Nudge: meaning and application

A pension fund wants members to check their choices, and an employer wants staff to complete a training course. Sometimes it helps more to arrange the choice differently than to add explanations or offer a reward. What exactly is a nudge, how well does it work and where are the ethical limits?

A nudge is a change in the choice architecture that alters behaviour in a predictable way without forbidding any options or significantly changing financial incentives.

Richard Thaler and Cass Sunstein popularised the concept in 2008. Defaults, reminders, social norm information, simplification and framing are all nudges. Published studies find a small to medium average effect, but much less remains after correcting for publication bias and at scale.

Richard Thaler and Cass Sunstein popularised the term. In their book Nudge (Thaler & Sunstein, 2008, p. 6), a nudge is any aspect of the choice architecture that alters behaviour in a predictable way, without forbidding any options or significantly changing economic incentives (Barton & Grüne-Yanoff, 2015, who quote the definition verbatim).

Choice architecture is how a choice is presented: which option is preset, the order of options and how much effort an action takes. It always exists, even if nobody designed it deliberately. The book built on the authors' earlier article on libertarian paternalism (Thaler & Sunstein, 2003).

A nudge forbids no options and does not significantly change financial incentives. A ban, fine or substantial reward is therefore not a nudge, though 'significantly' leaves room for interpretation.

A nudge is broader than one tool. A reminder is one type; a default, a shorter sign-up procedure or a different order of options are nudges too. Information can be part of a nudge when presented more simply, visibly or at a better moment.

The concept says nothing about direction: the same techniques can ease choices people want to make, but can also steer choices in a provider's interest.

Mertens and colleagues distinguish three categories, each aimed at a different barrier (Mertens et al., 2022):

• Decision information (limited access to information): making information understandable or visible, or offering a social reference point, such as what similar others do.
• Decision structure (limited capacity to evaluate options): defaults or active choice, the effort an option requires, and the composition and order of options, including simplification.
• Decision assistance (limited attention and self-control): reminders and support for making a commitment.

Framing fits best under decision information: describing a choice as a gain or a loss can shift preferences predictably (Tversky & Kahneman, 1981).

A default is what applies if someone does nothing. Johnson and Goldstein compared organ donation in countries where people must opt in with countries where everyone is a donor unless they opt out. Consent was in the high nineties in opt-out countries and far lower in opt-in countries (Johnson & Goldstein, 2003, as summarised by Jachimowicz et al., 2019). A large replication reproduced the effect in their scenario experiment (Chandrashekar et al., 2023). At a US employer, automatic pension enrolment raised participation, and many kept the preset contribution (Madrian & Shea, 2001).

A meta-analysis of 58 datasets (73,675 participants) found an average d = 0.68 (95% confidence interval 0.53–0.83), stronger for consumer and weaker for environmental decisions. Defaults worked mainly because they feel like a recommendation or like something already owned (Jachimowicz et al., 2019). A default changes a choice, but not automatically the outcome: for organ donation, opt-out is one factor among many (Steffel et al., 2019).

Mertens and colleagues' meta-analysis combines over 200 publications with over two million participants. The corrected version (2022) arrives at an average d = 0.43 (95% confidence interval 0.38–0.48): small to medium. Cohen's d expresses a group difference in standard deviations. Decision structure interventions scored highest (d = 0.54), decision information (0.34) and decision assistance (0.28) lower. Food choices showed the largest effect (0.65), financial decisions the smallest (0.24) (Mertens et al., 2022).

Such an average predicts little about one nudge in your context. Effects vary widely; many studies probably have a true effect near zero (Szaszi et al., 2022).

Studies with large effects are published more often. Mertens and colleagues found signs of this themselves; assuming moderate bias, their estimate fell to d = 0.31 (Mertens et al., 2022). Maier and colleagues corrected with a robust Bayesian meta-analysis, leaving an average of 0.04 (0.00–0.14). They found evidence against an effect for information and assistance nudges and for finance; for structure nudges such as defaults the evidence was undecided (Maier et al., 2022). So a general average effect is unproven, not every nudge ineffective.

DellaVigna and Linos analysed 126 experiments with over 23 million people by two US nudge units (government teams). The average effect was 1.4 percentage points (8.0 per cent above control) versus 8.7 percentage points (33.4 per cent) in academic journals. Selective publication explained about 70 per cent of that gap (DellaVigna & Linos, 2022).

Thaler and Sunstein call this libertarian paternalism: paternalistic because behaviour is steered towards people's own welfare, libertarian because freedom of choice remains (Barton & Grüne-Yanoff, 2015). This raises questions:

• Who decides what is good? Situations and preferences differ widely.
• Transparency. According to Barton and Grüne-Yanoff, nudges are only truly avoidable if both the type of intervention and the specific intervention are visible; covert nudges are harder to scrutinise.
• Manipulation. The more an intervention relies on inattention rather than understandable information, the closer it comes to manipulation.

Editorial advice: could you explain the nudge openly to the people it influences, and would they endorse its direction?

Choice architecture can also create barriers, which Thaler called sludge (Thaler, 2018). Sunstein describes sludge as frictions that separate people from what they want to get, such as long forms or a cancellation far harder than signing up (Newall, 2023).

Dark patterns go further: interface designs that coerce, steer or deceive users into unintended, potentially harmful decisions. Across about 11,000 shopping websites, Mathur and colleagues found 1,818 instances (Mathur et al., 2019). The line with a nudge lies at deception, coercion and undermining free choice. Read more in Neuromarketing and ethics.

These steps are editorial advice, not a validated protocol.

1. Make the behaviour concrete: who does what, when, and what outcome counts?
2. Map the choice architecture, including sludge.
3. Diagnose: is information missing, is the choice too complex, is attention lacking or are people unwilling? Interviews and association research show what the choice means to people.
4. Choose a fitting nudge and check it ethically: transparent, in the target group's interest, easy to avoid.
5. Test with a control group and pre-specify the outcome measure and analysis.
6. Measure behaviour and outcome, including by subgroup.
7. Expect a smaller effect when scaling up.

This example is fictional and contains no research results.

Situation. A pension fund wants members to log in and check choices such as contact details and a partner's pension allocation. After the annual letter, few log in.

Decision question. Which change makes logging in and checking likelier, without pressure or deception?

Available information. Log data show where members drop off; interviews show how they view pensions.

Suitable approach. First remove sludge: fewer login steps, a direct link. Then randomly assign three versions: the current letter, a simplified letter with one concrete action and the time needed, and that letter plus a reminder after two weeks. Pre-specified outcome: completed checks within thirty days, not just logins.

Possible interpretation. A difference applies to these members at this moment; it does not show better pension choices.

Next step. Repeat with another group. Without a difference, return to the diagnosis: do members not want to, or can they not?

• Equating nudge with one tool, such as a reminder.
• Using an average effect as a prediction for your target group.
• Measuring choices, forgetting outcomes: more sign-ups are not yet the desired result.
• Nudging without diagnosis: people who deliberately reject a behaviour will not change because of a nudge.
• Using covert or deceptive nudges, which undermine trust.
• No control group, so you cannot tell whether the nudge made the difference.

A nudge deliberately changes the choice architecture without forbidding options or significantly changing incentives. Defaults have the strongest evidence, but average effects of nudges are probably overestimated. Treat every nudge as a hypothesis: diagnose, test with a control group and only choose interventions you can justify openly.