attck-research

Evidence review: Choice Architecture for Ethical Persuasion

Article: Choice Architecture for Ethical Persuasion
Reviewed: November 2026 · Method version 2 · How these reviews work

How this review works

Each factual claim in the article was checked against the source it cites, then challenged by reviewers whose job was to find the strongest evidence against it: replication attempts, meta-analyses, later papers that cite the original work, and research from other fields that study the same mechanism. A separate reviewer weighed both sides, and a script re-checked every quoted source. Where a claim was overstated or misattributed, the article was corrected. The research below supports the article as it now reads.

Background research

Introduction

Economic models assumed a chooser who weighed every option on its merits, so a plan’s place on a list should not matter, and a pre-ticked box should matter only as much as the effort it takes to untick it.

Standard economics does allow small default effects through opt-out costs; the behavioral finding is that defaults matter far more than those costs explain (Madrian and Shea 2001).

In 2003 Thaler and Sunstein published ‘Libertarian Paternalism’; their 2008 book Nudge gave the job a name: choice architecture.

Title and year of confirmed from Crossref; the naming of choice architecture is confirmed in.

Twelve decision researchers led by Eric Johnson (2012) stated there is no neutral architecture; every presentation has a default, even when the default is that nothing changes.

The ‘no neutral architecture’ line was not verified verbatim (full text paywalled); the abstract supports the premise.

In 2018 Thaler gave the bad direction its own word: sludge, design that makes wise decisions harder to reach.

Friction is not always harmful: transparent friction can help some people (J Behav Decis Making 2026). Thaler’s term refers to friction that works against the chooser.

Every Screen Already Leans

In a 2015 column Thaler proposed three tests: the nudge should be transparent and never misleading, opting out should be as easy as possible, and there should be good reason to believe it improves the welfare of the person nudged.

The three principles restate guidelines from Nudge (2008). A 2025 field experiment found a transparent default worked better than a hidden one, which supports the transparency test.

A default that enrolls people in what most of them would pick after an hour of comparison saves them the hour, as long as the people it does not fit can see it and change it easily.

Defaults are sticky even for people they do not suit (Madrian and Shea 2001), and when needs vary widely, asking people to choose can beat any default (Carroll et al. 2009).

Finding 1

Johnson et al. (2012) map the toolkit: tools that structure the task (number of alternatives, decision aids, defaults, time windows, staging of configuration) and tools that describe options (grouping, attribute parsimony, scale of numbers).

Two-way split confirmed in the abstract; the sub-tool list was not verified verbatim.

In the paper’s example, drawn from work by Richard Larrick and Jack Soll, miles per gallon is non-linear in fuel used, so people undervalue upgrading inefficient cars; gallons per 100 miles corrects the misreading.

The MPG illusion originates with Larrick and Soll (2008); Camilleri and Larrick (2014) extend it to metric and scale design.

Finding 2

At the end of 2021 Mertens and colleagues published a PNAS meta-analysis of more than 200 studies and over two million participants, finding a small-to-medium overall effect (Cohen’s d = 0.45).

Mertens et al.’s own reply gives d = 0.31 under moderate and 0.08 under severe publication bias; a 2025 second-order meta-analysis finds near zero after adjustment.

In Mertens et al., interventions that changed the structure of a decision (defaults, arrangement of options) outperformed those that only added information or reminded people of their intentions; they also reported moderate publication bias.

The structure-versus-information gap is a pre-correction result; after correction structure evidence is undecided rather than positive (Maier et al. 2022).

Seven months later Maier and colleagues found, after correcting for publication bias, no evidence for an overall effect, evidence against information and assistance nudges, undecided evidence for structural ones, and wide heterogeneity meaning some nudges may still work.

Mertens et al. replied that a moderate-bias correction still leaves d = 0.31; the debate is over how severe the bias is.

Save More Tomorrow

In the first implementation (four annual raises), 78% of employees offered the plan joined, 80% of enrollees stayed through the fourth raise, and average saving rose from 3.5% to 13.6% over 40 months.

First implementation was a single firm and not randomized. Mertens et al. file commitment devices under decision assistance, the category Maier et al. found evidence against on average; that concerns the category, not these reported numbers.

The Federal Trade Commission’s 2022 staff report on dark patterns quotes a description of the cookie dialog that puts accepting front and center while refusing sits behind several screens.

Experimental support: removing Reject from the first consent layer raised consent by 22 to 23 points (Nouwens et al. 2020).

Plan selection and add-ons

He accepts required choices when a careless mistake would be costly. Adding a recommendation makes the nudge much stronger, and he faults this one because the cover was poor value for most travelers and the pitch was not transparent.

Thaler’s objection rests on value and transparency, not on combining required choice with a recommendation as such.

Free trials and cancellation

The column describes a £1 newspaper trial that rolled into £26 a month, needed 15 days’ notice to cancel and could only be cancelled by phoning London in British business hours; Thaler judged it failed all three tests.

The newspaper was The Times of London.

The FTC’s standard: cancelling should be at least as easy as signing up, through the same medium, without retention offers that cause unreasonable delay.

The report presents this as ROSCA’s requirement under the FTC’s 2021 policy statement; the later click-to-cancel rule’s legal status changed (not verified here).

What changed in the article

These edits were made on November 12, 2026 after the review.

Introduction: supported with caveats

Claim reviewed: Economic models assumed a chooser who weighed every option on its merits, so a plan’s place on a list or a pre-ticked box should not matter.

Now reads: Economic models assumed a chooser who weighed every option on its merits, so a plan’s place on a list should not matter, and a pre-ticked box should matter only as much as the effort it takes to untick it.

Why: This is a fair one-line summary of the rational-choice view, which treats the way options are presented as irrelevant. The caveat is that even standard models let a pre-ticked box matter a little, because unticking it takes effort. The sentence is uncited, but it is a textbook point.

Every Screen Already Leans: supported with caveats

Claim reviewed: A default that enrolls people in what most of them would pick after an hour of comparison saves them the hour.

Now reads: A default that enrolls people in what most of them would pick after an hour of comparison saves them the hour, as long as the people it does not fit can see it and change it easily.

Why: As a conditional statement it holds: a default that matches what people would pick after comparing saves them the effort. The catch is that a default also holds onto the people it does not fit, and retirement-plan research shows many keep defaults that few would have chosen on their own. Which default is best depends on how much people’s needs differ. Now cites Carroll, Choi, Laibson, Madrian & Metrick, Optimal Defaults and Active Decisions, QJE 2009.

Finding 1: supported with caveats

Claim reviewed: Miles per gallon is non-linear in fuel used, so people undervalue upgrading inefficient cars; gallons per 100 miles corrects the misreading, and costs over different periods are compared more accurately on a shared scale.

Now reads: In the paper’s example, drawn from work by Richard Larrick and Jack Soll, miles per gallon is non-linear in fuel used, so people undervalue upgrading inefficient cars; gallons per 100 miles corrects the misreading.

Why: The miles-per-gallon example is a real and well-replicated finding, and Johnson et al. do use it. It originally comes from Larrick and Soll’s 2008 Science paper, which the article does not credit. A later experiment confirms that the unit and scale of a number shift choices. Now cites Larrick & Soll, The MPG Illusion, Science 2008.

Finding 2 / Do’s and Don’ts: not supported

Claim reviewed: What evidence survives the bias correction sits with the structure of the decision: what is preselected, how many options, their order, the unit of the price; structure is where evidence is strongest and banners and reminders where it is weakest.

Now reads: Choice architecture is not a dial you can turn on any screen for a guaranteed lift. After the correction, the evidence for structural changes is undecided rather than negative, and defaults have the most support. Spend effort on what is preselected, how many options there are and their order, then test on your own screens. The unit of the price is a different kind of tool. The meta-analysis files it with changes to how information is described, the group where the corrected evidence points against an average effect, even though single field experiments on daily versus monthly amounts have found large differences. Treat it as a hypothesis to test, not a sure lift.

Why: The reanalysis the sentence cites does not say structural evidence ‘survives’ the bias correction; it says the evidence is undecided, and a 2025 synthesis of 14 nudge meta-analyses finds the average effect drops to about zero after the same kind of correction. The sentence also lists the unit of the price as a structural tool, but the meta-analysis files unit and scale changes under decision information, the group where the corrected evidence points against an average effect. The relative ranking is right: in the original meta-analysis, in newer field meta-analyses on food choice, and in the nudge-unit comparison, structural changes such as defaults outperform information and reminders. Now cites Hershfield, Shu & Benartzi, Temporal Reframing and Participation in a Savings Program: A Field Experiment, Marketing Science 2020.

Save More Tomorrow: not supported

Claim reviewed: Retention says as much about the ethics as uptake: a program that talked people into something against their interests would lose them as each raise was redirected; most stayed.

Now reads: Retention alone does not settle the ethics. Staying in was the default, and Thaler and Benartzi say plainly that one reason the plan works so well is that inertia is so powerful: once people enroll, few opt out. The stronger case is the design. People chose to join, knew in advance what each raise would do and could leave at any time, and those who left kept the increases already made rather than undoing them. The lesson transfers: schedule the cost of a good choice for the moment it hurts least, and say exactly what will happen.

Why: Staying in the plan was the default, and Thaler and Benartzi say so themselves: one reason the plan works so well is that inertia is so powerful, and once people enroll few opt out. High retention is therefore predicted by the design whether or not the plan helped people, and evidence that people keep default pension settings, subscriptions and gym contracts they no longer value points the same way. The plan may well have served participants, but that case rests on their choosing to join, being able to leave at any time, and keeping their increases even when they left, not on retention.

Claim reviewed: The FTC’s 2022 staff report describes the cookie dialog that puts accepting front and center while refusing sits behind several screens, and recommends avoiding unexpected data-collecting defaults, presenting choices when the decision is made, and not making one option more prominent than another.

Now reads: The Federal Trade Commission’s 2022 staff report on dark patterns quotes a description of the cookie dialog that puts accepting front and center while refusing sits behind several screens.

Why: The FTC report contains the cookie description and all three recommendations. The description of the cookie dialog comes from a workshop panelist the report quotes, not from the FTC’s own findings. HCI experiments confirm that burying Reject raises consent sharply. Now cites Nouwens et al., Dark Patterns after the GDPR, CHI 2020.

Plan selection and add-ons: supported with caveats

Claim reviewed: Thaler’s column describes an airline requiring a yes/no on trip insurance with ‘yes’ marked recommended, $20.13 of cover on a $300 fare; he accepts required choices when a careless mistake would be costly, and pairing one with a recommendation is what he questions.

Now reads: He accepts required choices when a careless mistake would be costly. Adding a recommendation makes the nudge much stronger, and he faults this one because the cover was poor value for most travelers and the pitch was not transparent.

Why: The airline example and numbers are accurate, and Thaler does accept required choices to prevent careless mistakes. But he does not say the pairing with a recommendation is the problem; he calls it a strong nudge and faults this one because the insurance was poor value and the pitch was not transparent. The article’s lesson still follows, but it should be credited to his reasons.