Insights / Consumer Goods

The conditions for staying

Trust in AI is not measured at adoption but in whether people stay. See how self-determination, curiosity and attachment decide it.

Team Stefanini ・ 14 Oct. 2026 ・ Estimated read: 23 min

Human Judgment, Self-Determination, Curiosity and the Lifecycle of AI Trust

Abstract: Most conversations about AI adoption concentrate on entry: acquisition rates, onboarding completion, and first-week activation. Far fewer examine the extended period between initial engagement and the moment a person stops using the tool. This article argues that AI abandonment is the terminal stage of a predictable psychological process involving progressively shrinking self-determination, unresolved distrust, and the collapse of whatever functional attachment had formed between the person and the system. A critical dimension of this process is the behavioral uncanny valley: AI systems that simulate care without preserving human agency generate attachment to a relationship that cannot sustain itself, producing structural trust failure when the gap between the experienced and the actual relationship becomes apparent. Using: Self-Determination Theory, Attachment Theory, The Extended-Self Framework, Speech-Act Theory, and an original empirical framework, the Process of Realizing Empathy, tested in a controlled human-AI agent study, this article examines what keeps people genuinely engaged over time. Why care-based trust requires human agency to become sustainable, and what organizations must build to design the conditions under which both can coexist. Sustainability is an output of this relationship to be consistently considered. Understanding this as an organization will lead us to build commercially durable forms of AI user loyalty.

Measuring the wrong moment

There is a pattern in AI product data that most organizations have not yet been able to act on. Adoption metrics capture entry. Session analytics capture presence. Neither measures the quality of engagement over time, and neither detects the specific moment at which a person who is technically still using a system has already decided, below the level of conscious articulation, that they do not trust it enough to bring their real work to it.

Disengagement from AI systems follows a recognizable shape in practice. The account remains open. The subscription continues. The dashboard logs an active user. The person is testing rather than trusting, verifying outputs rather than building on them, submitting only the tasks which are low stakes enough to risk being wrong. They are managing the relationship rather than investing in it. The behavioral signals are present in product data: declining interaction depth, narrowing task scope, increasing output verification, and decreasing investment of personal context. Most organizations are not instrumented to surface these signals, and so they watch for the exit while the decision to leave is being made weeks or months earlier.

Research on the abandonment of connected consumer technologies documents the shape of this pattern: 65 percent of activity tracker users stopped within two weeks (Shih et al., 2015), and participants in a study of self-selected smart devices abandoned almost 80 percent of them within two months, attributed not to technical failure but to the absence of the relational architecture that sustains genuine engagement (Lazar et al., 2015). These were not AI products; but they share a common mechanism, and the behavioral signals described above are its AI-native expression. The cost of that failure compounds: the cost of AI customer acquisition is substantial; the cost of rebuilding trust after structural design failures is disproportionate to what different upstream investment would have required (Mehrotra et al., 2024). The question this article addresses is what that upstream investment needs to contain, grounded in what motivational, relational, and communicative evidence tells us about how trust in AI systems forms, sustains, and fails.

The motivational architecture of genuine engagement

Self-Determination Theory proposes that sustained engagement in any domain depends on the satisfaction of three basic psychological needs: autonomy, the sense of being the originating agent of one’s own actions; competence, the sense of being effective and capable within the interaction; and relatedness, the sense of genuine connection and responsiveness in the exchange (Deci & Ryan, 2000; Ryan & Deci, 2017). When these needs are met, motivation is intrinsic and engagement sustains. When any of them is chronically frustrated, the person’s relationship with the activity degrades along a predictable trajectory.

For AI systems, each need carries precise design implications. Autonomy is frustrated when systems minimize the user’s deliberative space: when outputs arrive with confidence that forecloses reflection, when defaults guide rather than support the user’s own judgment. Research on AI decision-support systems found that reduced reflective capacity is produced when systems fail to signal their own failures, leaving users in an intuitive mode of flow precisely where deliberate scrutiny is required, and that restoring it calls for deliberate friction in the interaction (Buijsman et al., 2025), the structural source of the pattern in which people who use AI tools heavily report reduced confidence in their own judgment over time.

Competence is frustrated when systems are opaque: when errors are unexplained or framed in ways that transfer responsibility to the user, when behavior is inconsistent enough to prevent the user from developing a stable working model. Users who cannot build a reliable understanding of how a system behaves remain permanently provisional in their relationship with it. Provisional relationships do not deepen and do not survive meaningful alternatives.

Relatedness, the need AI products most commonly simulate and least reliably meet, requires that the interaction feel genuinely responsive to this specific person. Personalization as a technical feature does not satisfy this need. What satisfies it is the experience that what this person brings to the interaction actively shapes what they receive. This distinction, between simulated responsiveness and genuine relatedness, is the motivational precondition for the attachment dynamics examined below.

Curiosity as the bridge between entry and depth

Curiosity is one of the most reliable early predictors of sustained engagement with new systems. It functions as the motivational mechanism that converts initial exposure into the exploratory depth that produces genuine value, and it is the psychological condition that enables users to develop the interaction fluency that makes a system worth staying with (Litman, 2008). For AI products, it is also the condition most systematically undermined by how those products are designed.

Curiosity requires a specific environment to sustain itself: a productive gap between what is known and what might be known, wide enough to motivate exploration but narrow enough to remain within reach. Most AI systems close that gap as efficiently as possible, responding comprehensively and immediately, with confidence that signals the exploration is complete. Scholarship on creativity-oriented research argues that framing creative work as a problem to be solved renders invisible the situated, negotiated practice through which it is actually done, and warns that creative AI built on that framing threatens to supplant the practices it claims to support (Hsueh et al., 2024). Better AI outputs do not automatically produce more engaged users. The structure of the interaction, the cognitive work the system takes and what cognitive work it leaves for the person, determines the motivational trajectory of the relationship.

Curiosity in human-AI interaction is also operationally linked to expectation regulation: the process through which a person progressively calibrates their understanding of how a system behaves and builds toward matched expectations. Curiosity-driven testing behavior is one of the primary mechanisms through which that calibration occurs and through which the trust that sustains long-term engagement begins to form.

The object of trust: attachment, extended self, and the resilience of relationships

Trust does not operate uniformly across all relationships. The difference that matters most commercially is not the degree of trust at any given moment but its resilience: whether it bends or breaks under the conditions of failure, inconsistency, and disappointment that every AI system will inevitably produce. The mechanism that determines that resilience is attachment. Bowlby established that what determines whether a bond survives separation, disappointment or failure is not its intensity but the security of the expectations built into it (Bowlby, 1982). That work has since been extended well beyond human bonds, to places, objects, and the tools people work with including AI.

The conditions under which functional attachment forms are predictable: consistent availability, demonstrated responsiveness to specific individual signals, and the accumulation of a shared relational history that makes the relationship feel specific to this person. When these conditions are present, the relationship acquires what trust theorists describe as goodwill-based trust: a form of relational confidence that persists through momentary failures (Lewicki et al., 2006). Russell Belk’s foundational work on possessions and the extended self established that objects and tools a person invests cognitive and emotional energy in become incorporated into their sense of self and capability (Belk, 1988), and his subsequent account of the digital extended self found the same logic applies to digital tools and AI systems, with additional dimensions specific to data-mediated intimacy (Belk, 2013). AI systems generate the conditions for extended self incorporation from the earliest stages of meaningful use. The question for product design is whether the system has been designed to honor the relationship it is generating.

The complexity of attachment in human-AI contexts deserves acknowledgment. Empirical research on affective attachment in human-AI agent contexts found that attachment indicators were present across both active and passive collaborative conditions, attachment forms across conditions of use. What differentiated the conditions was qualitative richness: in active collaborative settings, users used plural first-person language spontaneously and described shared outcomes, while in passive conditions the predominant perception was of the AI agent as “individual and independent” (García Corretjer et al., 2022). Attachment is not an automatic consequence of prolonged exposure. It deepens specifically through active collaboration, a finding established early in the research on long-term human-computer relationships, where relational behaviour rather than duration predicted whether people chose to continue wanting to engage with artificial actor. Organizations whose AI products are designed on a passive model are generating surface attachment that will not withstand significant failure.

The clinical research on the working alliance provides the most empirically rigorous anchor for the relationship between attachment quality and outcome resilience. Bordin established that the therapeutic alliance consists of agreement on goals, agreement on tasks, and the quality of the relational bond (Bordin, 1979). Wampold and Imel’s meta-analyses found that the bond component accounts for outcome variance significantly beyond what specific techniques predict (Wampold & Imel, 2015). The same pattern holds against a machine: in a month-long study with 101 participants, a relational agent differed from an equivalent task-oriented one on the bond dimension alone, and that difference held at four weeks (Bickmore & Picard, 2005). When extended self incorporation occurs without design that honors the relationship, the consequences of system failure exceed what organizations typically anticipate. At scale, extended self disruption, through major AI failures, privacy violations, or service discontinuation, converts product failures into brand events and trust violations into churn that resists conventional retention intervention (Belk, 2013).

A framework for realizing empathy: empirical evidence for active AI design

AI organizations face a specific and commercially urgent design question: given what motivational psychology and attachment theory tell us about why trust forms and erodes, what does an interaction architecture that builds those conditions actually look like? Original research on long-term human-technology relationships offers a precise and empirically tested answer. The Process of Realizing Empathy (RE) is a structured framework for designing interactions that generate the conditions under which trust, affective attachment, and genuine mutual understanding develop over time (García Corretjer, 2022).

The applicability of this framework to AI product design rests on direct equivalence. The empirical study testing the RE model used an AI agent as the interactive counterpart, a system capable of natural language exchange, adaptive response, affective expression, and collaborative negotiation. Social AI agents of this kind and the conversational AI, decision-support, and recommendation systems at the center of current commercial investment share the same underlying computational architecture, the same requirements for natural language understanding, and the same relational dynamics. The psychological mechanisms under examination operate at the level of human cognition, not at the level of the interface form through which the AI presents itself. The evidence transfers directly.

The RE framework identifies three stable conditions that must be consistently present: an interaction environment suited to the context and the person, a common language or communication platform through which genuine exchange can occur, and an open posture that preserves what the user brings without predetermining the outcome. Within those stable conditions, the RE process moves through four steps, cycling repeatedly as the relationship develops.

The process of realizing empathy: four steps applied to AI interaction design

REGULATE EXPECTATION

Each actor enters the interaction having reflected on what they are about to engage with, allowing their initial assumptions to remain flexible. For AI systems: transparent capability framing, honest onboarding, and the clear signal that the system’s outputs are a starting point for the user’s judgment rather than a conclusion that forecloses it. When absent users arrive with unmanaged expectations that the first failure confirms, generating immediate trust debt that is rarely recovered.

LISTEN

Active, wide collection of the other’s expression without judgment or preemptive response. For AI systems: interaction architectures demonstrating genuine responsiveness to this specific user’s inputs rather than pattern-matching to a statistical average. When absent: the perception most commonly described by disengaging users is that the system is “not responsive to me”, the precursor to progressive withdrawal.

CONSIDER

An internal process of reflection through which the received information is processed against the actor’s own understanding before a response is formed. For AI systems: interaction rhythms that leave space for the user’s own deliberative process, pacing and restraint that signal the user’s reasoning is expected to participate, not merely receive. When absent: the structural source of autonomy diminishes; the system fills the space where reflection once lived.

ACT

The external expression of what the consideration process has produced. For AI systems: output calibrated to this specific user’s context, with appropriate acknowledgment of uncertainty. When absent: outputs arrive without relational context, appearing technically accurate but experientially hollow, the mechanism behind adoption plateaus that resist conventional product interventions.

The controlled study testing this framework compared two AI agent interaction conditions: an active collaborative strategy in which the human and AI agent negotiated decisions through dialogue and reflective exchange, and a passive strategy in which each acted in alternating turns without joint deliberation. Participants in the active condition showed significantly higher levels of expectation regulation, and perceived closeness with the AI agent reached a mean score of 5.11 compared to 3.83 in the passive condition (p = 0.0045). In the qualitative record, participants in the active condition described shared ownership of outcomes while participants in the passive condition described the AI as “individual and independent”, the language pattern that precedes disengagement in commercial AI product data (García Corretjer et al., 2022).

The organizational implication is architectural. AI systems designed for active collaboration produce significantly higher user closeness, stronger trust indicators, and better expectation regulation than systems designed for passive delivery. Those figures map directly onto the engagement depth, interaction quality, and long-term retention that determine whether AI product investment generates compounding return. Organizations whose systems are built on a passive model are systematically generating the conditions for the abandonment curve documented at the opening of this article.

The behavioral lifecycle: from engagement to withdrawal

Early engagement is characterized by active exploration: broad task scope, high investment of personal context, and willingness to test the system in domains that matter. Users at this stage are extending trust provisionally, gathering evidence about the system’s reliability, consistency, and genuine responsiveness. This early period is where expectation regulation is most active: users engage in curiosity-driven testing, progressively calibrating what they can reasonably expect. When expectation regulation is supported by appropriate transparency, trust begins to accumulate. When the system forecloses that process through overconfident outputs or opaque behavior, expectations crystallize prematurely, producing a brittleness that the first significant failure will expose.

Algorithm aversion, the documented tendency for users to lose confidence in AI systems after observing a failure, even when those systems remain more accurate than human judgment, has been replicated consistently enough to be treated as a design constraint (Dietvorst et al., 2015). Users evaluate AI systems not primarily on aggregate accuracy but on the quality of the relational exchange. Where attachment has formed through genuine active collaboration, this revaluation is more forgiving. Where only instrumental use has developed, it is decisive.

Kahneman’s dual-process account provides the cognitive frame: System 1 processes adapt to AI assistance rapidly and can become dependent in ways that precede any conscious decision to rely on the system (Kahneman, 2011). The experience of reduced cognitive load feels like efficiency. The long-term effect on deliberative capacity is invisible until the system’s absence registers as a loss of one’s own capability, precisely the extended self-disruption whose commercial cost was described in the preceding section.

Agency as a design variable: the condition that makes care genuine

The position a person occupies within an AI system, whether they function as decision-maker, collaborator, approver, or passive recipient, is a design determination made through the cumulative effect of defaults, interface language, response framing, and the presence or absence of mechanisms through which the person can exercise meaningful override. Current governance frameworks have begun to formalize this: levels of human control over AI systems are now considered a designable and auditable dimension of AI deployment (Knight First Amendment Institute, 2025).

Care-based trust creates the conditions in which trust can persist. But it does not make trust sustainable on its own, and this is the critical distinction that this article introduces across the compendium. Care that preserves human agency produces relational trust that compounds over time. Care that simulates genuine orientation toward the user’s interests while systematically foreclosing their participatory contribution produces a specific and commercially dangerous pattern: the behavioral uncanny valley of AI relationship design.

The uncanny valley concept, originally introduced to describe the discomfort produced by near-human robots that approach but do not achieve human likeness (Mori, 1970/2012), has been extended to the perception of mind and agency in artificial systems (Gray & Wegner, 2012). In behavioral and relational AI, the same structural mechanism operates: a system close enough to genuine relationship to generate attachment, but absent the reciprocal human agency that makes the relationship real. Users form emotional investment in something that does not exist as they experience it. When the gap becomes apparent, through system failure, behavioral change, or the user’s growing awareness, the response is the specific disorientation of discovering that something genuinely incorporated was hollow. This is not a mild trust violation. It produces the structural, irrecoverable distrust that precedes the most commercially damaging form of AI abandonment: the kind that generates negative advocacy rather than silent churn.

The design response is to treat care and agency as inseparable requirements rather than competing values, or worse, as values to be traded against each other in pursuit of engagement optimization. Care creates the conditions in which trust can persist; human agency in the interaction, the preserved space for the user’s own judgment, deliberation, and participatory contribution, is what makes that trust sustainable over time. This is not an ethical constraint on product design. It is the condition for the only form of product loyalty that is commercially durable (Buijsman et al., 2025; Calvo et al., 2020). Shneiderman’s HCAI framework makes the structural claim explicit: every interface, every default, every moment of system intervention or restraint is a decision about where agency resides (Shneiderman, 2022). Organizations that do not make this decision explicitly have made it implicitly, and the consequences appear in product data as the behavioral patterns this article has described throughout.

Design as performative language

Understanding why these interaction architecture decisions carry disproportionate commercial weight requires a frame beyond usability. Speech act theory establishes the necessary premise: language does not merely describe states of affairs; it performs actions within them (Austin, 1962; Searle, 1969). Every element of a designed interface is a communicative act with behavioral consequences that extend beyond its surface content. A confirmation dialog does not only inform; it assigns responsibility. A recommendation phrased as a conclusion does not merely suggest; it constitutes authority. An onboarding flow does not just orient; it establishes the norms through which all subsequent interactions will be interpreted, including whether the user experiences themselves as an active participant or a passive recipient.

These messages are carried through the full multimodal register of the system: visual language, the rhythm and length of responses, vocabulary, tolerance for user ambiguity, and treatment of the system’s own limitations (Kress, 2010). When design does not govern these channels with intention, they default to whatever engineering and product optimization produces. The behavioral consequence is the loss of trust pattern this article has described: users who begin verifying rather than trusting, narrowing rather than deepening, a pattern invisible in session data but decisive in 12-month retention curves and in the quality of interaction data available for model improvement.

Paul Dourish extends this into the embodied register: meaning in designed systems is constituted through physical and procedural engagement, not only symbolic representation (Dourish, 2001). As AI moves beyond screen-based interfaces into voice, ambient, and spatial environments, the channels through which agency is communicated multiply and the organizational requirement to govern them intentionally intensifies.

The design response: building the conditions that sustain

The interventions that reduce abandonment risk are structural before they are cosmetic, and they begin with the organizational commitment to treat relationship quality as a product metric alongside the technical ones, and to measure both care and agency as dimensions of that quality, not one at the expense of the other.

For product design, Self-Determination Theory and the RE framework together translate into operational constraints. Autonomy requires designing systems that structurally preserve the user’s deliberative space: explicit communication of uncertainty, routes through which users can challenge and override outputs, and interaction rhythms that leave room for the person’s own reasoning to participate rather than simply receive. Competence requires designing for the development of genuine mastery over time, systems legible enough to learn from, that explain their reasoning in ways users can evaluate. Relatedness and attachment require designing for responsiveness that feels specific to this person: interaction architectures that adapt to demonstrated preferences and acknowledge what the user brings as genuinely shaping what they receive.

Expectation regulation requires specific design attention because it is the mechanism through which provisional trust converts into durable trust. Providing the transparency and consistency that allows calibration to proceed, the space for curiosity-driven testing, and honest acknowledgment when expectations have not been met are the interaction design decisions on which the motivational architecture of long-term engagement depends. Organizations that instrument to detect when expectation calibration is failing intervene when recovery is still possible. Those that do not encounter the consequences in churn data.

The sustainability argument is the final and most commercially significant layer of this article’s case. Organizations that design for care and agency simultaneously are building both a more ethical version of the product and a fundamentally different commercial relationship: one in which users perceive genuine value through their own active participation, rather than value delivered to them despite their passivity. Research on genuine versus coerced customer loyalty documents the financial consequences of that difference. Loyalty built on real value shows up in Net Promoter Score and in long-term retention, and it recovers after a failure. Loyalty held in place by exit friction shows up the same way until the moment a credible alternative appears, and then it does not come back (Reichheld, 2003; Edelman, 2024).

Sustainable AI product businesses will be defined by this understanding; the recommendation is to act when things are still being designed.

By Dr. Marialejandra García Corretjer, PhD

References

  • Austin, J. L. (1962). How to Do Things with Words. Oxford University Press.
  • Belk, R. W. (1988). Possessions and the extended self. Journal of Consumer Research, 15(2), 139–168.
  • Belk, R. W. (2013). Extended self in a digital world. Journal of Consumer Research, 40(3), 477–500.
  • Bickmore, T., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327.
  • Bordin, E. S. (1979). The generalizability of the psychoanalytic concept of the working alliance. Psychotherapy: Theory, Research and Practice, 16(3), 252–260.
  • Bowlby, J. (1982). Attachment and Loss, Vol. 1: Attachment (2nd ed.). Basic Books.
  • Buijsman, S., Carter, S. E., & Bermúdez, J.-P. (2025). Autonomy by design: Preserving human autonomy in AI decision-support. Philosophy & Technology, 38, 97. https://doi.org/10.1007/s13347-025-00932-2
  • Calvo, R. A., Peters, D., Vold, K., & Ryan, R. M. (2020). Supporting human autonomy in AI systems. In C. Burr & L. Floridi (Eds.), Ethics of digital well-being. Springer.
  • Deci, E. L., & Ryan, R. M. (2000). The “what” and “why” of goal pursuits. Psychological Inquiry, 11(4), 227–268.
  • Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126.
  • Dourish, P. (2001). Where the Action Is: The Foundations of Embodied Interaction. MIT Press.
  • Edelman. (2024). Edelman Trust Barometer. Edelman.
  • García Corretjer, M. A. (2022). On the relationship between people, objects, & interactive technologies [Doctoral dissertation]. La Salle, Ramon Llull University.
  • García Corretjer, M. A., Ros, R., Mallol, R., & Miralles, D. (2022). Empathy as an engaging strategy in social robotics: A pilot study. User Modeling and User-Adapted Interaction, 33(2), 221–259. https://doi.org/10.1007/s11257-022-09322-1
  • Gray, K., & Wegner, D. M. (2012). Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition, 125(1), 125–130.
  • Hsueh, S., Ciolfi Felice, M., Fdili Alaoui, S., & Mackay, W. E. (2024). What counts as “creative” work? Articulating four epistemic positions in creativity-oriented HCI research. Proceedings of CHI 2024. https://doi.org/10.1145/3613904.3642854
  • Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
  • Knight First Amendment Institute. (2025). Levels of autonomy for AI agents (K. Feng, D. W. McDonald, & A. X. Zhang, Essay 25-15).
  • Kress, G. (2010). Multimodality: A Social Semiotic Approach to Contemporary Communication. Routledge.
  • Lazar, A., Koehler, C., Tanenbaum, J., & Nguyen, D. H. (2015). Why we use and abandon smart devices. Proceedings of UbiComp 2015, 635–646.
  • Lewicki, R. J., Tomlinson, E. C., & Gillespie, N. (2006). Models of interpersonal trust development. Journal of Management, 32(6), 991–1022.
  • Litman, J. A. (2008). Interest and deprivation factors of epistemic curiosity. Personality and Individual Differences, 44(7), 1585–1595.
  • Mehrotra, S., Degachi, C., Vereschak, O., Jonker, C. M., & Tielman, M. L. (2024). A Systematic Review on Fostering Appropriate Trust in Human-AI Interaction. ACM J. Responsib. Comput., 1(4), Article 26.
  • Mori, M. (1970/2012). The uncanny valley (K. F. MacDorman & N. Kageki, Trans.). IEEE Robotics & Automation Magazine, 19(2), 98–100.
  • Picard, R. W. (1997). Affective Computing. MIT Press.
  • Reichheld, F. F. (2003). The one number you need to grow. Harvard Business Review, 81(12), 46–54.
  • Ryan, R. M., & Deci, E. L. (2017). Self-Determination Theory. Guilford Press.
  • Searle, J. R. (1969). Speech Acts. Cambridge University Press.
  • Shih, P. C., Han, K., Poole, E. S., Rosson, M. B., & Carroll, J. M. (2015). Use and adoption challenges of wearable activity trackers. iConference 2015 Proceedings.
  • Shneiderman, B. (2022). Human-Centered AI. Oxford University Press.
  • Wampold, B. E., & Imel, Z. E. (2015). The Great Psychotherapy Debate (2nd ed.). Routledge.
LET'S TALK

From reflection to practice

How do these questions connect to your business right now? Talk to Stefanini about strategy, technology and customer experience.

Talk to Stefanini