The feed is the experiment,and the experiment has concluded.

Strip out the brand differences and a feed-ranked platform is doing four things in a tight cycle. It shows you content. It measures what you do with that content — scrolls, dwell time, taps, shares, screenshots, the precise pixel coordinates where your thumb lifted. It updates a model of you. And it picks the next piece of content to show, choosing whichever option the model predicts will maximise some target metric, usually a weighted blend of time-in-app and a few interaction counts.

That's it. That's the whole machine. What makes it powerful is not any single decision in the loop; it's that the loop runs perhaps a hundred times per session, hundreds of millions of sessions a day, with the model continuously retraining on the results. There is no human content editor in this process. The optimiser is the editor.

Once you see the system this way, a number of mysteries resolve. The mystery of why every platform's feed slowly converges on the same texture — short, emotional, slightly outraged, visually busy — resolves into "because that texture wins the optimiser's metric." The mystery of why platforms can't simply fix the problems they cause without crippling the product resolves into "because the problems and the product are produced by the same mechanism." The mystery of why teenagers in particular seem so flattened by the experience resolves into "because the optimiser has more training data on them per unit time than on anyone else."

The strongest evidence that the conditioning works is not from critics. It is from the platforms' own internal research, which was designed to answer a much narrower question — does our product change user behaviour in the direction we want? — and which kept coming back with the answer "yes, and also several directions we don't want."

The 2021 Facebook Files made some of this public. Internal Instagram research found teen users self-reporting that the app made them feel worse about their bodies, and that they kept using it anyway because the algorithmic feed kept surfacing exactly the comparisons that had hurt them. The notable detail is not the harm. The notable detail is that the company had quantified it, internally, with the same rigour it used to quantify ad revenue, and concluded that the harm and the engagement were the same variable measured from two sides.

YouTube's recommendation team published similar work, less dramatically, in machine learning conference proceedings — papers describing how watch-time optimisation produced systematic drift toward longer, more extreme content, and how various dampening interventions affected the metric. The dampening interventions worked too. They simply traded one number for another. Engineers don't write papers about systems that don't work.

The platforms' own data scientists have been publishing the proof-of-concept for two decades. We just kept reading it as marketing.

The clearest sign that software-layer conditioning has matured into a reliable industrial process is that the gambling industry adopted it wholesale. Online sports betting and slot apps now ship with notification timing, loss-recovery mechanics, near-miss visualisations, and reward variability schedules that are functionally identical to the engagement design of consumer social platforms. They are not similar by coincidence. The same consultancies, the same playbooks, sometimes the same engineers, have moved between the two industries.

Gambling is a useful case because the metric is unambiguous. The slot app's job is to extract money. If the conditioning patterns it shares with Instagram and TikTok didn't work, the gambling sector — which is ferociously empirical about anything that moves the revenue line — would have dropped them. It hasn't. It has doubled down.

The same loop, deployed against the same neurology, with a different revenue model bolted to the end. That the loop works in both deployments is exactly the point. It is portable, generalisable, and at this stage, well understood by its operators.

When I say the conditioning works, I am being deliberately narrow. I mean: the systems reliably produce, at population scale, the behavioural changes their operators are paid to produce. They increase time in app. They increase the probability of a click, a share, a deposit. They reduce churn. They lift the metric the optimiser was pointed at. There is no longer any serious internal debate about this inside the firms that operate them.

What the systems do not do is honour any side constraints that weren't built into the metric. They do not, by default, protect adolescent mental health. They do not protect a healthy information environment. They do not, on their own, refuse to amplify the most enraging item available. These are real failures, but they are failures of the metric, not failures of the technology. The technology is doing what it was built to do. It is the metric that is impoverished.

This is why "the platforms have lost control of the algorithm" is the wrong frame. The platforms have not lost control. The platforms have exactly the control they purchased. The thing they have not done is purchase a metric that lines up with the public interest, because no one is paying them to.

This strand splits into three sub-literatures of very different evidential strength.

(a) The experimentation infrastructure — settled

Continuous large-scale controlled experimentation is ordinary industry practice. Ron Kohavi, Diane Tang and Ya Xu — experimentation leaders at Microsoft, Google and LinkedIn — report in Trustworthy Online Controlled Experiments (Cambridge University Press, 2020), and in "Online randomized controlled experiments at scale" (Trials 21:150, 2020), that those firms run "over 20,000 controlled experiments/year" (they caution that counting methods vary). The feed is, literally, a permanent A/B test.

(b) Direct efficacy evidence — settled, and strong

The cleanest proof that instrumented design produces operator-intended behaviour comes from dark-pattern experiments. Jamie Luguri and Lior Strahilevitz, "Shining a Light on Dark Patterns" (Journal of Legal Analysis 13(1):43–109, 2021), ran a nationally representative US randomized trial: mild dark patterns more than doubled sign-ups for a dubious identity-protection service versus a neutral interface; aggressive patterns roughly quadrupled them; effects compounded when stacked. The FTC's staff report Bringing Dark Patterns to Light (P214800, 2022) elevates this to the regulatory record and documents Credit Karma selecting an allegedly false "pre-approved" claim because A/B testing showed it maximised clicks. (Caveat: the doubling effect is one twice-run experiment on a single service type — robust, but generalising from one well-controlled setting.)

(c) The adolescent-mental-health controversy — the open wound

Here the literature does not converge. This is the field's live methodological war over effect sizes, and it is unresolved as of this writing.

Both camps actually agree the correlation exists and that teen mood-disorder rates rose sharply in the early 2010s. They disagree on magnitude, causation, and whether self-reported screen time is a valid measure. This is the precise locus of the "does it work at scale" question.

(d) Primary-source leaks

The Facebook Files / Frances Haugen disclosures, entered into the US House Energy & Commerce Committee record (22 September 2021), are the field's key primary documents — distinct from peer-reviewed scholarship. A March 2020 internal slide reported that "32% of teen girls said that when they felt bad about their bodies, Instagram made them feel worse"; a 2019 slide stated "We make body image issues worse for one in three teen girls"; another reported that teens "blame Instagram for increases in… anxiety and depression… unprompted and consistent across all groups." Meta disputes the framing, not the existence of the figures.

Consumer platforms are the easy case. The conditioning is visible. You can feel it in your own thumb. The reason this paper treats them as Case 01 rather than the whole story is that the same loop has now been deployed in two domains where the subjects are less free to leave: gig labour and state compliance. Those are Theme 02 and Theme 03.

The platforms taught a generation of engineers how to bend behaviour with software. Those engineers did not stay at the platforms. They went to logistics companies, fintech startups, government IT contractors, police technology vendors. They took the loop with them. The loop still works. It works on warehouse pickers. It works on benefit claimants. It works on people awaiting an immigration hearing. The mechanism does not care.