The scandal is telling a user they are in charge of what comes next, measuring the moment they try to be, and shipping it anyway.

If the harm only exists across a thousand videos the user chose to watch, who should be allowed to overrule the choice — the platform, the regulator, or the person holding the phone? Write to us via our contact page.