Submitted:
23 October 2026
Posted:
27 October 2026
You are already at the latest version
Abstract
Artificial intelligence (AI) now selects, personalises, writes and voices marketing communications, placing an algorithmic agent at the interface between firms and consumers. Yet the evidence on how consumers respond is fragmented across three largely separate literatures — studies of AI disclosure, of personalisation and privacy, and of algorithm appreciation and aversion — which imply conflicting managerial defaults. This paper integrates these literatures through a single interpretive mechanism drawn from the Persuasion Knowledge Model: persuasion-knowledge activation. We argue that interface design choices — whether AI involvement is disclosed, whether data collection is overt or covert, how humanlike the agent appears, and how the task is framed — do not act on trust directly. They act by changing what consumers infer about the persuader's goals and competence, an inference process in which privacy calculus, uniqueness neglect and identity threat also operate. The resulting Trust at the Interface framework links interface signals to consumer appraisals, to trust attributions directed separately at the AI system and at the firm, and to behaviour, and adds a failure-and-repair loop explaining asymmetric blame and recovery after AI errors. Seven formal propositions specify the framework's testable claims, with implications for interface design, consumer protection and future research.
Keywords:
artificial intelligence
; consumer trust
; persuasion knowledge
; disclosure
; personalisation
; privacy calculus
; algorithm aversion
; trust repair
1. Introduction
Marketing’s point of contact with the consumer is increasingly a machine. Recommender systems decide which products a consumer sees; targeting algorithms decide who sees which advertisement and when; generative systems draft the copy; chatbots conduct the conversation itself. Puntoni et al. (2021) map this shift as four consumer experiences with AI — data capture, classification, delegation and social experience — each asking something of the consumer: data, attention, reliance, or social engagement. What the consumer gives in return depends less on the technology’s capability than on whether the consumer trusts it. Trust, not capability, is the binding constraint on AI-mediated marketing.
The evidence about that trust, however, arrives in fragments that point in different directions. Disclosure experiments suggest that telling consumers a chatbot is a machine can devastate purchase behaviour (Luo et al., 2019) — yet the personalisation literature shows that concealed data practices destroy trust when discovered (Aguirre et al., 2015). The appraisal literature is split against itself: consumers often prefer algorithmic to human judgement (Logg et al., 2019) yet abandon algorithms after a single visible error (Dietvorst et al., 2015). And when AI causes harm, consumers can be more forgiving of an algorithm than of a human employee (Srinivasan & Sarial-Abi, 2021) — awkwardly beside the anger triggered by an overly humanlike chatbot (Crolic et al., 2022). Each literature studies one design choice in isolation and treats trust as that choice’s direct output.
This paper’s claim is that the fragments cohere once we attend to the consumer’s own interpretive machinery. The Persuasion Knowledge Model (Friestad & Wright, 1994) holds that consumers accumulate knowledge about persuaders’ goals and tactics and deploy it to cope with persuasion attempts. We argue that persuasion-knowledge activation is the integrating mechanism of AI-mediated marketing: interface signals are first interpreted — as honesty or as a confession of manipulation, as competence or as intrusion, as a tool’s malfunction or as a firm’s betrayal — and trust follows the interpretation, not the signal. The paper makes three contributions. First, it integrates the disclosure, personalisation–privacy and algorithm-appraisal literatures into one framework. Second, it separates two trust targets the literatures routinely conflate — trust in the AI system and trust in the firm behind it — and shows why design choices can move them in opposite directions. Third, it extends the analysis across time, adding a failure-and-repair loop explaining why trust in AI interfaces is characteristically asymmetric: quickly granted, sharply withdrawn, and repaired by different means than interpersonal trust.
2. Approach and Theoretical Anchors
This is an integrative, conceptual review. It is not a systematic review: no search protocol, screening flow or study-count claim is made. The evidence base is a curated set of peer-reviewed studies bearing directly on how consumers appraise AI at the marketing interface — experimental and field evidence on disclosure, personalisation and chatbot interaction, and the experimental literature on algorithm appreciation and aversion — read against three theoretical anchors: the Persuasion Knowledge Model, research on trust in AI, and the privacy calculus. The synthesis is interpretive: it identifies the mechanism connecting findings across these bodies of work and derives propositions at the level at which managers design interfaces.
The first anchor is the Persuasion Knowledge Model (Friestad & Wright, 1994). Consumers hold beliefs about persuasion agents — their motives, tactics and appropriateness — and activate this knowledge when they recognise a persuasion attempt, using it to cope: to resist, discount or, sometimes, appreciate the attempt; the resulting evaluations cover both the offer and the marketer. A meta-analysis of 148 papers and 171 data sets confirms that persuasion knowledge reliably increases coping responses and less favourable evaluations of marketer actions, while cautioning that its explanatory power reaches only about half that of persuasion itself and depends on characteristics of the persuasion process (Eisend & Tarrahi, 2022). Persuasion knowledge is thus a real but partial, design-sensitive brake, not a switch.
The second anchor is trust in technology. Glikson and Woolley (2020), reviewing empirical work on human trust in AI, show that trust in AI is built and lost through different cues than interpersonal trust, with transparency and perceived capability as recurring levers. A consumer following a recommendation is appraising a system’s capability, not a person’s character; yet she also appraises the firm that deployed the system, and the two judgements need not coincide.
The third anchor is the privacy calculus. Dinev and Hart (2006) model willingness to provide personal information in e-commerce as a risk–benefit assessment in which privacy concerns and trust beliefs are central. Every personalised interface implicitly asks the consumer to redo this calculus: what does the system know about me, and what do I get in return? Together, the three anchors supply the appraisal machinery through which, we argue, all interface signals must pass.
3. Three Literatures
3.1. Disclosure: Telling the Consumer It Is a Machine
The strongest causal evidence on disclosure comes from Luo et al. (2019), a field experiment with more than 6,200 customers of a firm using AI chatbots in sales conversations. Undisclosed chatbots performed as well as proficient human sales agents. Disclosing the chatbot’s identity before the conversation, however, reduced purchase rates by more than 79.7%, an effect traced to disclosed bots being perceived as less knowledgeable and less empathetic. Two mitigations matter: disclosing after value had been demonstrated (late disclosure) and customers’ prior AI experience both softened the penalty. Disclosure was not evaluated as an abstract virtue; it was interpreted through beliefs about what a machine can do.
Subsequent work shows the interpretation is role-contingent. Wu et al. (2025), across three experiments, distinguish what the AI did: consumers were more willing to share an advertisement placed by AI than one created by AI, when they believed AI capable of high-complexity tasks — mediated by perceived task objectivity feeding a “machine heuristic”, a rule-of-thumb crediting machines with mechanical competence. Disclosure is thus read against the task. It can also be read against the consumer’s identity: Xu et al. (2024) find that disclosing AI integration into service delivery can itself backfire, through fear of AI, identity threat and existential threat. The composite lesson is that transparency is a signal whose meaning the receiver supplies: the same disclosure can register as honesty, as a competence warning, or as an existential one.
3.2. Personalisation and Privacy: The Covertness Problem
The personalisation literature locates the trust problem not in personalisation itself but in how the underlying data were obtained. Aguirre et al. (2015), combining field and experimental evidence, show that personalised advertising collapses when consumers realise their data were collected covertly: felt vulnerability drives the drop. Overt collection reverses the effect, and trust-building strategies offset it. Boerman et al. (2017), reviewing online behavioural advertising, similarly find that knowledge and disclosure shape consumers’ responses, with persuasion knowledge and privacy concern as the recurring explanatory constructs — the pre-AI baseline that algorithmic persuasion inherits and intensifies.
Prior trust acts as a licence. Bleier and Eisenbeiss (2015) show that retargeting effectiveness hinges on prior trust in the retailer: trusted retailers gain from deep personalisation without triggering reactance or privacy concerns, while less-trusted retailers trigger both. Personalisation depth is not a dial with a fixed trust price; it is spent against a balance of accumulated trust.
Inside the interface, the privacy calculus keeps running. Kim et al. (2023), in two experiments on highly personalised chatbot advertising, find that promotion-focused consumers respond favourably — attending to benefits — while prevention-focused consumers respond unfavourably, attending to risks; risk–benefit perceptions mediate both paths. Critically, high privacy concern produced dislike regardless of regulatory focus: above a threshold of concern, benefit framing stops working. The calculus is a gatekeeper, not a smooth trade-off.
3.3. Algorithm Appraisal: Appreciation, Aversion and the Subjective Domain
The appraisal literature explains why baseline trust in AI advice is often surprisingly high — and why it is fragile. Logg et al. (2019), across six experiments, find that lay people frequently adhere more to advice believed to come from an algorithm than to identical advice from a person, though appreciation wanes against the consumer’s own judgement and with domain expertise. Appreciation supplies the high starting point. Dietvorst et al. (2015) supply the cliff edge: after seeing an algorithm err, people lose confidence in it faster than in a human forecaster who made the same mistake, and avoid it even when it outperformed the human.
Whether consumers resist the algorithm in the first place depends on the task and on the self. Castelo et al. (2019) show aversion is task-dependent: consumers resist algorithms for subjective tasks, and framing a task as benefiting from quantitative analysis increases trust in and reliance on algorithmic advice. Longoni and Cian (2022) document the corresponding “word-of-machine” effect: consumers believe AI recommenders are more competent than humans on utilitarian attributes and less competent on hedonic ones, so attribute salience decides preference — and the effect is eliminated in human–AI hybrid decision making. Longoni et al. (2019) identify a deeper brake in personal domains: “uniqueness neglect”, the belief that an AI cannot account for one’s unique characteristics and circumstances, which drives resistance to AI providers and falls when the AI is framed as personalised or as supporting rather than replacing a human. For marketing this is pointed: personalisation’s core promise — this was made for you — is exactly what uniqueness neglect denies, unless the interface credibly signals individual consideration.
Anthropomorphism cuts across appraisal. Crolic et al. (2022), in five studies including real-world telecom data, show that humanising a service chatbot — a name, an avatar — backfires for customers who arrive angry: anthropomorphism inflates pre-encounter expectations of efficacy, and the expectancy violation lowers satisfaction, firm evaluation and purchase intentions; non-angry customers show no such penalty. Humanlikeness is a promise the interface makes; whether it is kept decides whether it builds or burns trust.
4. The Trust at the Interface Framework
The framework organises the evidence into four stages and one loop (Figure 1, described textually).
Interface signals. Firms control a set of signals: whether and when AI involvement is disclosed; whether data collection is overt or covert, and how much prior trust the firm brings; how humanlike the agent appears; and how the task is framed — objective or subjective, utilitarian or hedonic, AI alone or human–AI hybrid. Puntoni et al.’s (2021) four experiences locate where these signals bite hardest: data capture and classification are the most trust-sensitive moments.
Consumer appraisals. Signals do not move trust directly; they are interpreted. The integrating appraisal is persuasion-knowledge activation: the consumer’s inference about what the persuader is doing and why. It has two characteristic outputs. A machine-heuristic reading — the system is competent, mechanical, impartial — supports trust, as in algorithm appreciation and the placement-disclosure findings (Logg et al., 2019; Wu et al., 2025). A manipulation-inference reading — the firm is watching me, steering me, hiding behind a machine — undermines it, as in covert-collection reactance (Aguirre et al., 2015). Three further appraisals run alongside: the privacy calculus, capable of vetoing favourable readings once concern crosses a threshold (Dinev & Hart, 2006; Kim et al., 2023); uniqueness neglect (Longoni et al., 2019); and identity threat (Xu et al., 2024). Which dominates is set by the signal configuration and by the consumer — privacy concern, regulatory focus, AI experience and emotional state all moderate the reading (Kim et al., 2023; Luo et al., 2019; Crolic et al., 2022).
Trust attributions — two targets. Appraisals resolve into trust judgements about the system (is it competent?) and about the firm (is it honest, does it respect me?). The two targets explain otherwise puzzling combinations: a consumer can trust a retailer’s algorithm because she trusts the retailer (Bleier & Eisenbeiss, 2015), or distrust a flawless system because its covert data practices indict the firm. Design can move the targets in opposite directions — covert personalisation may raise short-run confidence in the system’s recommendations while destroying trust in the firm when discovered.
Behaviour. Trust attributions drive purchase, advice adherence, data sharing and continued use.
The failure-and-repair loop. AI interfaces fail in public, and attribution of the failure runs through perceived agency. Srinivasan and Sarial-Abi (2021), across eight experiments, find consumers respond less negatively when a brand-harm error was caused by an algorithm than by a human, because the algorithm is perceived as having lower agency and hence lower responsibility. But the protection is conditional — it reverses when the algorithm is anthropomorphised, when it uses machine learning, or when it works on subjective, interactive tasks — and publicising technological (rather than human) supervision softens responses. Set against algorithm aversion (Dietvorst et al., 2015), the dynamics are asymmetric: errors attributed to a low-agency system are forgiven as malfunctions; errors attributed to a humanlike or firm-directed agent are prosecuted as betrayals. Repair must therefore work on the attribution — evidencing system-level correction and oversight — not merely on the apology. The loop closes back into signals: how a firm handles failure becomes part of the interface the next consumer reads.
5. Propositions
The framework yields seven testable propositions.
P1 (Role-contingent disclosure). The effect of disclosing AI involvement on consumer trust depends on the AI’s role and the task’s perceived objectivity: disclosure builds or preserves trust for placement, targeting and other objective-task roles, and erodes trust for creative and interpersonal roles, because disclosure activates persuasion knowledge whose valence is set by lay beliefs about machine competence.
P2 (The covertness penalty). Trust loss from AI personalisation is driven primarily by the perceived covertness of data collection, not by personalisation depth: overt collection and trust-building cues neutralise reactance even at high depth, whereas discovered covert collection damages trust in the firm beyond the single interaction.
P3 (Privacy calculus as gatekeeper). Perceived manipulation rises sharply once the sensitivity of the data an AI interface is seen to use exceeds the consumer’s privacy-calculus threshold; above that threshold, high privacy concern overrides benefit framing and message-level moderators such as regulatory focus.
P4 (Anthropomorphism’s expectancy debt). Anthropomorphic interface design raises pre-encounter expectations of competence and empathy; the more humanlike the agent, the larger the trust penalty when performance falters or the customer is already in a negative emotional state, relative to a plainly mechanical agent delivering the same outcome.
P5 (Asymmetric trust dynamics). Trust in AI-mediated marketing is asymmetric over time: it is granted quickly on competence cues (algorithm appreciation) and withdrawn disproportionately after a single observed error — except where the system is framed as low-agency, in which case blame after failure is buffered, but so is credit for success.
P6 (Uniqueness and identity as the subjective-domain brake). In subjective, identity-relevant and hedonic domains, resistance to AI persuasion is mediated by uniqueness neglect and identity threat rather than by competence doubt; hybrid human–AI arrangements and personalisation framing that credibly signal individual consideration attenuate this resistance.
P7 (Repair belongs to the system, accountability to the firm). After an AI-caused failure, trust is repaired more effectively by communications that evidence system-level correction — oversight, monitoring, changed behaviour of the model — than by apologies that foreground human supervision of the algorithm; and sequencing disclosure after demonstrated value (late disclosure) reduces the initial disclosure penalty without incurring the trust cost of discovered concealment.
6. Discussion
6.1. Contributions
The framework contributes in three ways. First, integration: disclosure, personalisation–privacy and appraisal findings, usually reported as separate effects with conflicting implications, become configurations of one interpretive process. The apparent contradiction between “disclosure punishes” (Luo et al., 2019) and “concealment punishes” (Aguirre et al., 2015) dissolves once both are read as persuasion-knowledge events — agent disclosure harms trust when it activates an unflattering competence belief; concealment of the data practice harms trust when its discovery activates a manipulation inference. Second, target separation: distinguishing trust in the system from trust in the firm explains cross-effects the single-target literature leaves opaque, including why prior firm trust licenses personalisation (Bleier & Eisenbeiss, 2015) and why discovered covertness damages the firm, not merely the campaign. Third, dynamics: the failure-and-repair loop connects first-encounter trust formation to the longitudinal problem firms actually face — keeping trust after the inevitable error.
6.2. Implications for Marketers
Four design rules follow. (1) Disclose by role, not by blanket policy: disclosure is cheapest — sometimes trust-positive — for objective tasks, and costliest in creative and interpersonal roles; disclosure after demonstrated value reduces the penalty (Luo et al., 2019). (2) Collect overtly, and spend prior trust deliberately: covertness, not depth, is the penalty; firms with thin trust reserves should personalise shallowly until trust is earned (Aguirre et al., 2015; Bleier & Eisenbeiss, 2015). (3) Treat anthropomorphism as a promise: humanlike design borrows against expectations and is repaid — with interest — at the first failure or the first angry customer (Crolic et al., 2022). (4) Design the failure path in advance: communicate system-level correction and technological oversight, and resist anthropomorphising the failing system (Srinivasan & Sarial-Abi, 2021).
6.3. Implications for Regulators and Consumer Protection
The framework reframes familiar regulatory instincts. Disclosure mandates assume transparency is unconditionally protective; the evidence shows disclosure’s effects are role-, timing- and consumer-contingent, so mandates are better aimed at data-practice transparency — overt collection, discoverable data use — where concealment’s trust harm is unambiguous, than at uniform agent-identity labels whose effects vary. The covertness penalty (P2) and privacy-calculus gatekeeper (P3) also imply that consent obtained without salience is functionally covert: regulators concerned with manipulation should attend to whether collection is noticed, not only whether it is disclosed somewhere. These implications are offered as interpretation of consumer evidence, not as legal analysis.
6.4. Limitations and Future Research
The review has limits. Its evidence base is curated rather than exhaustive, and it leans on experimental studies — many in North American and online contexts, several in adjacent domains such as healthcare and hospitality — whose transfer to live marketing interfaces requires care. The framework is conceptual: persuasion-knowledge activation is inferred from the pattern of findings, not measured as a mediator across them, and the propositions await direct tests. Three priorities follow. First, studies measuring persuasion-knowledge activation, privacy calculus and trust in both targets within the same design, so the framework’s mediating claims can be falsified rather than illustrated. Second, longitudinal and field work on the failure-and-repair loop. Third, boundary work: cultural variation in disclosure norms and machine heuristics, consumers’ growing AI experience (already a moderator in Luo et al., 2019), and generative interfaces whose fluency may blur the signals — humanness, authorship, targeting — on which the appraisals depend.
7. Conclusion
AI-mediated marketing does not face a trust problem so much as an interpretation problem. Consumers read every interface signal — a disclosure, a suspiciously apt recommendation, a chatbot’s name and face — through what they know and suspect about persuaders, and extend or withdraw trust according to the reading. Firms that design for the reading, not the signal in isolation, will disclose where disclosure reads as honesty, collect data where collection reads as fair exchange, keep their agents’ promises proportionate to their competence, and repair failures at the level of the system. Trust at the interface is not a feature to be added; it is the medium in which every other feature is judged.
Funding
None. This research received no external funding.
Conflicts of Interest
The author is affiliated with Sensegrass. No other conflicts of interest are declared.
AI Assistance Disclosure
The author used an AI assistant (Muse) for literature organisation and drafting support; all arguments, propositions and final text are the author’s responsibility.
Data Availability
Not applicable — this is a conceptual and integrative review paper reporting no new data.
References
- Aguirre, E.; Mahr, D.; Grewal, D.; de Ruyter, K.; Wetzels, M. Unraveling the personalization paradox: The effect of information collection and trust-building strategies on online advertisement effectiveness. Journal of Retailing 2015, 91(1), 34–49. [Google Scholar] [CrossRef]
- Bleier, A.; Eisenbeiss, M. The importance of trust for personalized online advertising. Journal of Retailing 2015, 91(3), 390–409. [Google Scholar] [CrossRef]
- Boerman, S. C.; Kruikemeier, S.; Zuiderveen Borgesius, F. J. Online behavioral advertising: A literature review and research agenda. Journal of Advertising 2017, 46(3), 363–376. [Google Scholar] [CrossRef]
- Castelo, N.; Bos, M. W.; Lehmann, D. R. Task-dependent algorithm aversion. Journal of Marketing Research 2019, 56(5), 809–825. [Google Scholar] [CrossRef]
- Crolic, C.; Thomaz, F.; Hadi, R.; Stephen, A. T. Blame the bot: Anthropomorphism and anger in customer–chatbot interactions. Journal of Marketing 2022, 86(1), 132–148. [Google Scholar] [CrossRef]
- Dietvorst, B. J.; Simmons, J. P.; Massey, C. Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General 2015, 144(1), 114–126. [Google Scholar] [CrossRef] [PubMed]
- Dinev, T.; Hart, P. An extended privacy calculus model for e-commerce transactions. Information Systems Research 2006, 17(1), 61–80. [Google Scholar] [CrossRef]
- Eisend, M.; Tarrahi, F. Persuasion knowledge in the marketplace: A meta-analysis. Journal of Consumer Psychology 2022, 32(1), 3–22. [Google Scholar] [CrossRef]
- Friestad, M.; Wright, P. The Persuasion Knowledge Model: How people cope with persuasion attempts. Journal of Consumer Research 1994, 21(1), 1–31. [Google Scholar] [CrossRef] [PubMed]
- Glikson, E.; Woolley, A. W. Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals 2020, 14(2), 627–660. [Google Scholar] [CrossRef]
- Kim, W.; Ryoo, Y.; Lee, S.; Lee, J. A. Chatbot advertising as a double-edged sword: The roles of regulatory focus and privacy concerns. Journal of Advertising 2023, 52(4), 504–522. [Google Scholar] [CrossRef]
- Logg, J. M.; Minson, J. A.; Moore, D. A. Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 2019, 151, 90–103. [Google Scholar] [CrossRef]
- Longoni, C.; Bonezzi, A.; Morewedge, C. K. Resistance to medical artificial intelligence. Journal of Consumer Research 2019, 46(4), 629–650. [Google Scholar] [CrossRef]
- Longoni, C.; Cian, L. Artificial intelligence in utilitarian vs. hedonic contexts: The “word-of-machine” effect. Journal of Marketing 2022, 86(1), 91–108. [Google Scholar] [CrossRef]
- Luo, X.; Tong, S.; Fang, Z.; Qu, Z. Frontiers: Machines vs. humans: The impact of artificial intelligence chatbot disclosure on customer purchases. Marketing Science 2019, 38(6), 937–947. [Google Scholar] [CrossRef]
- Puntoni, S.; Reczek, R. W.; Giesler, M.; Botti, S. Consumers and artificial intelligence: An experiential perspective. Journal of Marketing 2021, 85(1), 131–151. [Google Scholar] [CrossRef]
- Srinivasan, R.; Sarial-Abi, G. When algorithms fail: Consumers’ responses to brand harm crises caused by algorithm errors. Journal of Marketing 2021, 85(5), 74–91. [Google Scholar] [CrossRef]
- Wu, L.; Dodoo, N. A.; Wen, T. J. Disclosing AI’s involvement in advertising to consumers: A task-dependent perspective. Journal of Advertising 2025, 54(1), 20–38. [Google Scholar] [CrossRef]
- Xu, Y. W.; Zhou, G. M.; Cai, R. R.; Gursoy, D. When disclosing the artificial intelligence (AI) technology integration into service delivery backfires: Roles of fear of AI, identity threat and existential threat. International Journal of Hospitality Management 2024, 122, 103829. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.