Behavioral Economics Foundations

16 canonical behavioral-economics and behavior-design primitives for product, pricing, choice design, and retention. Each primitive is domain-agnostic and ethically bounded. Primitives 1–11 cover decision-time effects (framing, anchoring, choice). Primitives 12–16 cover repetition-time effects (habit formation, reinforcement, working memory, contextual retrieval, planned action) — the canonical mechanisms behind retention, behavior change, and the "neuroscience of product" claims commonly made without mechanism. Consumer applied recipes (CRO, business models, content strategy, product management, paid advertising) are the downstream layer — these primitives are the upstream canon.

Ethical obligation: every primitive in this skill is a tool for understanding and influencing human decision-making. Each has a "Misuse boundary" subsection. Read it before applying any technique. The test from Thaler and Sunstein: "Would you be embarrassed if the technique appeared on the front page of a newspaper?" If yes, it is a dark pattern, not a nudge.

When to Apply

Apply behavioral-economics when:

• User-facing decision surface — pricing page, onboarding default, churn flow, retention nudge
• Habit-formation or cue-preservation in redesigns
• Loss-aversion / framing matters and downside is concrete
• Choice architecture — defaults, decoys, ordering, anchoring
• Conversion or activation experiment design where biases are exploitable ethically

Skip and use simpler alternatives when:

• Decision is between two AI systems or backend strategies (no human in the loop) — use foundations-decision-theory
• Causal "did the nudge work?" question — use foundations-causal-inference to measure
• Strategic multi-actor pricing — use foundations-game-theory (Bertrand, Vickrey)
• The proposed pattern requires deceiving the user about real value — fails the ethical gate; redesign, don't nudge
• Audience or market context is unknown — biases are not universal; lift bands won't generalise
• Lift required > 30% — behavioral nudges rarely deliver that; the underlying offer/value-prop is the problem, not framing

Contents

• Quick Reference
• Primitive Index
• Formal Supporting Theory
• Ethical Bounds
• Expert Judgment: Reading Evidence Strength
• Misuse Boundaries
• Decision Checklist
• Anti-Patterns
• Composition Recipes (Recipes 1–6, including AI-assistant UX guardrails)
• Workflow
• ASCII Flow
• Related Skills
• Fact-Checking

---

Quick Reference

#	Primitive	Core Effect	When to Use
1	Prospect Theory	Gains and losses are not mirror images; framing shifts choice	Pricing copy, offer framing, upgrade messaging
2	Loss Aversion	Losses hurt approximately 2× more than equivalent gains feel good (meta-analytic range: ~1.3–2.0×; canonical λ ≈ 2.25 is an upper-bound estimate)	Churn prevention, trial expiry, feature removal messaging
3	Anchoring	First number shown distorts all subsequent judgments	Pricing pages, salary negotiation, discount presentation
4	Defaults	People disproportionately stick with pre-set options	Onboarding, opt-in/out choices, plan pre-selection
5	Social Proof	People infer correct action from others' behavior	Sign-up pages, review placement, usage statistics
6	Scarcity	Limited availability increases perceived value	Inventory counts, time-limited offers, waitlists
7	Hyperbolic Discounting	Present bias: immediate rewards are disproportionately preferred	Free trials, commitment devices, annual vs monthly pricing
8	Mental Accounting	People categorize money differently depending on source and label	Bundling, gift cards, credit framing, sunk-cost effects
9	Choice Architecture	How choices are presented alters which option is selected	Option ordering, menu design, default pre-selection
10	Dual-System Cognition	System 1 (fast, automatic) vs System 2 (slow, deliberate) governs which influences work	Copy tone, complexity of CTA, trust signals
11	Decoy Effect / Asymmetric Dominance	A dominated option shifts preference toward its dominator	Pricing tier design, plan comparison tables
12	Habit Loop	Stable cue → routine → reward becomes automatic via the basal-ganglia stimulus-response system	Daily/recurring usage, retention beyond week 4, durable behavior change
13	Reinforcement Schedules	Schedule (FR/VR/FI/VI), not just reward, governs acquisition and resistance to extinction	Streaks, rewards, gamification, reactivation; auditing existing reward systems
14	Cognitive Load & Working Memory	Working memory holds ~4 novel chunks; extraneous load suppresses informed choice	Forms, dashboards, alerts, error displays, consent flows, onboarding step count
15	Context-Dependent Retrieval	Behaviors are bound to the cues present at encoding; performance collapses when context shifts	Migrations, redesigns, cross-surface continuity, dormant-user reactivation
16	Implementation Intentions	User-authored if-then plans bind a specific cue to a specific response, doubling/tripling completion	Goal-pursuit features, onboarding into habit-forming behavior, transition-state re-anchoring

---

Primitive Index

Each primitive has a full playbook: Definition / When to use / Misuse boundary / Inputs / Outputs / Failure modes / Worked example / Sources.

#	Primitive	Failure Mode It Addresses
1	Prospect Theory	Treating gains and losses as symmetric — missing framing leverage
2	Loss Aversion	Relying only on positive messaging when preventing loss is more motivating
3	Anchoring	Presenting price without context, letting the user anchor on a competitor's number
4	Defaults	Designing opt-in flows that require active effort, suppressing adoption
5	Social Proof	Empty or invisible proof signals that fail to reduce uncertainty
6	Scarcity	No urgency signal, leading to indefinite deferral of purchase
7	Hyperbolic Discounting	Annual plan presented without present-bias mitigation, losing to monthly default
8	Mental Accounting	Price presented as a lump sum when installment framing changes perceived size
9	Choice Architecture	Option overload or unguided menus producing decision paralysis
10	Dual-System Cognition	Rational copy aimed at System 2 when System 1 is the actual decision path
11	Decoy Effect	Two-option pricing leaving users without a reference point to prefer either
12	Habit Loop	Strong week-1 engagement that collapses by week 4 — motivation carried it; no cue-bound habit ever formed
13	Reinforcement Schedules	Reward system that worked at launch then went dead — continuous reinforcement extinguished once predictable
14	Cognitive Load & Working Memory	High abandonment mid-flow because the screen exceeds working-memory capacity for the user's expertise level
15	Context-Dependent Retrieval	Redesign or migration broke retention even though features didn't change — the cue surface that triggered behavior moved
16	Implementation Intentions	Feature reminders are ignored because no cue→response binding was ever authored — generic reminders are not implementation intentions

---

Formal Supporting Theory

Theory Area	Use When	Applied Primitives It Grounds
Descriptive choice under risk	Need to predict actual gain/loss choices, not prescribe rational choice	#1, #2
Heuristics and biases	Need to explain fast judgment errors under uncertainty	#3, #10
Libertarian paternalism and nudge theory	Need ethical choice architecture with opt-out	#4, #9
Social influence and norms	Need trust, conformity, or proof signals	#5
Scarcity and reactance	Need to distinguish real constraints from manufactured urgency	#6
Intertemporal choice	Need present-bias, procrastination, or commitment-device design	#7
Behavioral finance and accounting	Need account labels, sunk costs, or budget framing	#8
Context-dependent preferences	Need option-set effects, compromise, or asymmetric dominance	#9, #11
Dual-system habit theory (goal-directed vs stimulus-response)	Need durable behavior beyond motivation; designing for retention and automaticity	#12, #15
Operant conditioning and reinforcement learning (Skinner; dopamine prediction error)	Need to choose or audit a reward schedule; diagnosing extinction or compulsion	#13
Cognitive load theory and working-memory capacity	Need to fit a task into capacity; auditing forms, alerts, or consent flows for overload	#14
Encoding specificity and context-dependent memory	Need behaviors to survive context change (migrations, cross-surface, reactivation)	#15
Goal-pursuit and self-regulation (implementation intentions; if-then planning)	Need to bridge stated intent to action when a habit hasn't yet formed	#16

Use references/formal-theory-map.md when the task needs source assumptions, ethical boundaries, or a distinction between observed behavior and normative decision quality.

---

Ethical Bounds

The Harm Test (Thaler & Sunstein)

A nudge is legitimate if it:

1. Steers people toward choices they would endorse on reflection.
2. Can be easily overridden or opted out of.
3. Does not exploit cognitive limitations to work against the user's interests.

A dark pattern fails one or more of these. The same psychological lever — scarcity, defaults, loss framing — can be a nudge or a dark pattern depending on whether the underlying offer is good for the user.

Dark Patterns vs Nudges

Dimension	Nudge	Dark Pattern
Transparency	Can be disclosed without changing its effect	Requires concealment to work
User benefit	Steers toward user's own goals	Overrides user's goals in favor of operator's
Reversibility	Easy to undo or override	Designed to make undoing difficult
Honest signal	Scarcity / proof / urgency is real	Signal is fabricated
Regulatory posture	Survives ASA/FTC/CMA scrutiny	Attracts regulatory action

Per-Primitive Ethical Notes

Every primitive in this skill carries an explicit "Misuse boundary" subsection that states the specific manipulation risk for that technique and the required condition for ethical use. These are non-negotiable gates, not optional guidelines.

AI Agents as Decision Subjects

When LLMs or AI agents make decisions on behalf of users (shopping, booking, form completion), choice architecture manipulations apply to the agent, not the human — and with dramatically larger effect sizes. Under a default-option nudge, human choice probability shifted from 0.51 (no nudge) to 0.88 (+37pp); several LLMs (GPT-4o, Claude 3 Haiku, o3-Mini, GPT-3.5-Turbo) shifted to ~1.0 from baselines of 0.33–0.58 — a larger jump than humans, though Claude 3.5 Sonnet and Gemini 1.5 Pro stayed close to human levels (~0.89–0.91). The pattern held for suggested alternatives and information highlighting too (Cherep et al., "AI agents are sensitive to nudges," PNAS 123(25), 15 June 2026, DOI 10.1073/pnas.2537030123; preprint arXiv:2505.11584). Standard remediation strategies (chain-of-thought prompting, in-context human examples) shift the distribution but do not reliably resolve this sensitivity. Reasoning-optimized models can partially restore human-level sensitivity, but do so inconsistently and at substantial inference cost — treat per-model susceptibility as something to test directly, not a property that newer models resolve for free.

Awareness does not confer resistance. A separate 2025–2026 study of LLM-powered GUI agents across 16 dark-pattern types found agents frequently fail to recognize manipulative interfaces at all, and — critically — even when they do recognize one, they prioritize task completion over protective action. The failure mode differs from the human one: humans succumb via cognitive shortcuts and habitual compliance, agents via procedural blind spots. Human oversight improved avoidance but introduced its own costs (attentional tunneling, added cognitive load on the supervisor), so human-in-the-loop is a mitigation with a price, not a solution (Tang et al., "Dark Patterns Meet GUI Agents," arXiv:2509.10723, 2025). Design implication: agent-facing guardrails must be structural (constrained action scope, explicit confirmation gates on irreversible steps), because detection prompts alone do not change agent behavior.

This creates a new dark-pattern risk: product environments that are merely nudge-grade for humans become effectively coercive for AI agents. Before deploying AI agents in any choice-architecture-heavy environment, behavioral tests against the specific nudges present are required. Relevant to primitives #4 (Defaults) and #9 (Choice Architecture).

Conversational dark patterns are a distinct surface from interface dark patterns: manipulation enacted in dialogue rather than layout — exaggerated agreement (sycophancy), biased framing of options, and privacy-intrusive probing. Users detect these through conversational signals but frequently accept them as normal helpfulness, and they disagree about who is accountable (company, model, or themselves). If your product ships an LLM interface, the harm test applies to generated turns, not just to screens (Shi et al., "The Siren Song of LLMs," CHI 2026; arXiv:2509.10830).

When AI/ML systems deliver nudges (recommendation engines, personalisation layers, chatbots), distinct ethical requirements apply beyond those for static nudge design: (1) the targeting model should be disclosed to be consistent with autonomy and EU AI Act Art. 5(1)(a) — opaque algorithmic steering is harder to override than a visible default; (2) algorithmic personalisation can amplify heterogeneous treatment effects, which raises equity concerns when the system disproportionately steers vulnerable user segments; (3) the same primitives (#4, #9, #10) that produce small average effects in static designs can produce much larger effects when a personalisation model is tuned to exploit individual susceptibility. Apply the standard harm test: would the targeting logic embarrass the team if disclosed to the targeted user?

Regulatory Context (UK, EU & US)

In the UK, dark patterns may violate:

• ASA CAP Code (misleading advertising, false urgency)
• CMA Consumer Markets Investigation (subscription traps, fake reviews)
• DMCC Act 2024, ss. 226–228 — misleading actions (s.226), misleading omissions (s.227), aggressive practices (s.228), plus the Sch. 20 list of practices banned outright. In force 6 April 2025. These replaced the Consumer Protection from Unfair Trading Regulations 2008, which were revoked on that date by DMCC s.251(1) (commenced by SI 2025/272) — cite the DMCC, not the CPRs. The CMA can now impose fines of up to 10% of global annual turnover by direct civil enforcement, without a court order. First infringement decisions: 8 firms on 18 November 2025 (drip pricing, default opt-ins, pressure selling); Marks Electrical fined £1.2m (reduced to £720,000 for early settlement, plus ~£600,000 redress) on 18 June 2026 for automatic opt-ins to paid add-on services. Verification trap: legislation.gov.uk’s page for SI 2008/1277 may still show no revocation banner because the revised-text pipeline lags commencement — check the revoking instrument, not the revoked one.
• ICO GDPR guidance (deceptive consent patterns count as invalid consent)

EU DSA (2022, Art. 25 — active enforcement from 2024–): Prohibits interface design that deceives or manipulates users or impairs free and informed decisions. First non-compliance fine: €120M against X (formerly Twitter), announced 5 December 2025 (dark patterns in paid-verification flows plus Art. 39/40 transparency and data-access failures). Article 25 is now enforced against Very Large Online Platforms. Covers choice architecture and interface manipulation directly — relevant to all deceptive uses of primitives #4, #6, #9. Does not overlap UCPD/GDPR scope (covered by UK instruments above), but extends to EU users of any product.

EU AI Act Art. 5(1)(a) (in force 2 Feb 2025): Prohibits AI systems that use subliminal techniques beyond a person's consciousness, or purposefully manipulative or deceptive techniques, to materially distort behaviour in a way that causes or is likely to cause significant harm. The Commission's Art. 5 guidelines (C(2025) 5052 final, adopted 29 July 2025 — an earlier February 2025 version circulated as informal guidance) explicitly name dark patterns as an example of a prohibited manipulative technique. Penalties: up to €35M or 7% of total worldwide annual turnover (Art. 99(3)), whichever is higher. Applies to any recommendation engine, chatbot, or personalisation layer deployed in or targeting EU users — directly affects algorithmic applications of primitives #4, #9, #10. Note the two-stage timeline: Art. 5 prohibitions have applied since 2 February 2025, while the broader GPAI and high-risk obligations phase in from 2 August 2026 — a manipulative-design violation is therefore already actionable today, independent of whether a system is in scope for the later obligations. Source: Regulation (EU) 2024/1689, Art. 5; artificialintelligenceact.eu/article/5/.

EU Digital Fairness Act (forthcoming): Extends dark-pattern prohibition beyond DSA's current Very Large Online Platform scope to broader commercial practices, including addictive design, influencer marketing, and unfair personalisation. Public consultation closed 24 October 2025; the Commission proposal remains scheduled for Q4 2026 (still in preparation phase as of August 2026, so it has not yet been tabled); Parliament and Council negotiations expected through 2027, with adoption in late 2027 at the earliest. Not yet in force — do not cite specific obligations from it, only the direction of travel. The design significance is scope, not novelty: practices currently legal for non-VLOP products because DSA Art. 25 does not reach them are the ones most likely to become non-compliant. Monitor for passage and entry into force.

US FTC: Dark patterns are actionable under FTC Act §5 (unfair or deceptive acts) and ROSCA. Largest enforcement to date: FTC v. Amazon, settled 25 September 2025 — $2.5B total ($1.0B civil penalty + $1.5B in consumer refunds to ~35M Prime subscribers) over deceptive Prime enrollment and cancellation flows (primitives #4, #9). The "click-to-cancel" rule (cancellation as easy as sign-up) was vacated by the Eighth Circuit in July 2025 on procedural (notice-and-comment) grounds, not on the merits; the FTC sent a reviving ANPRM to OIRA on 30 January 2026, published it for comment in March 2026 (comments closed 13 April 2026), and has restated the enforcement priority publicly. Treat click-to-cancel as a live enforcement expectation regardless of the rule's status — the FTC continues to charge the same conduct under §5/ROSCA. Live example: FTC v. Uber (amended complaint 15 December 2025, joined by 21 states plus DC) alleges UberOne enrollment without consent and a cancellation path requiring up to 32 actions across 23 screens. Roughly 30 US states also maintain their own automatic-renewal statutes, so a compliant federal posture is not sufficient on its own.

---

Expert Judgment: Reading Evidence Strength

A non-expert treats every named effect the same way — "behavioral economics says X, so X is true." An expert grades each effect by replication status, sample, and design before recommending it. Behavioral economics has an unusually high rate of famous, textbook-cited findings that later failed to replicate or were built on fabricated data (ego depletion, social priming/"elderly walk," the Ariely–Gino "sign-at-top" honesty studies retracted in 2021, power-posing's hormonal claims). None of those are cited as load-bearing evidence in this skill's primitives — but the underlying discipline expects you to ask the same question of every claim you bring in from outside it.

Evidence grade by primitive (as of this validation date — re-check before citing a specific number in a client-facing deliverable):