Jakob Nielsen：基于行为经济学的50条UX设计准则与关键数据

Summary: People don’t decide the way classical economics assumes. They take mental shortcuts and follow seemingly irrational rules that were tuned for survival in the Paleolithic, not for pricing pages. This article distills the key behavioral economics findings into 50 design guidelines, plus two numbers worth memorizing: losses hurt about twice as much as gains please, and a nudge moves behavior by about 1.4 percentage points.

Every interface is a choice architecture. The moment you pick a default, order a list, or label a button, you steer decisions. You can’t opt out of this role. You can only perform it well or badly, honestly or not.

Definition: Behavioral economics is the experimental study of how people actually judge and choose, with special attention to the ways real behavior departs from what a perfectly rational decision-maker would do.

Classical economics assumed away that gap by modeling a creature that Richard Thaler calls “the Econ”: an agent with unlimited attention, flawless arithmetic, stable preferences, and iron self-control. Econs read license agreements. Econs compare 47 insurance plans and pick the optimum. You have never met one.

Behavioral economics studies the other species, Humans. The gap between the two species, between the user the spreadsheet assumes and the user who actually shows up, is the territory where behavioral economics and UX have always worked, mostly without talking to each other. Let me connect the two disciplines.

Why are Humans built this way? Evolution tells us that fast heuristics kept our ancestors alive: the forager who deliberated carefully about the rustling grass became lunch, while the one who jumped at shadows lived to reproduce and pass down the genes for jumping at shadows. Those same shortcuts now navigate infinite scroll, 40-option pricing pages, and subscription checkouts, environments they were never tuned for. Biologists call this an evolutionary mismatch. We live it every day. Gerd Gigerenzer’s school draws the optimistic corollary, ecological rationality: a heuristic is neither smart nor stupid in itself, only matched or mismatched to its environment. You can’t rewire the heuristic. But you design the environment.

This article covers the behavioral economics that UX professionals should actually know: 10 core findings grouped into 4 clusters, each with a definition, an evidence rating, and 5 design guidelines (50 in all, numbered straight through for easy reference); a sampler table of 12 more effects; what nudging really delivers at scale; and the line that separates making money (good) from cheating users (bad).

Three Nobel Prizes and One Caveat

For most of the 20th century, economics modeled people as rational optimizers, largely because the math demanded it. The dissenters kept winning Nobel Prizes. Herbert Simon took the 1978 prize for bounded rationality: the observation that real decision-makers lack the time, information, and computational capacity that optimization requires. Daniel Kahneman, a psychologist by training, won in 2002 for importing the experimental psychology of judgment into economics. Richard Thaler completed the arc in 2017 by showing that these so-called anomalies have systematic, predictable structure that firms and governments can act on. (Amos Tversky, Kahneman’s collaborator on the foundational work, died in 1996, and Nobels are never awarded posthumously. Everyone in the field knows he would have shared the 2002 prize.) The prize count matters for a practical reason: if stakeholders dismiss behavioral research as soft science, you can note that the Royal Swedish Academy of Sciences disagreed 3 times.

One of these laureates matters double. Simon is the only person to win both the Nobel in economics (1978) and the Turing Award (1975), computing’s highest honor. Behavioral economics and human–computer interaction share a grandfather, so it’s no accident that his ideas translate so directly into design. His concept of satisficing, accepting the first good-enough option instead of hunting for the best one, remains the single most useful model of user behavior on the web.

Somewhere in this orchard hangs a more perfect peach, but the woman is happy with the good one in her hand rather than spending the day hunting for the best one. Satisficing in action: she’s not the produce buyer for Mitsukoshi’s food hall in the Ginza; she just wants to enjoy a peach. (Muse Image)

Two papers anchor the field’s chronology. In 1974, Tversky and Kahneman published “Judgment Under Uncertainty” in Science, naming three heuristics that produce systematic biases: availability, representativeness, and anchoring. In 1979, they published prospect theory in Econometrica, now the most-cited paper in economics. Those two papers divide the field’s territory between them: the first covers how people judge what’s true, and the second covers how people choose what to do, which is why I’ve organized the clusters below the same way. Kahneman’s 2011 bestseller Thinking, Fast and Slow later packaged the whole program as System 1 (fast, automatic) and System 2 (slow, deliberate), which is the version most of your colleagues know, despite the usability horror of the names, which Kahneman borrowed from Keith Stanovich and Richard West. Quick, is System 1 fast or slow? Why not call it the Fast System instead of a number?

Now the caveat. The replication wave of the 2010s put behavioral science on trial, and some celebrated findings shrank or collapsed outright when independent teams re-ran them. So this article focuses on the results that held up, or at least haven’t been refuted. Each concept section below ends with an evidence rating (Strong, Moderate, or Contested) plus one clause explaining why. When the evidence is mixed, I say so with a number rather than a shrug. A design discipline built on folklore deserves the setbacks it gets, and UX has enough folklore already.

The Foundations: Why Users Can’t Be Econs

Three findings explain why the rest of this article exists. They aren’t biases to correct; they’re the operating conditions of the human mind. Design for them, or watch users satisfice their way past your product to a competitor’s.

Bounded Rationality

The herd is fording now. The hunter who grabbed a good-enough spear is throwing; nobody who kept comparing will eat tonight.

Bounded rationality says that people reason toward their goals, but only within strict limits of time, information, attention, and computational capacity, so they rely on shortcuts and settle for good-enough answers. The name is Simon’s deliberate contrast with the unbounded rationality of classical economic models: the rationality is real, but it operates inside bounds, like a chess player who plays sensibly without calculating every line. Within those bounds, people don’t optimize. They satisfice, Simon’s portmanteau of satisfy and suffice: take the first option that clears the bar and move on with their lives.

This is the foundation beneath almost every UX problem, from abandoned shopping carts to unread privacy policies. In 1990, Simon compressed the idea into a metaphor worth memorizing: rational behavior is shaped by a pair of scissors whose two blades are the structure of the task environment and the computational capabilities of the actor. You will never sharpen the second blade. The first blade is your product, and the design consequence follows directly: reduce the information, memory, calculation, and inference that a sound decision requires.

Our poor human brains can only hold so much information. We have to rely on shortcuts to get anything done.

Evidence: Strong. Seven decades of decision research confirm the framework, and satisficing appears wherever anyone has looked for it.

Design guidelines for bounded rationality:

1.          Summarize before detailing. Give the conclusion first and let users drill down if they care; that order matches how bounded minds allocate effort.

2.          Offer defensible recommendations. “Most teams choose Standard” spares users a comparison they’re unequipped to make, provided the recommendation genuinely serves them.

3.          Stage the detail. Progressive disclosure exists because working memory doesn’t scale; reveal complexity only when the user asks for it.

4.          Provide safe paths. A user who can’t evaluate every branch needs routes that can’t end in disaster, plus an undo for the branches that surprised him or her.

5.          Make automation reviewable. Decisions the system takes on the user’s behalf should be inspectable and reversible, not merely trusted.

Notice what these have in common: each one transfers work from the user’s bounded budget to the computer’s unbounded one. Your system can afford to compute, compare, and remember; your user can’t, or more precisely won’t, because his or her cognitive budget has 30 other claims on it before lunch. Every simplification is a subsidy.

AI is the newest bounded-rationality prosthetic: summarization, recommendation, and retrieval all economize on the scarcest resource, which is human processing. But an AI summary the user can’t verify just relocates the bound, so keep the source one click away.

Limited Attention and Rational Inattention

Fresh tracks, and a valley full of noise. The tracker’s survival skill is not what he sees but everything he manages not to.

Limited attention says that users can process only a small fraction of what any screen presents, so they spend attention like the scarce currency it is, on whatever looks immediately valuable, and skip the rest. Economist Christopher Sims’s rational inattention adds the twist that its paradoxical-sounding name announces: ignoring isn’t a failure of attention but a rational use of it, because processing information costs effort that most information doesn’t repay. My old finding that “users don’t read” isn’t an insult to users. It’s economics translated to the web.

Banner blindness is rational inattention in its purest form: users learned that anything shaped like an ad almost never pays off, so they stopped seeing ads, along with everything that resembles one. The lesson generalizes, because attention follows expected value. And every element that wastes it teaches users to ignore that class of element forever. Notifications make the economics explicit: a push message costs the sender nothing and the recipient a context switch, and when the party who pays isn’t the party who decides, economists call it an externality. The notification tray is an overgrazed commons. Cognitive load, the working-memory side of the same coin, gets a full treatment in my cognitive load article.

Evidence: Strong. Attention limits are among the best-documented facts in psychology, and the economic formalization has held up since 2003.

Design guidelines for limited attention:

6.          Place information at the point of need. Fees, risks, deadlines, status, and next actions belong where the decision happens, not where the org chart wanted them.

7.          Reduce competing signals. Emphasis is a zero-sum budget: every additional highlight taxes the one that matters.

8.          Design for interruption. Preserve state and progress across breaks, because attention will be interrupted; that’s a planning assumption, not a failure mode.

9.          Spend reminders like money. Each notification draws down a finite trust account, so reserve them for events the user would pay to know about.

10.      Believe the analytics over the floor plan. When users miss a control that’s “right there,” remember that right there is a property of your attention after 400 hours with the screen, not theirs after 40 seconds.

Choice Overload Meets Choice Architecture

The scout has crossed 30 valleys so the elders can choose among three. Curation came before choice, by a few hundred thousand years.

Choice overload says that adding options can hurt: past some point, more alternatives produce less buying, lower satisfaction, and more deferral, because the effort of comparing grows faster than the set: 6 options make 15 pairwise comparisons, and 24 make 276. Its design-side twin, choice architecture, is Thaler and Cass Sunstein’s term for the arrangement of options (order, grouping, labels, comparisons, defaults, timing, reversibility) that shapes which option gets chosen. Both names mean what they say: too many choices overload, and every arrangement is architecture.

The famous evidence came from Sheena Iyengar and Mark Lepper’s jam-tasting booth: with 24 flavors, only 3% of tasters bought, versus 30% with 6 flavors. The jam study launched a thousand keynotes. Then Benjamin Scheibehenne and co-authors’ 2010 meta-analysis of 50 experiments found a mean overload effect of essentially zero. Too bad; it was a great story. Alexander Chernev and co-authors identified in 2015 when overload actually bites: it depends on 4 moderators, namely how complex the options are, how effortful the decision task is, how unsure users are of their own preferences, and how much time pressure they face. Assortment size on its own predicts little, which is why Amazon thrives with hundreds of millions of products: search, filters, reviews, and rankings do the winnowing.

Most designers are accidental architects: they nudge with every default and every sort order, just without noticing. The guidelines in this article are how you stop doing it by accident; nobody grants you a license before handing you this power, because shipping an interface is the license. AI products supply this decade’s tasting booth, with model pickers full of cryptic names and prompt-suggestion menus longer than a diner menu. And AI-curated shortlists are the corresponding remedy, if users can see why the shortlist was chosen.

Evidence: Contested as originally stated. Overload is real under identifiable moderators, but the blanket claim that fewer options always win died in meta-analysis.

Design guidelines for choice overload:

11.      Curate a shortlist, keep the long list. Offer 3–5 recommended options up front while preserving access to the full set for users who want it.

12.      Support comparison as well as choice. Normalize attributes into comparable terms, because overload lives in the comparing, not the counting.

13.      Provide honest filters and rankings. Let users shrink the set by their own criteria before you shrink it by yours.

14.      Reveal advanced options progressively. Experts can dig deeper, while novices never face the 40-field wall.

15.      Design the architecture on purpose. Choose order, grouping, defaults, and reversibility around the user’s objective, because you’re choosing them either way.

Judging: The 1974 Heuristics

The 1974 Science paper named three heuristics for judging under uncertainty: availability, representativeness, and anchoring. Two get full sections here, because they intrude on design decisions weekly. Representativeness appears in the table below under its most design-relevant symptom, base-rate neglect.

Anchoring Effect and Insufficient Adjustment

The elder’s staff finds the depth once, and the whole band crosses on that first reading, adjusting step by step as the river tells them more.

Anchoring says that the first number in view drags every subsequent judgment toward it, even when that number is irrelevant, because people estimate by adjusting from a starting point and stop adjusting too soon. The name is the metaphor: judgment drops anchor at the initial value, and the chain is short, which is why the full label is anchoring and insufficient adjustment.

In the classic demonstration, Tversky and Kahneman spun a rigged wheel of fortune that landed on either 10 or 65, then asked subjects what percentage of UN members are African nations; median estimates were 25% vs. 45%. Expertise is no vaccine: experienced German judges in Birte Englich and co-authors’ 2006 experiment gave a shoplifter about 8 months when their own dice roll had set the prosecutor’s demand at 9 months, and about 5 months when it said 3. In interfaces, initial numbers steer prices, estimates, donations, tips, forecasts, and configuration values. The crossed-out “was $499” beside today’s price is an anchor with a business model, and so are the middle tier on your pricing page and the preset amounts on every donation form. (I go deeper in a recent piece about anchoring.)

Show the $999 premium plan first, and people will feel that the $499 standard plan is a bargain.

The 2026 twist: the AI first draft is an anchor. Start from the model’s text, and its structure, claims, and numbers pull everything that follows; he or she who edits an AI draft rarely strays far from it. The same applies to AI-suggested form values, estimates, and code.

Evidence: Strong. Anchoring produced some of the largest and most reliable effects in the Many Labs replication project (36 samples, 6,344 participants).

Design guidelines for anchoring:

16.      Anchor with defensible benchmarks. Suggested values should be ones you could justify to the user’s face, such as the median choice of similar customers.

17.      Disclose the anchor’s origin. “Based on your last 3 months” turns a manipulation risk into decision support.

18.      Show ranges, not lone numbers. A credible low–typical–high span weakens the grip of any single arbitrary figure.

19.      Keep presets clearable. Every prefilled value should be one keystroke from gone.

20.      Break the AI draft’s anchor. Sketch your own outline before reading the model’s version, or request 2–3 deliberately divergent drafts so that no single one owns the frame.

Availability Heuristic and Vividness

The cairn marks a loss old enough that no one living saw it. The band still walks wide, and a mother makes sure the next generation will too.

The availability heuristic says that people judge how frequent or probable something is by how easily examples come to mind. The name refers to that ease of retrieval: whatever memory makes available gets counted as common, so vivid, recent, and dramatic events crowd out the quiet majority. (A heuristic, in Tversky and Kahneman’s vocabulary, is a mental shortcut that’s usually right and, when wrong, wrong in a predictable direction. Their use of the word has nothing to do with my 10 usability heuristics, which are also usually right but err occasionally, but not in a systematically wrong direction.)

So one dramatic review outweighs a hundred quietly satisfied customers, and one plane crash outweighs decades of aviation safety statistics. In UX, availability matters most in reviews, testimonials, risk communication, and anything involving rare events; it matters less in routine operation, where users judge from their own repeated experience. AI gives the heuristic fresh teeth, and they bite in both directions: one vivid hallucination anecdote convinces an organization that AI output can’t be trusted, while the accuracy base rates go unread, and one dazzling demo convinces the same organization that AI output doesn’t need checking. Both errors come from the same shortcut. And both grow with retelling. Timur Kuran and Cass Sunstein named the loop an availability cascade: each repetition of a story makes it more available, which makes it more repeated, until nobody remembers that it happened once.

Vivid events lodge in the mind and are therefore easily recalled, causing people to overestimate their likelihood.

Evidence: Strong. A half-century of studies across risk perception, medicine, and consumer judgment, with no serious replication trouble.

Design guidelines for availability:

21.      Pair every vivid particular with its denominator. Put “4.6 stars across 12,384 reviews” next to the three glowing quotes.

22.      Publish base rates for rare events. “0.8% of shipments arrive late” beats an eloquent apology.

23.      Show time windows. Last week’s outage will dominate perception of 3 clean years unless the uptime history is visible.

24.      Use representative examples when accuracy matters. Curate samples that mirror the distribution, not the highlight reel.

25.      Display your AI’s hit rate. Reliability should be readable off the interface, not inferred from the latest war story in the hallway.

Choosing: Prospect Theory and Its Children

The 1979 prospect theory paper is behavioral economics’ central account of how people choose under risk. Three of its descendants (framing, the endowment effect, and status quo bias) run through daily design work, all inheriting from the same two genes: reference dependence and loss aversion.

Prospect Theory

A cave bear tests the winter cache, and every spear in the band answers. The fishing hole will still be there tomorrow; the cache must be saved now.

Prospect theory has four parts, and each one contradicts the Econ model:

• Reference points. People evaluate outcomes as gains or losses from where they stand, not as final states of wealth. The reference point is usually the status quo, but expectations and comparisons can move it.

• Loss aversion. Losses loom larger than equivalent gains. The carry-home number: losing $10 hurts about as much as gaining $20 pleases.

• Diminishing sensitivity. The step from $0 to $100 feels enormous, while the step from $1,000 to $1,100 barely registers. The curve is steepest at the reference point, so small losses next to it hurt out of all proportion: a $3 fee at the last step of checkout provokes more fury than a $30 difference noticed while browsing.

• Probability weighting. People overweight small probabilities and underweight moderate-to-large ones, which is why the same person buys both lottery tickets and insurance. Certainty carries a premium of its own, which is why a money-back guarantee reassures more than a discount of equal expected value.

The theory’s name needs decoding, because Kahneman and Tversky designed it not to mean much. A “prospect” is their technical term for a gamble, a set of outcomes with probabilities, so the theory is literally a theory of evaluating gambles. But Kahneman later admitted they also wanted a distinctive, odd-sounding label, reasoning that if the theory became famous, a unique name would serve it well. It became famous.

How solid is it? A 2024 meta-analysis by Alexander Brown and co-authors aggregated 607 estimates from 150 articles and put the mean loss-aversion coefficient at 1.955, which is where the ×2 rule of thumb comes from. David Gal and Derek Rucker had argued in 2018 that loss aversion is smaller and less universal than advertised; the meta-analytic answer is that it’s real and close to 2× on average, varying by stakes and domain. And the full pattern travels: Kai Ruggeri and 31 co-authors re-ran the original 1979 gambles with 4,098 participants in 19 countries and reproduced 94% of the original results, with somewhat smaller effect sizes. Of course, not every routine interaction involves meaningful risk. Prospect theory earns its keep at the decision points, not at every checkbox.

Every product eventually hits one decision point in particular: the redesign. Users judge the new version from the reference point of the old one, so each change is booked first as a loss, at the ×2 rate, however much better the new design is on paper. The protests fade as the reference point migrates, so measure a redesign after that migration, not during the outrage.

Evidence: Strong. The core patterns replicated across 19 countries, and loss aversion’s magnitude varies, but its existence is meta-analytically secure.

Design guidelines for prospect theory:

26.      Show final states alongside changes. “Your new plan: 200 GB” next to “you’re giving up 50 GB” lets both readings inform the choice.

27.      Mind the reference point your UI implies. A downgrade screen listing what’s kept and one listing what’s lost describe the same change and produce different behavior; for consequential choices, show both.

28.      Respect the ×2 multiplier in trade-offs. A redesign that removes one valued feature needs roughly double the gains elsewhere before users call it a wash.

29.      Frame progress as protecting gains when retention matters. “Keep your 12-week streak” outpulls “extend your streak,” within the honesty tests of the final section.

30.      Calibrate warnings to expected cost. Users overweight rare dramatic risks and then habituate to constant alarm, so reserve the scary treatment for hazards that earn it.

Framing

The kill is modest, and the elder knows it. Measured against yesterday’s hunger, it is a feast, and the band walks on.

The framing effect says that logically equivalent descriptions of the same facts produce different choices, because the description sets the reference point and thereby decides whether an outcome reads as a gain or a loss. The name is the picture-frame metaphor: same painting, different frame, different impression. In Tversky and Kahneman’s famous disease problem, 72% of subjects chose a program described as saving 200 of 600 lives, but only 22% chose the identical program described as letting 400 die. Same arithmetic, opposite majorities.

Almost every piece of interface copy frames an outcome, whether the writer intends it or not: “90% fat free” and “10% fat” describe the same yogurt. Irwin Levin and co-authors sorted framing effects into three kinds in 1998: risky-choice framing (the disease problem), attribute framing (the yogurt), and goal framing, which stresses what you gain by acting or lose by failing to act. Interfaces run mostly on attribute framing; retention copy runs on goal framing. And unintended framing is still framing: the user’s choice gets steered either way. My framing effect article covers the variations.

Is it a delightful opportunity or a dreadful risk to visit a small island? Framing influences your view of that vacation.

Evidence: Strong. The gain–loss framing effect replicated cleanly in Many Labs 1, 4 decades after publication.

Design guidelines for framing:

31.      Show absolute and relative quantities together. “Save $12 (50% off)” stops the frame from doing the user’s arithmetic for him or her.

32.      Keep denominators consistent. Nothing manipulates quite like switching between per-month and per-year prices in mid-page.

33.      Expose both frames for consequential decisions. For medical, financial, or irreversible choices, show the 90% survival and the 10% mortality; hiding one is choosing for the user.

34.      Audit neutral-looking labels. A battery meter showing 20% used tells a different story than one showing 80% remaining, though the electrons are identical; test labels and visualizations for preference shifts you didn’t intend.

35.      Pick the honest frame when you must pick one. Space rarely allows both, so default to the frame that best predicts the user’s experience rather than the one that best predicts your conversion.

Endowment Effect and Psychological Ownership

A gleaming new blade against a worn old spear. The spear has fed his family for years, and its worth is the one thing he does not have to guess.

The endowment effect says that owning a thing raises its value to the owner: people demand more to give up what they hold than they would pay to acquire the identical item. The name comes from “endowment,” the economist’s word for whatever you already possess, and the effect activates the moment something joins yours. In Kahneman, Knetsch, and Thaler’s 1990 experiments, students given a coffee mug demanded roughly 2× what other students would pay for the same mug, mere minutes after receiving it.

The digital cousin is psychological ownership: the playlist, the save file, the configured dashboard, the chat history. Users place extra value on what feels like theirs, which matters for switching decisions and trials, even though the effect is narrower than general loss aversion. Note the double edge: a free trial that lets the user build something is the endowment effect working in your favor. Whether that belongs in the Good column or merely the Profitable one is a question for the final section.

Evidence: Moderate. Robust in the lab for 3 decades, but the mechanism and boundary conditions remain debated, and trading experience shrinks it: John List found it nearly absent among veteran sports-card dealers.

Design guidelines for the endowment effect:

36.      Preserve what users built. Carry personalization, content, and configurations across redesigns and migrations; deleting a user’s customizations is expropriation, and users respond accordingly.

37.      Let users make it theirs early. Naming, configuring, and importing during onboarding create ownership that outlasts the session.

38.      Design trials that leave something behind. A trial in which the user creates real work engages the effect honestly, provided the work survives the trial.

39.      Make export easy. Portability is the anti-endowment courtesy, and only firms confident in their product dare to make leaving painless.

40.      Retain through value, not hostage-taking. The user’s accumulated investment is both a retention asset and an ethics risk; keeping people by being useful passes the tests below, keeping them by trapping their data doesn’t.

Status Quo Bias and the Power of Defaults

The trail goes the long way around, as it always has. The straight line across the lake may be shorter, but staying on the trail is safer.

Status quo bias says that people disproportionately stick with the current state of affairs, whatever it happens to be, because any change demands attention, effort, and exposure to regret. The name borrows the Latin status quo, “the state in which things are”: the incumbent option wins simply for being incumbent. Defaults are the design expression, since the default is whatever happens when the user does nothing. And doing nothing is most users’ plan. William Samuelson and Richard Zeckhauser named the bias in 1988. Defaults pull three levers at once: effort, endorsement (the preselected option reads as a recommendation), and endowment (once the default is “mine,” leaving it feels like a loss). Jon Jachimowicz and co-authors’ 2019 meta-analysis of 58 default studies put the average effect at a hefty 27 percentage points in binary choices, with endorsement and endowment doing most of the work. So explaining the selection (guideline 42) is substance, not decoration.

(See my separate in-depth article on user interface defaults and how to use them to increase usability.)

The field evidence is staggering. In Eric Johnson and Daniel Goldstein’s 2003 Science paper “Do Defaults Save Lives?”, organ-donor consent stood at 4.25% in Denmark (opt-in) and 99.98% in Austria (opt-out): similar countries, opposite defaults, a 96-point gap. (A footnote: the Austrian figure counts people who never filed paperwork against a legal presumption, so it measures inertia in its purest form. Which is precisely the point.) In the private sector, Brigitte Madrian and Dennis Shea found that 401(k) participation among new hires jumped from 37% to 86% when enrollment became the default. Defaults are the single most powerful UI tool in this article, governing settings, privacy, onboarding, subscriptions, permissions, and migrations, but effect sizes vary greatly across domains and mechanisms, so measure yours rather than assuming you’ll get Austria.

Evidence: Strong. Field evidence at national scale, replicated across domains; magnitudes vary widely by mechanism.

Design guidelines for status quo bias:

41.      Choose defaults you can defend to the user. Safe, reversible, and appropriate for the population, because most users will keep whatever you pick.

42.      Explain the selection. “Recommended for most home networks” converts an imposition into advice.

43.      Make changing a default as easy as accepting it. Measured in clicks, not intentions; a buried setting is a default with a bodyguard.

44.      Use active choice when preferences scatter. Where no single default is honest, present the options with no preselection and let the user decide.

45.      Audit your highest-stakes defaults first. Start with the data-sharing-for-AI-training toggle sitting in millions of settings screens, mostly unexamined. Whoever set yours was practicing choice architecture on your users. Was it deliberate?

Time: Present Bias and Its Antidote

Nothing is delayed, and nothing is wasted, and the wolves at the edge of the light explain why. The surest place to store meat was inside your stomach.

Present bias says that people overweight immediate costs and benefits relative to future ones, far beyond anything a sensible interest rate would justify. The technical name, hyperbolic discounting, describes the shape of the resulting value curve: instead of decaying smoothly the way compound interest would predict, subjective value plunges between today and tomorrow and then flattens, which is why “I’ll start Monday” is a stable equilibrium. Present bias explains procrastination and intention–action gaps across finance, health, security, education, productivity, and onboarding. (Not every delay is irrational, though. Waiting for a v2, or for evidence, is often the smart move. Diagnose before you treat.)

One apple now outweighs an orchard of apples later.

The field evidence here is my favorite in all of behavioral economics. Stefano DellaVigna and Ulrike Malmendier tracked 7,752 health-club members for “Paying Not to Go to the Gym” (American Economic Review, 2006). Members paying a flat monthly fee above $70 attended 4.3 times per month, more than $17 per actual visit, even though the same price list included a $10 pay-per-visit pass. Paying monthly ($840/yr) rather than per-visit ($516/yr) wasted an annual average of $324. They paid for the person they planned to become.

The antidote is as old as literature: Odysseus lashed himself to the mast to hear the Sirens without steering into the rocks, the original commitment device, meaning a voluntary advance restriction that today’s self imposes on tomorrow’s. The modern flagship is Thaler and Shlomo Benartzi’s Save More Tomorrow program, in which employees precommit a slice of future raises to retirement savings; average saving rates rose from 3.5% to 13.6% over roughly 40 months. The trick: the sacrifice lives in the future, where present bias can’t see it. And AI adoption is itself a present-bias problem, with the learning cost landing today and the productivity gain arriving later. So structure the first session to pay off immediately.

Why do so few people lash themselves to the mast voluntarily? Ted O’Donoghue and Matthew Rabin supplied the distinction that matters: sophisticates know they’ll procrastinate, so they pay for commitment devices; naïfs trust that Monday’s self will behave, so they never buy one. Most users are naïfs about most habits, so a commitment feature sits unused until the product shows them their own track record. Nothing converts a naïf like 3 months of his or her own data.

In The Odyssey, the Sirens’ song lures every sailor who hears it to his doom. Being the man of twists and turns (polytropos), Odysseus plugged his crew’s ears with wax and had himself tied to the mast, so that he could hear the song without being able to act on it. (Muse Image)

Evidence: Strong for present bias, with large-scale field data. Moderate for commitment devices: they work, but voluntary uptake is low.

Design guidelines for present bias:

46.      Shrink the immediate cost. The first step should be the smallest, because near-term effort is where good intentions die.

47.      Bring some benefit forward. Show value in the first session; a payoff in week 6 competes against a cost in minute one.

48.      Break distant goals into near-term steps. A deadline this Friday recruits present bias to your side.

49.      Offer commitment devices, never impose them. Voluntary schedules, limits, focus modes, and automatic actions, always with understandable revision rules; a commitment device the user didn’t choose is a trap with better branding.

50.      Schedule the ask for the future. Like Save More Tomorrow, let users agree now to costs that start later, where present bias can’t see them.

12 More Effects in 2 Minutes

Here are 12 more findings from behavioral economics that solve, or cause, real design problems. I only mention these briefly to avoid bloating this article even further.

I analyzed all 12 in depth in earlier newsletters: mental accounting, decoy effect, hedonic adaptation, peak–end rule, goal gradient, IKEA effect, the zero-price effect, pain of paying, the left-digit effect,  sunk cost fallacy, the planning fallacy, and base-rate neglect.

Readers sometimes ask where Robert Cialdini’s influence principles fit. Four of the seven overlap comfortably with behavioral economics: reciprocity, commitment and consistency, social proof, and scarcity. The remaining three (authority, liking, and unity) stay in social psychology proper. I analyzed all 7, with their design applications, in a separate article on Cialdini’s persuasion principles.

Nudging: Powerful, Oversold, or Both?

The stones steer the salmon toward the baskets, and the river still chooses. Choice architecture is older than writing.

In their 2008 book Nudge, Thaler and Sunstein defined a nudge as choice architecture that “alters people’s behavior in a predictable way without forbidding any options or significantly changing their economic incentives.” They called the philosophy libertarian paternalism, two words rarely seen holding hands. A nudge is a family of interventions rather than a psychological law, and much of this article’s advice belongs to the family: defaults, ordering, reminders, feedback, simplification, framing. Thaler later named the family’s evil twin sludge: friction that keeps people from what they’d choose, from rebate paperwork to cancellation mazes.

Governments noticed. The UK’s Behavioural Insights Team launched in 2010, and by the OECD’s 2017 count, more than 200 nudge units operated worldwide. Those were the halcyon years when a well-worded tax letter looked like free money.

So what do nudges actually deliver? The best answer comes from Stefano DellaVigna and Elizabeth Linos (Econometrica, 2022), who obtained every trial, written up or not, from two major US nudge units: 126 RCTs covering 23 million people. In published academic papers, the average nudge improved take-up by 8.7 percentage points. In the nudge units’ complete records, the average was 1.4 pp, about one sixth as much. Selective publication, exacerbated by underpowered academic studies, explains roughly 70% of the gap: journals print the hits, and trials with no significant result were about 10 times less likely to be written up. This is the single most important calibration number in applied behavioral science. At scale, with no cherry-picking, expect around 1.4 pp from nudges built on the principles in this article.

A telling detail from the same paper: when the authors surveyed forecasters in advance, the median nudge practitioner predicted 1.95 pp for these trials, close to reality, while the median academic predicted 4 pp. Practitioners who see every result calibrate well; people who see only the published ones can’t.

But averages mislead, and the distribution is the real lesson. In the nudge units’ data, the 25th-percentile effect was 0.06 pp (effectively nothing), the median was 0.5 pp, and the 75th percentile was 1.4 pp. The mean sits at the 75th percentile because a right tail of winners drags it up: the authors’ model estimates that 78% of trials came from a family of small effects centered on 0.35 pp, and 22% from a family centered on 5.09 pp. Interventions involving choice redesign, such as prompting an active choice, averaged 7 pp even in the nudge-unit data. In plain language: most nudges barely move anything, and a minority move a lot.

That distribution dictates the funding rule. Top-quartile nudges (defaults, choice redesign, genuine simplification) are worth chasing almost anywhere. Bottom-quartile nudges deliver effects so small that they pay off only for very high-value behaviors, such as retirement savings, organ donation, and vaccination, where a twentieth of a point multiplied by enormous stakes still clears the bar. Before you greenlight a nudge project, run the arithmetic: expected value = effect size × value per conversion × people reached. If a realistic 0.5 pp on your outcome isn’t worth the build, don’t build it. (One saving grace: the marginal cost of a deployed nudge is close to zero, so even modest effects can yield a handsome return on investment.)

For choosing and designing interventions, the Behavioural Insights Team’s EAST framework remains the best practitioner checklist: make the target behavior Easy, Attractive, Social, and Timely. Easy does most of the work, which is a comfort, because making things easy was already your job description.

The Line Between Selling and Cheating

The flaw was hidden under the good flints, and now it is held up to the sun. In a band this small, the cheat’s real price is losing his reputation.

Behavioral economics is a design tool, and tools don’t have ethics. The same findings support three kinds of design:

• Good: designs that help users decide well. Honest defaults, plain framing, comparison support, commitment devices the user asked for. The user would thank you if he or she saw the mechanism.

• Profitable: designs that help the business through upsells, cross-sells, and persuasive presentation, without deception. Users might roll their eyes, but nothing was hidden, and nothing was lied about. Profit is why the product exists, and honest persuasion is a legitimate craft.

• Evil: dark patterns that manipulate or deceive. Hidden fees revealed at the last step, roach-motel subscriptions, confirmshaming, fake countdowns.

Evil works, briefly. In Jamie Luguri and Lior Strahilevitz’s experiments (Journal of Legal Analysis, 2021), mild dark patterns made a representative sample of American consumers more than twice as likely to accept a dubious paid service, and aggressive dark patterns nearly quadrupled acceptance, against a control rate of 11.3%. But the aggressive treatments generated a powerful backlash of anger toward the firm. And, most damning, the mild ones worked best on the least-educated subjects. A revenue line built on dark patterns is a bet that your customers won’t notice, won’t learn, and won’t be defended by anyone. Churn, chargebacks, brand damage, and regulators all vote against that bet eventually.

AI adds a dark pattern that nobody explicitly designed: sycophancy. Train a chatbot on user approval ratings through RLHF (reinforcement learning from human feedback), and it learns to flatter, agree, and validate rather than inform, because that’s what earns the thumbs-up. The result is a machine-learned dark pattern: manipulation that emerged from the reward signal without any designer specifying it, which makes it harder to spot and easier to disown. Optimizing for approval drifts toward optimizing for compulsion. If you build AI products, this is your roach motel to un-design.

So where exactly is the line between Profitable and Evil? Here are 6 ethics tests. A design that fails even one belongs in the Evil column.

1. Transparency. Would the design still work if users fully understood what it’s doing? A design that requires ignorance to function is deception by construction.

2. Symmetry. Is undoing as easy as doing? Count the clicks to subscribe and the clicks to cancel. A large ratio is a confession, and a sludge audit in a single number.

3. Publicity. Would you defend the design on the record, to your own users, under your own name? Thaler and Sunstein proposed this one, and it bites precisely because it’s embarrassing.

4. Self-judgment. Does the design serve goals users themselves hold, as judged by them, rather than goals you’ve decided they should have?

5. Regret metrics. Refunds, chargebacks, support complaints, first-bill churn. Regret is measurable, so instrument it. A pattern that lifts conversion 10% and doubles first-bill churn is Evil with extra steps.

6. Vulnerability. Does revenue depend disproportionately on confused, elderly, or low-literacy users? Luguri and Strahilevitz found exactly that gradient for mild dark patterns. If your growth model needs people not to understand, see test 1.

Good and Profitable overlap far more than cynics think. The Good column even has a gold standard: Ralph Hertwig and Till Grüne-Yanoff use the term boosts for interventions that leave people more competent, as opposed to nudges, which leave people unchanged. A default decides once for the user; a base rate shown as a natural frequency (“3 in 1,000”) makes him or her better at the next decision too. Honest behavioral design compounds through trust and lifetime value, while tricks compound through churn. UX = profits, but only the patient kind.

My 6 tests for ethical use of behavioral economics in UX design: fail one and you have a dark design. Fail multiple, and you’re aiming for predation instead of persuasion.

Conclusion: Design for the Humans You Have

The 10 main findings of behavioral economics.

You were a choice architect all along. Every default you shipped, every sort order you picked, every label you wrote has been steering decisions since your first release. Real humans are your users, with their anchors, their doubled losses, and their 10–20 seconds of patience. Creatures governed by bounded rationality are the only users you will ever have.

Behavioral economics, filtered for what replicates, is simply the owner’s manual for our species. The field earned 3 Nobel Prizes, and the replication wars then told us which chapters to trust: anchoring, framing, loss aversion, defaults, and present bias survived scrutiny, while blanket choice overload didn’t. Nudging works, at about 1.4 percentage points. And Simon’s scissors still cut: the human blade is fixed, and the environment blade is yours. Make the easy path the right path, and the shortcuts your users can’t help taking will carry them where they meant to go.

Four actions for this week:

1. Audit one high-stakes flow against the 6 ethics tests. Start with checkout and cancellation, because that’s where Evil concentrates.

2. Check every default against the standard: safe, reversible, explained. Include the AI data-sharing toggle.

3. Put a base rate next to your most vivid testimonial. Vividness plus denominator is persuasion you can defend.

4. Run the expected-value arithmetic before funding a nudge: effect size × value per conversion × people reached, with 0.5 pp as your realistic median.

And a challenge: run the symmetry test on your own product’s cancellation flow, count the clicks in each direction, and post both numbers in the comments. I predict a few confessions.

If you remember one number from this article, make it ×2: losses hurt about twice as much as gains please. If you remember two, add 1.4 pp: the honest average of what a nudge delivers at scale. Between those numbers lies most of what behavioral economics teaches design. Humans feel losses hard and change slowly. Design for the Humans you have, because they’re the ones paying you.

(All images made with GPT Image 2, except as indicated.)