Confirmshaming: It Works, By About Five Points, Until Money Is Involved

No thanks, I like paying full price. No thanks, I hate saving money. No thanks, I hate fun & games.

Those three are not invented examples. They are real decline buttons, recorded by researchers crawling eleven thousand shopping sites, and they are what confirmshaming looks like: writing the “no” option so that clicking it costs you a small amount of dignity.

For a pattern this recognisable, the evidence is remarkably thin - but it is no longer as thin as it was. Two experiments have now isolated it. Together they say something more precise, and more interesting, than either the design-blog version or the sceptical version.

It works. By about five percentage points. And in the one study that made people commit real money, it stopped reaching significance.

Where the word came from

Worth correcting a common attribution. Harry Brignull coined “dark patterns” on 28 July 2010, registering darkpatterns.org. The site now runs as deceptive.design, renamed on advice from the World Wide Web Foundation’s Tech Policy Design Lab, with the vocabulary shifted to “manipulative, deceptive and coercive patterns.”

He did not coin “confirmshaming.” His own taxonomy entry credits it to an anonymous source in 2016 - a Tumblr blog that collected examples. The term came from the internet, was adopted into a research vocabulary, and became canonical from there.

That origin is not incidental. A pattern named by a joke blog arrives with cultural confidence and no data attached.

How common is it, actually?

One measurement exists. Mathur and colleagues (2019) crawled about 11,000 shopping websites and 53,000 product pages, finding 1,818 dark-pattern instances of 15 types.

Confirmshaming: 169 instances across 164 websites - roughly 1.5% of the sites crawled. By instance count it ranked fourth of fifteen, behind low-stock messages (632), countdown timers (393) and activity notifications (313).

Two caveats. The crawler only visited product, cart and checkout pages, and only analysed text-based interfaces - so newsletter and exit-intent pop-ups, where confirmshaming is most familiar, are systematically under-sampled. The authors describe their figures as a lower bound. And they classified confirmshaming as asymmetric but not covert: unlike hidden costs or a trick question, it does not conceal what it is doing. You can see it.

That turns out to matter.

The first experiment, with the numbers it took a while to find

Luguri and Strahilevitz (2021) tested several manipulative techniques on 3,777 US participants, census-weighted. Their confirmshaming manipulation made the decline button read: “I don’t care about protecting my data or credit history.”

The main text reports only a chi-square. The acceptance rates are in a table further down:

Condition	Accepted	n
Control	14.8%	191 / 1,289
Scarcity	14.3%	91 / 635
Confirmshaming	19.6%	120 / 612
Social proof	22.1%	140 / 634
Hidden information	30.1%	183 / 607

Confirmshaming lifted acceptance by 4.8 percentage points - from 14.8% to 19.6%, a relative increase of about a third. χ²(1, N = 1,901) = 6.96, p = .008. In the cleanest cell-to-cell comparison, 20.5% against 13.2%, p = .02.

So it is real, and it is the weakest of the techniques that worked. Hidden information doubled acceptance. A trick question, elsewhere in the same study, moved it from 19.2% to 33.4%. Scarcity and countdown timers did nothing at all (p = .78).

The second experiment, which is the reason to revisit this

Zac and colleagues (2025), in Behavioural Public Policy, ran a preregistered randomised experiment with 2,252 participants, 451 of them in the confirmshaming arm. Crucially, they measured the decision twice: once when participants chose an offer, and again when they had to actually commit to paying for it.

Stage	Confirmshaming	Control	Statistic
First choice	0.5543	0.4876	χ² = 3.99, p = .05
Payment choice	0.1131	0.0790	χ² = 2.98, p = .084 - not significant

The authors put it plainly: confirmshaming “was effective at the first-choice stage” but “was weaker when participants were given a ‘second chance’ and directly confronted with a need to pay for their decision.”

That is the most useful finding in the whole area. A guilt-worded button can nudge a costless click. Give people a moment and a bill, and the effect thins out below the threshold. False hierarchy, tested in the same study, stayed strong at both stages.

Two nulls from the same paper: effectiveness was not moderated by income or education, and financial literacy did not moderate the first-choice effect.

Does it backfire? The evidence disagrees with itself

Sha, Xu and Wang (2026) ran the only study built specifically around confirmshaming: 151 participants retained of 170, with the decline button reading “No, I prefer to pay full price” against a control reading “Cancel.” The manipulation landed hard - perceived sarcasm 5.01 against 1.69, F(1,151) = 338.10, p < .001.

Two mechanisms then pulled in opposite directions:

• Financial cost perception → acceptance: β = 1.43 [0.77, 2.54]
• Experienced coercion → acceptance: β = -0.73 [-1.60, -0.21]
• Combined indirect effect on acceptance: β = 0.43, not significant.

Where the effect did not cancel was downstream. On willingness to recommend the brand, the combined indirect effect was β = -0.44, p < .05, driven by coercion at β = -0.69 [-0.94, -0.44].

But Luguri and Strahilevitz measured something similar and found nothing. Their participants’ post-exposure mood was essentially the same in the confirmshaming condition (M = 3.82, SD 2.05) as in the control (M = 3.72, SD 1.91), and the manipulation had no effect on willingness to take part in follow-up research, F(4, 2112) = 0.96, p = .43.

So: one study finds a reputational cost, a much larger one finds no mood cost. Both are single measurements of a slippery construct. The backfire claim is plausible and unsettled, and anyone stating it as established is going beyond the evidence.

People see it coming, and think it does not work on them

Bongard-Blanchy and colleagues (2021) showed nine interface mockups to 406 UK participants, representative by age, gender and ethnicity, and asked whether they spotted a manipulation and whether it would influence them.

71% detected the confirmshaming interface - the second most-detected of the nine. And they rated themselves as less likely to be influenced by it than by any other pattern shown, at -0.85 on a scale running from -2 to +2.

This is self-report, so it measures perceived resistance rather than actual resistance, and the authors note their findings concern one specific implementation. But set it beside the experiments and there is a real tension: the pattern people are most confident they can see through still moved acceptance by nearly five points.

What the general psychology predicts - and it is not what you would expect

Confirmshaming presumably works, if it works, through guilt or through the discomfort of being pushed. Both mechanisms have been meta-analysed, and neither points where the design literature assumes.

Guilt appeals. Turner and Rains (2021) pooled 26 studies. More intense guilt appeals did produce more guilt (r = .29) - and also more anger (r = .24). But the amount of guilt produced was not related, linearly or curvilinearly, to the appeal’s effect on attitudes or behavioural intentions. Peng and colleagues (2023) pooled 26 studies, 127 effect sizes and 7,512 participants, and found a pooled effect of g = 0.19, 95% CI [0.10, 0.28] - small.

Reactance. Li and Shi (2026) pooled 33 studies and 146 effect sizes. Freedom-threatening language reliably produced anger (r = .21) and counterarguing (r = .17) - and those in turn were reliably associated with less persuasion (r = -.23 and r = -.18).

Put together, the mechanism literature predicts that a shaming decline button should be roughly a wash: it induces guilt, which does not convert, and anger, which actively works against you. That the two isolating experiments nonetheless found small positive effects is a point in their favour, not a reason to inflate them - and it means the general literature cannot be borrowed to prop this up. If anything it argues the other way.

A paper we are not counting

Löschner and Pannasch (2023) tested preselection, confirmshaming and disguised ads together and separately, and is frequently listed as confirmshaming evidence.

We are not treating it as such. It appeared as a late-breaking poster, its full text sits behind a paywall we could not get through, and its abstract singles out disguised ads as the manipulation that raised error rates and dwell time. No standalone confirmshaming effect is reported in anything we could read, and the systematic review below codes it as contributing exactly one experimental unit. Citing it would be citing a title.

What the review says, read carefully

Schaffner, Heysen and Chetty (CHI 2026) reviewed 148 experimental units across 27 peer-reviewed papers, 2019 to 2025, and record confirmshaming as appearing in 28 experimental instances.

That number needs reading carefully. They are experimental units within studies, not 28 separate studies - Luguri and Strahilevitz alone contribute 17 units to the corpus. The review reports no confirmshaming-specific significance rate.

Two of its wider findings apply here. 15 of 101 experiments testing a pattern against a control found no significant effect, with large variation in effect sizes indicating strong context dependence. And the authors state that some degree of publication bias should be assumed across a literature where 85% of such experiments came out significant.

Why the gap between fame and evidence

Confirmshaming is unusually visible. It is written in words, it is often funny, and it is easy to screenshot. That makes it the pattern most likely to be collected, shared, complained about and taught - and none of those activities generate data.

Obstruction, by contrast, is boring. It is an extra screen, a button one layer down. Nobody screenshots a slightly longer cancellation flow. And in Luguri and Strahilevitz’s own data it was the stronger effect.

There is a general lesson in that asymmetry. Visibility and effectiveness are not the same property, and the patterns that get named are selected for the first. The things doing the most work in an interface tend to be structural and unremarkable: what is on the first screen, what is already ticked, how many steps the path you did not take requires.

So what should you do about it?

Less than the outrage suggests, and the numbers say why.

If a decline button insults you, you are being moved - by about five points, on a costless click. That is a real effect and a small one. And the best evidence available says it thins out at exactly the moment that matters: when the decision involves handing over money, the same manipulation stopped reaching significance.

You are also, statistically, seeing it. Seven people in ten spot it. What they get wrong is the next part - believing that spotting it means it is not working, when the experiments say a small effect survives being noticed.

The practical move is the one the whole evidence base points at. When an interface makes you feel something on the way to a decision, that feeling is usually the least effective thing being done to you. Look instead at what has been moved, hidden, pre-selected, or placed one screen further away - and, if there is a price attached, take the second look that made the effect disappear in the laboratory.

Confirmshaming is one of a family of interface patterns, most of which work by structure rather than sentiment - see the persuasive design guide for the wider set, the roach motel for the obstruction pattern that outperforms it, and should you accept cookies for the largest field measurements available.