Designing AI for Willing Action: Engagement Without Coercion
AI engagement is strongest when action remains genuinely willing: users understand the recommendation, can decline it, and still feel that the system serves their goals.
The right engagement target is willing action, not maximum compliance
AI systems can recommend, persuade, personalize, remind, and increasingly act. That makes a simple product metric such as “did the user comply?” inadequate. A high conversion rate can be produced by clarity and value, but also by confusion, pressure, hidden alternatives, or fatigue. Self-Determination Theory offers a better motivational lens because it distinguishes autonomous motivation from more controlled forms and emphasizes the role of autonomy, competence, and relatedness.[1] Designing for willing action means the person understands what is being proposed, sees a realistic alternative, and still chooses to proceed because the action fits a goal they endorse. Engagement becomes evidence of alignment rather than victory over resistance.
Consent is not the same as absence of refusal
A user can click “yes” after being cornered, exhausted, or misled. Willing action requires a path to “no” that is visible, understandable, and proportionate.
Dark patterns reveal how easy it is to preserve formal choice while undermining real choice
The Federal Trade Commission’s report on dark patterns documents interface tactics that can trick or manipulate consumers, including deceptive presentation, hidden information, difficult cancellation, and false urgency.[2] AI interfaces can reproduce these tactics in language. An assistant can repeatedly reframe a refusal, bury the downside of an action, overstate consensus, or create artificial urgency in a personalized tone. Because the system may appear helpful or socially responsive, users can interpret pressure as advice. Product teams should therefore test conversational flows for the same properties they would test visual interfaces: truthful framing, symmetry of options, easy exit, and clear disclosure of commercial interests.
Respectful systems do not negotiate against the user’s boundary by default
When the user declines an optional action, the assistant should usually accept that answer. Re-persuasion should be justified by new information, not by the product’s desire for conversion.
The AI Act shows that manipulative design can cross from ethics into regulation
The EU AI Act prohibits certain AI systems that use subliminal, purposefully manipulative, or deceptive techniques to materially distort behavior by appreciably impairing informed decision-making when significant harm results or is reasonably likely.[3] The exact legal scope depends on context, but the design implication is broader: AI persuasion is not ethically neutral simply because the user remains capable of clicking. Systems that learn how to exploit attention or vulnerability can create risks that ordinary static interfaces did not have at the same scale.
Risk grows when optimization is individualized
A system that tests thousands of messages can discover which one moves a specific person. Governance must decide which objectives the system is allowed to optimize and which tactics are prohibited even if they work.
Human oversight and determination should remain visible at consequential boundaries
UNESCO’s Recommendation on the Ethics of Artificial Intelligence emphasizes human rights, dignity, transparency, fairness, and human oversight and determination.[4] In product terms, meaningful human control requires more than an emergency stop hidden in settings. Users should know when the AI is recommending versus executing, which actions need approval, what can be reversed, and who is accountable for the outcome. Consequential choices—financial commitments, public communication, data sharing, destructive operations, or major life decisions—should have friction proportional to the stakes.
Friction can protect autonomy
Behavioral design often aims to reduce friction, but zero friction is not always desirable. Confirmation and reflection are useful when an action is expensive, irreversible, public, or identity-changing.
Persuasive technology ethics distinguishes convincing from manipulation
Philosophical work on persuasive technology has examined the boundary between communication that helps people make decisions and strategic influence that undermines autonomy.[5] AI products live directly on that boundary because generated language can adapt arguments in real time. A responsible system can explain reasons, surface tradeoffs, and invite reflection. A manipulative system can exploit asymmetry: it knows the user’s history, tests persuasive variants, and has no fatigue while the user does. Design should therefore preserve the user’s ability to understand and contest the influence attempt.
Defaults should express user benefit and remain easy to reverse
AI products need defaults because asking users to configure every choice defeats the value of assistance. The question is whose interest the default serves. A good default reflects the user’s stated goal, common safety practice, or reversible convenience. It should not quietly maximize data collection, spending, or engagement at the user’s expense. Personalized defaults can be even better when the system explains the basis and provides a one-step override. The test is whether the product would still endorse the default if the user fully understood how it was chosen.
Urgency should correspond to real-world urgency
AI can generate extremely persuasive urgency, so the system should be grounded in actual deadlines, capacity limits, risks, or commitments. “Your application closes tomorrow” can be helpful if true. “Act now or lose your chance” is manipulative when the chance is artificial. Even true urgency should be calibrated to the user’s goals and emotional state. The system should inform, not catastrophize. When stakes are uncertain, it should communicate uncertainty instead of converting possibility into pressure.
Measure regret, trust, and goal attainment alongside engagement
A product designed for willing action needs metrics that can detect when engagement is being purchased with future dissatisfaction. Teams can measure opt-out friction, reversal rates, complaint rates, post-action regret, perceived control, trust, and progress toward user-defined goals. Qualitative research should ask whether people felt helped or pushed. The strongest signal is not that users always accept the AI’s suggestion. It is that they can reject it comfortably and still return because the system remains useful. Willing action creates a durable relationship: the AI earns influence through relevance, transparency, and respect rather than by making resistance progressively harder.
Works Cited
- 01Self-Determination Theory — The Theory selfdeterminationtheory.org
- 02
- 03European Union — Artificial Intelligence Act, Regulation (EU) 2024/1689 eur-lex.europa.eu
- 04
- 05Spahn — Persuasive Technology and the Ethics of Communication pubmed.ncbi.nlm.nih.gov
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.