Reactance

Brehm's theory is that a person who believes they are free to do something, and is then told they must or must not, is pushed into a state directed at restoring the freedom — by doing the forbidden thing, liking it more, derogating the source, or denying there was a problem. Rosenberg and Siegel's fifty-year review finds that most of the progress since has been in measurement rather than mechanism. Dillard and Shen made it measurable: reactance is an amalgam of anger and negative cognitions, empirically inseparable, triggered by perceived threat to choice. Rains's meta-analysis of 20 studies found that intertwined model fitted best, with each component correlating with attitude toward the message at r = .20 and .16 (via Reynolds-Tylus 2019). That is the size of the construct — real, consistent and small: enough to change how a screen feels, not enough on its own to predict what gets clicked.

What raises it is reasonably well mapped, and what is supposed to lower it is not. Controlling language — “you must,” imperatives, mandates — raises perceived freedom threat compared with qualified phrasing, and so does framing the request as a loss. Shen's web experiment (N = 814) crossed high and low threat, gain and loss framing, choice and no choice: threat and the loss frame increased reactance, a gain frame and the presence of a behavioural choice lowered it, and the advantage of offering a choice was largest exactly where the threat was high. Give two real options rather than one instruction, and give them where the ask is heaviest. The copy remedy is weaker than its reputation: the “but you are free” postscript came out at g = 0.44 across Fillon, Souchet, Pascual and Girandola's pre-registered meta-analysis of 52 experiments and 19,528 participants, but the seven studies rated low risk of bias showed no effect at all, g = 0.11, CI [−0.18, 0.40], and the technique worked best face to face rather than in text. Freedom-restoring postscripts in health messages are similarly uneven, reducing threat in some tests and doing nothing in others (Reynolds-Tylus 2019). Treat “you can always change this later” as worth a line of copy and nothing more until your own test says otherwise.

The uncomfortable part is what happens when someone actually builds the pushy screen. Luguri and Strahilevitz ran a representative US sample (N = 1,963) through a consent flow for an identity-theft protection plan. Acceptance was 11.3% in the control, 25.8% with mild dark patterns and 41.9% with aggressive ones. Reactance was plainly there in the aggressive arm: aggravation rose from 2.96 to 3.94 on a seven-point scale, felt freedom to refuse fell from 6.21 to 4.74, and 65 participants abandoned the flow against 9 in the mild one. None of it stopped the sign-ups, and the mild patterns more than doubled acceptance with no detectable mood cost at all.

Their second study (N = 3,932) isolated single patterns, and confirmshaming — the shamed decline button — raised acceptance from 13.2% to 20.5%, repeated that gain alongside a preselected default and alongside added obstruction, and did not differ from control on mood. It is also common: Mathur and colleagues' crawl of roughly 11,000 shopping sites logged 1,818 dark pattern instances, 169 of them confirmshaming, mostly in email-capture popups where the decline read “No thanks, I like paying full price.” Their classification is the useful part — confirmshaming is asymmetric but not covert, so users can see it working on them. The case against shipping it is that it manipulates, and increasingly that it is unlawful, not that it costs you conversions.

Where resistance does cost money is when the product contradicts something the user has already decided. Fitzsimons and Lehmann had participants form impressions of granola bar formulations and then receive a recommendation. When an expert recommended against the option participants liked, they chose it more (2.71 of 3) than when the same expert recommended for it (2.12), and a non-expert source moved nothing. Where participants expected their stated choice to determine what they actually received, they picked the disparaged option 95.0% of the time against 44.4% when the choice was inconsequential, and reported lower satisfaction (5.45 against 7.37). Advice that argues with the user's own read of the options produces a contrary choice and an unhappy one, so a recommendation belongs next to their selection rather than on top of it.

Two results point the other way, toward control being worth paying for. In Tucker's randomised field experiment a nonprofit's personalized Facebook ads performed no better than generic ones until the platform gave users more granular privacy controls; after that, users were nearly twice as likely to click them, while targeted-but-not-personalized ads were unchanged. One advertiser, click-through rates of a fraction of a percent and a policy change that was not itself randomised make that a direction rather than a measurement. Bruns and colleagues found that telling people a default was there to influence them, or why, did not significantly reduce its effect on carbon-offset contributions, and reactance did not interact with transparency. Labelling your own nudge is cheap, and does not appear to trigger the backlash designers fear.

So the state of the field is split, and the split is what you design around. Reactance as a construct replicates: it is measurable, it has a stable two-component structure, and controlling language and removed choice reliably raise it. The design rule drawn from it — that pressure backfires at the point of conversion — does not replicate in interfaces, where the best controlled experiment shows pressure raising compliance and anger at the same time. Build against reactance because of what it does to trust, to churn and to your standing with a regulator, and measure those. Expecting the A/B test on the pushy variant to go your way is the part the evidence will not support.