Nudge… a Lot

What Bock argues

The claim is that you can move behavior at scale with small, cheap interventions — an email, a checklist, a default, a well-timed reminder — and that you should find out which ones work by running experiments rather than guessing. Bock's examples are the small stuff: a note that nudges new hires to set up faster, a default that improves savings enrollment, a checklist that makes onboarding stick. The deeper move is cultural: treat people decisions as testable. Don't roll out the wellness program to everyone and declare victory; run it on some and not others and measure the difference. Nudges are the cheap lever; experimentation is the discipline that tells you which lever actually moved anything.

The instinct is right and the cheap lever is real. The discipline is where 2015's optimism needs a cold compress.

What the research actually says (and where 2015 needs an update)

Choice architecture is real: Thaler and Sunstein (Nudge, 2008) established that the framing, ordering, and default of a choice systematically changes what people pick, without removing any option. Defaults in particular are powerful — the classic retirement-savings and organ-donation results show enrollment swinging enormously on whether the good choice is opt-in or opt-out. That much is settled, and it is genuinely cheap leverage: you are not paying people or retraining them, just designing the choice better.

Here is the update Bock's 2015 enthusiasm needs, and it is the whole reason this is a measurement chapter. Nudge effects are modest and highly heterogeneous, and the published record overstates them. When researchers looked hard at the nudge literature for publication bias, the average effect shrank substantially once you account for the studies that never got published, and the effects vary wildly by domain, population, and design — a nudge that moved the needle at one company in one quarter routinely does nothing at the next. The lesson is not "nudges don't work." It is "you cannot import a nudge's effect size from someone else's paper" — which is exactly Bock's other point (run your own experiments) made non-negotiable. The cheap lever is real; the transferable effect size is mostly a mirage. So the rigor moves from "nudge" to "measure the nudge, on your own people, honestly."

And measurement here has a cost-benefit of its own. Running a clean experiment isn't free — it takes a holdout, time, and the discipline to not peek — so the meta-question is whether the experiment is even worth running: how much would knowing the true effect change your decision, versus just acting? That is value of information, and it is the portfolio's recurring move: sometimes the honest answer is don't run the study, the decision is the same either way; sometimes it's this one is worth a clean test.

How you run it

• Design a real comparison. A/B or a proper field experiment with a holdout — not a before/after that confounds the nudge with everything else that changed.
• Report honest confidence intervals. A nudge result without a CI is a press release. Most nudge effects are small; a small effect with a wide CI is "we don't know yet," not "it works."
• Price the experiment first. Use value-of-information (EVPI/EVSI) to decide whether the test is worth its cost — would the result actually change what you do? If not, skip the study and act.
• Watch for the dark side. The same machinery that nudges toward savings can nudge toward the company's interest against the employee's; a nudge people would object to if they saw it is a dark pattern, not a kindness.

The analysis you can execute

The experiment / nudge analysis: forecasting's EVPI/EVSI to price the experiment before running it, and calculus to estimate the effect with honest confidence intervals after. Mostly composition over existing spokes. The headline outputs are two numbers most "we ran a pilot" conversations never produce: was this experiment worth running (before) and what's the effect, with its uncertainty (after).

The AI-era turn

AI makes nudges cheap to design, target, and personalize — which raises both the upside and the hazard. Personalized nudges measured against a holdout are a genuine improvement; personalized nudges assumed to work, or optimized against the company's metric without the employee's knowledge, are manipulation at scale. The Chapter 2 transparency rule applies: a nudge program people can see and opt out of is choice architecture; one they can't is a dark pattern with a friendly UI. Measure, disclose, and let people leave.

What to do Monday

1. Pick one people decision you're about to roll out broadly. Don't roll it out broadly. Run it on half.
2. Before you run it, ask the value-of-information question: would the result actually change what I do? If no, skip the experiment and decide. If yes, run a clean one.
3. Report the effect with a confidence interval, and refuse to import an effect size from anyone else's paper.
4. For any AI-personalized nudge, apply the Ch-2 test: would the recipient object if they saw the mechanism? If yes, you built a dark pattern, not a nudge.

Cross-refs: Ch 2 (transparency — the line between a nudge and a dark pattern); Ch 13 (honest null reporting — most of your experiments won't work, and that's the point); the portfolio value-of- information doctrine (price the certainty before you buy it).