A design review puts two versions of a subscription screen on the table. The aggressive one pre-checks the consent box, buries cancellation three menus deep, and labels the button Continue. The mild one merely writes the choice clearly. On the dashboard, the aggressive version wins by a factor approaching four against the neutral control. Someone in the room calls it unethical; someone else points at the conversion curve and says nothing. Both are right, and that is the uncomfortable part. The aggressive variant did not fail. It worked almost four times as well. Then the complaints arrived.

In 2021, Luguri and Strahilevitz published "Shining a Light on Dark Patterns" in the Journal of Legal Analysis. Across a series of experiments, an aggressive dark-pattern subscription flow raised the likelihood of subscribing to almost four times the control condition. A mild version — equally manipulative, only quieter — raised it to more than twice. The finding that tends to vanish from the slide is the remainder: only the aggressive version produced what the authors called a "powerful backlash." The mild one did not. The extra conversion came bundled with a reaction; the milder lift did not carry the same price. That asymmetry is not a footnote. It is the whole argument about the ethics of growth compressed into one experiment.

Set the wider evidence base beside it and the conversion side and the bill side become visible at once:

Evidence anchorSource and yearWhat was measuredReported figureWhat the figure does not prove
Aggressive vs. mild subscriptionLuguri & Strahilevitz 2021, Journal of Legal AnalysisExperimental subscription likelihoodAlmost 4× aggressive, >2× mildThat only the aggressive version triggered a powerful backlash
Consumer exposure surveyEuropean Commission 2025, Consumer Conditions ScoreboardShare reporting encounters over the past year66% fake reviews, 61% misleading discounts, 93% worry about targeted ads, 45% online scamsSurvey perception, not measured behavior or causality
Post-purchase reactionREBR 2025, Romania70 participants, self-report61.4% regret, 47.1% abandoned the store, ~70% annoyance, manipulation or distrustA small, correlational sample
Trust premiumUserTesting 2025, global4,000 respondents, stated willingness68% would stay with a trusted brand through a price rise, +25% average willingness to pay; a second group 60% would pay more, 82% tolerate at most two bad experiencesStated intent, not revealed preference
Transparent-nudge meta-analysisBruns et al. 2025, ScienceDirect23 papers, 117 effect sizesPositive on behavioral outcomes, none on non-behavioral; 88% of the effect from default nudges, 89% from online studiesA general victory for transparency, rather than gains concentrated in defaults

Two of those rows deserve a slow reading. The Commission's 66% is the share of consumers who reported encountering fake reviews, not a finding that 66% of reviews are fake — a distinction that blurs fastest inside a deck. The REBR study ran on 70 people in Romania, so its 61.4% regret rate is a signal about how a dark pattern feels after the click, not a population estimate. The UserTesting numbers are stated willingness, the kind that reliably sits above what a checkout actually shows. None of that makes the reaction imaginary. It makes the measurement honest, and an honest measurement is the only basis for a decision worth defending.

The Bruns meta-analysis is the most useful of the set and the most easily misread. Transparent nudges did move behavior, but did nothing for non-behavioral outcomes such as attitudes, and 88% of the behavioral effect came specifically from default nudges. Put plainly, the gain lives in making the default legible, not in appending a paragraph of disclosure. The ethics review attached to the same paper is equally telling: across 23 frameworks, autonomy appeared in 22 (95.7%), and reflection-based nudges were better received than covert defaults. Transparency is not a feature that costs conversion. In the data it reads more like a precondition for the nudge to be defensible at all.

The temptation is to stage this as a morality play — dark patterns bad, transparency good. The evidence will not carry that. Luguri and Strahilevitz showed that the aggressive version wins on the metric a growth team is paid against. Calling it ineffective would be the comfortable lie. The honest position is that it converts and then invoices: the bill arrives as regret, abandonment, and the attention of a regulator. A mild nudge that is fully legible keeps more of the conversion than a purist expects while sidestepping the backlash the aggressive version bought. That is a tradeoff worth taking seriously, not a sermon. The literature cannot price the bill exactly, but it can stop a team from hiding behind a claim that the aggressively designed flow does not work.

The bill is also becoming literal. A Digital Fairness Act cleared the committee stage on 11 November 2026 with a proposal to ban dark patterns outright, estimated to save EU consumers up to €13.7 billion a year — a projection, not a measured saving. A revision to the CPC framework contemplates fines of up to 6% of global turnover. Those figures turn what used to be a design-review argument into a line in a risk register.

For an organization running an open directory, the practical move is unglamorous. Make the choice architecture itself inspectable: which surfaces arrive pre-selected, which defaults are opt-out, how many steps separate subscribing from stopping. Keep those surfaces versioned in the shared runtime so a mild, legible variant can be tested against an aggressive one on the same footing, and so the backlash shows up in the same dashboard as the conversion. Cross Mini App keeps that comparison reproducible from the directory side, which is the only way a claim about trust survives contact with a growth target. The aggressive flow will still win the first week. What matters is what the second week looks like.

Sources