§ L.0The shape of the field
The literature divides cleanly on a question it rarely states plainly: not whether operators build instrumented feedback loops to steer behaviour — that is uncontested — but whether those loops reliably produce, at population scale, the changes their operators intend. On that question the evidence is asymmetric. It is strong and experimental where the intervention is a discrete interface choice, and genuinely unsettled where the intervention is an engagement-ranked feed acting on something as diffuse as adolescent well-being. A faithful review has to hold both findings at once.
The literature strongly supports that the apparatus exists and is engineered to modify behaviour; strongly supports that it works in controlled interface experiments; but shows no settled consensus on the effect size of engagement-ranked feeds on population-scale well-being.
§ L.1Foundations & theory
The field's conceptual scaffolding is older than the platforms. Its load-bearing move is to treat instrumented design as deliberate behavioural engineering rather than neutral tooling. The behaviourist root is B. F. Skinner's work on operant conditioning, and in particular the finding that variable-ratio reinforcement — reward delivered on an unpredictable number of responses — produces the highest, most extinction-resistant response rates. That is the slot-machine schedule, and it reappears, undisguised, in the design literature on feeds.
Natasha Dow Schüll's Addiction by Design: Machine Gambling in Las Vegas (Princeton University Press, 2012) is the canonical bridge from that tradition to machine-mediated design: a roughly fifteen-year ethnography of how slot-machine design engineers a continuous "machine zone" of play. It remains the most-cited demonstration that addictive engagement is a design target, not a side effect.
The persuasion strand runs through B. J. Fogg's Persuasive Technology (Morgan Kaufmann, 2003), which founded "captology" — the study of computers as persuasive technology — and its trade descendant, Nir Eyal's Hooked: How to Build Habit-Forming Products (Portfolio, 2014), whose "Hook Model" (trigger → action → variable reward → investment) reads in retrospect as much a confession of intent as an instruction manual.
The political-economy strand reframes all of this as an extractive system. Shoshana Zuboff's The Age of Surveillance Capitalism (PublicAffairs, 2019) names behavioural surplus — the exhaust of user behaviour — as the actual commodity, traded in "behavioural futures markets." Nick Couldry and Ulises Mejias's The Costs of Connection (Stanford University Press, 2019) advances the data colonialism thesis: that "the historic appropriation of land, bodies, and natural resources is mirrored today in this new era of pervasive datafication," in which apps and platforms "capture and translate our lives into data… and [sell it] back to us," framed explicitly as "designs for controlling our lives."
This strand broadly agrees that the techniques are intentional and extractive. It disagrees on register — "addiction," "persuasion," "surveillance," "colonialism" carry different causal and moral weight — and it is largely theoretical rather than effect-measuring. Which is precisely why the empirical strands below became the battleground.
§ L.2Consumer platforms — the empirical core
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.
§ L.3Algorithmic management & gig labour
This strand establishes — qualitatively and legally — that platforms engineer feedback loops as labour control, while framing workers as autonomous entrepreneurs.
The foundational empirical work is Alex Rosenblat and Luke Stark, "Algorithmic Labor and Information Asymmetries: A Case Study of Uber's Drivers" (International Journal of Communication 10, 2016): a nine-month qualitative study finding that "Uber does leverage significant indirect control over how drivers do their jobs," rooted in "the information and power asymmetries produced by the Uber application" — "soft" control beneath an entrepreneurial veneer. Rosenblat's book Uberland (University of California Press, 2018) extends this ethnographically.
The sharpest recent contribution is Veena Dubal, "On Algorithmic Wage Discrimination" (Columbia Law Review 123(7), 2023), defining the practice as workers "paid different hourly wages — calculated with ever-changing formulas using granular data on location, individual behavior, demand, supply, or other factors — for broadly similar work," where firms "calculate the exact wage rates necessary to incentivize desired behaviors, while workers can only guess." It imports price discrimination into the wage relation. The mechanism is contested — Uber denies per-driver fare personalisation; Lyft called the paper "biased" — so it is best read as Dubal's well-cited but disputed thesis, the gig-labour analogue of the platform "does-it-work" dispute.
Surrounding this core: Mateescu & Nguyen's Data & Society explainer on algorithmic management (2019) typologises the techniques across sectors; Mary Gray and Siddharth Suri's Ghost Work (Houghton Mifflin Harcourt, 2019) documents the invisibility of platform-mediated labour; and Vallas & Schor, "What Do Platforms Do? Understanding the Gig Economy" (Annual Review of Sociology, 2020) provides the cross-sector synthesis. On the legal side, Uber BV v Aslam [2021] UKSC 5 is the judicial acknowledgement of the control relationship, the UK Supreme Court holding drivers to be "workers."
§ L.4State & compliance systems
This strand documents the public sector importing the private sector's conditioning techniques — and is where "operator-intended behaviour change" most often shades into documented system failure. Virginia Eubanks's Automating Inequality (St. Martin's Press, 2018) is canonical: digital welfare tools perform the rationing that legislatures decline to authorise openly. Australia's Robodebt scheme — automated income-averaging debt recovery later ruled unlawful — is documented in the Royal Commission into the Robodebt Scheme, Final Report (2023), whose value is its record of what officials knew, and when: the system continued because, by the operator's metric, it was "working" — it raised debts and recovered money — even as the harms accumulated off-dashboard.
In Europe, the Dutch SyRI welfare-fraud risk system was struck down on human-rights grounds in NJCM et al. v. The Netherlands (District Court of The Hague, 2020); see the analysis in the Human Rights Law Review (22(2), 2022). Sarah Brayne's Predict and Surveil (Oxford University Press, 2020) is the leading ethnography of predictive policing, drawn from the LAPD. On immigration, reporting on SmartLINK and ICE's "Alternatives to Detention" programme documents the check-in app as a compliance instrument, and the AI Now Institute's reports (2017–2022) survey public-sector deployment and accountability.
A distinctive tension lives here. What makes the state case qualitatively different — the loop aimed at a captive population that cannot uninstall — is also where the "it reliably works" thesis is most strained: the harms are frequently logged as errors and false positives, not as smoothly achieved operator-intended conditioning. That is itself a finding about the limits of the thesis.
§ L.5The central dispute, stated plainly
Read across the strands, the disagreement is not really about whether the apparatus exists or whether anyone intends it. It is about magnitude, and about where the burden of proof sits. The skeptic pole treats small measured effects as the honest ceiling of what we know; the harm pole treats them as a methodological floor, suppressed by the choices of measurement. Both are defensible. Neither has won.
The working paper's own position — set out in the Method — sides with neither effect-size estimate directly. It makes the narrower argument: that the cumulative loop, run thousands of times per subject over years against a metric the operator keeps tuning, is a different object from any single intervention an academic study can isolate, and that the operators' internal numbers, where they have leaked, are not modest.
§ L.6Gaps & open questions
Four questions remain genuinely open across the literature:
1. Effect-size resolution. Can the Orben/Przybylski–Haidt dispute be settled by methodological convergence — agreed social-media-specific exposure measures, objective rather than self-reported screen time, sex- and puberty-stratified analysis — or is it irreducibly a disagreement about which magnitude counts as policy-relevant?
2. Causation versus correlation. Where trends and correlations are agreed, what experimental or quasi-experimental designs could establish that engagement-ranked feeds cause population-scale affective change rather than merely correlate with it?
3. Mechanism opacity in labour markets. Is Dubal's individualised-wage contention empirically verifiable when the pay formulas are proprietary and firms deny per-worker personalisation — and what disclosure or audit regime would resolve it?
4. Does the thesis survive the state strand? Where harms present as failures and false positives rather than as successful operator-intended change, does "it reliably works at population scale" hold — or does it need narrowing to "the apparatus is built and, in controlled cases, demonstrably effective"?