Abstract
AI-generated advertising raises a disclosure problem that differs from conventional sponsorship disclosure because the relevant concealment concerns both authorship and persuasive intent. This manuscript proposes a field-test design for studying whether disclosure of AI generation changes ad credibility, consumer understanding, and downstream behavioral responses under realistic platform conditions. The evidence base is intentionally targeted rather than exhaustive: it draws on experiments in pharmaceutical social-media advertising, price disclosure in pharmaceutical television advertising, and alcohol warning and counter-advertising studies, together with transferable methodological lessons from a systematic review of emoji research and a social-media advertising study that modeled content characteristics, psychological mediators, and platform context. The proposal specifies a factorial design with experimentally varied disclosure type, placement, and salience; an algorithmic decision sequence for stimulus selection and exposure logging; and a measurement set that separates immediate comprehension from trust, message processing, and sharing or purchase intention. A central operational rule is that disclosure exposure should be coded only when the disclosure is visible before the first evaluative response, and salience should be treated as a manipulable display characteristic rather than a binary presence variable. The protocol also defines validation checks for manipulation fidelity, attrition bias, and heterogeneity across platform type and ad category. The resulting design is intended to support falsifiable tests of when disclosure informs audiences, when it reduces persuasion, and when it is ignored or counteracted by other ad features. It remains a conceptual proposal; no empirical results are claimed.
1. Introduction and research question
The policy problem addressed here is not whether AI can generate advertising copy, images, or video, but how disclosure of AI generation should be tested when the ad appears in ordinary platform environments. Existing disclosure debates in digital marketing often concern source affiliation, hidden sponsorship, price transparency, or warning content. Those studies are useful because they show that disclosure and counter-messaging can alter perceived source motives, trust, and downstream intentions, while also showing that effects depend on media form and outcome choice. In a social-media pharmaceutical experiment, affiliation cues and comment-control cues changed perceived source connection and reduced trust, recommendation likelihood, and sharing intentions. In a pharmaceutical television experiment, some price-disclosure formats increased perceived informational sufficiency in particular conditions but did not change treatment-seeking or comparison-shopping intentions. In alcohol studies, warning or counter-advertising changed immediate affective or attitudinal responses and, in one field experiment, also shifted sponsor-brand evaluations and purchase intentions. These findings motivate a focused question for AI-generated advertising: under what disclosure conditions does identifying AI authorship change consumer response, and through which measurable mechanisms? [1] [2] [3] [4]
The research question is therefore narrowly framed: in realistic exposure settings, does disclosure of AI generation change comprehension of message provenance, perceived credibility, perceived manipulation, and behavioral intentions relative to otherwise identical ads without disclosure? A secondary question concerns moderation by platform context and ad domain. A third question concerns method: which outcome set and measurement sequence are sensitive enough to detect disclosure effects without conflating disclosure with unrelated design features such as ad quality, message tone, or platform norms? The protocol below proposes a field-test architecture rather than an empirical result, because the main contribution is methodological specification for future evaluation.
2. Evidence scope and methodological foundations
The evidence base used for this protocol is a targeted narrative synthesis of studies directly relevant to disclosure, persuasion, and platform-mediated advertising. It is not a systematic review and does not claim exhaustive coverage. The cited social-media pharmaceutical experiment is relevant because source-affiliation cues and comment-control cues changed trust and sharing through altered perceptions of source relations. The price-disclosure experiment is relevant because it tested whether policy-mandated transparency improved understanding or consumer behavior and found that informational gains did not automatically translate into action. The alcohol warning and counter-advertising studies are relevant because they used randomized exposure and measured both proximal responses and downstream intentions or brand-related judgments. On that basis, this manuscript adopts a general design principle: disclosure should be evaluated with both immediate cognitive outcomes and downstream behavioral proxies, not with attitudinal measures alone. [1] [2] [3] [4]
Two additional methodological considerations guide the design. First, the systematic review of emoji research describes a fragmented literature with diverse theories, varied methods, emphasis on short-term outcomes, and limited use of field experiments; this supports a platform-aware design rather than one focused only on message text or imagery. Second, the social-media advertising study provides an example of a model that links perceived content characteristics, psychological mediators, and platform environment, while also stating that its implications for content design are inferential because the study did not experimentally manipulate ad production. This manuscript uses that structure only as an organizing template for a future disclosure trial. Accordingly, disclosure is treated here as one experimentally varied content feature, and platform context is specified as a boundary condition to be modeled rather than ignored. [5] [6]
3. Proposed design and experimental conditions
The proposed study is a randomized field test with a factorial structure. The core design is 2 × 3 × 2: disclosure presence (AI disclosure vs no disclosure), disclosure salience (low-salience footer vs medium-salience label near the ad frame vs high-salience pre-roll or opening-screen disclosure), and platform context (short-form feed platform vs search/social commerce interface). The design can be expanded by ad domain, but the protocol requires only one domain per deployment wave to preserve interpretability. A field test means that participants encounter stimuli in an environment that approximates ordinary platform use, whether through a recruited online panel, browser-based simulation, or in-platform ad exposure with consented instrumentation. The ad creative itself must remain constant across all cells except for the disclosure treatment.
Operationally, disclosure presence refers to a textual or visual statement that the ad content was generated wholly or partly by an AI system. Disclosure salience refers to the perceptual prominence of that statement. The protocol does not equate salience with typography alone; instead, it defines salience as the combination of spatial position, timing, contrast, and pre-response visibility. The stimulus should be prepared so that the underlying message, offer, and visuals remain identical across conditions. This prevents the disclosure manipulation from becoming a proxy for creative quality. A practical comparator is a human-authored version with no AI disclosure and matched layout, which allows the study to distinguish effects of authorship disclosure from effects of disclosure as such.
The target sample should be recruited from adults with ordinary exposure to digital advertising in the chosen platform environment. The protocol does not require a probabilistic sample, but it requires quota balancing for age, gender, and device type if these factors shape platform use. Because exposure conditions differ between simulated and native delivery, the analysis plan must separate assignment, verified visibility, and substantive response. The field-test unit is the individual exposure event, but the estimand is the average causal effect of disclosure assignment on outcomes within a defined platform-ad domain combination. For native platform delivery, the preregistered default rule is to include a case in the primary analysis only when server-side impression timestamps, viewport-entry logs, and client-side response timestamps jointly indicate that the disclosure entered the visible screen area before the first evaluative item was rendered; otherwise the case remains in the intention-to-treat analysis but is coded as unverified visibility for compliance-sensitive analyses. If a platform does not permit these synchronized logs, the deployment should be classified as partial-verification and this limitation should be reported explicitly.
4. Variables and exact operational definitions
The primary independent variable is disclosure condition. The primary dependent variables are: (1) comprehension of message provenance, measured as correct identification of whether the ad was AI-generated; (2) perceived credibility, measured by a multi-item scale; (3) perceived manipulation or deceptiveness; and (4) behavioral proxy outcomes such as click intention, sharing intention, purchase intention, or willingness to seek more information, depending on the ad domain. Secondary outcomes should include ad recall, ad annoyance, and policy support for mandatory disclosure. Where the ad concerns health, a domain-specific outcome such as treatment-seeking intention may be appropriate, but the protocol does not assume that every domain requires identical endpoints.
At minimum, the mediation model should include perceived transparency and perceived source authenticity as candidate mediators. If the ad domain is health-related or socially sensitive, perceived risk should also be included. These mediators are chosen because the disclosure question concerns whether audiences interpret the message as properly attributable and candid about its production. A concrete operational rule is required. Exposure is coded as verified only when the disclosure was visible before the first evaluative response according to preregistered log data. In a browser-based simulation, visibility time should be defined as the first timestamp at which the disclosure is rendered within the active viewport for at least a fixed minimum duration, with 1000 ms as the default threshold and 500 ms and 1500 ms reserved for sensitivity analyses. In native platform delivery, the same rule should be approximated by synchronized impression, viewport, and survey-render timestamps; if that approximation is unavailable, visibility cannot be treated as verified. Formally, valid_exposure = 1 if d_visible_before_response >= d_min, and valid_exposure = 0 otherwise, where d_min is the preregistered minimum visible duration. Cases with valid_exposure = 0 should be retained for intention-to-treat estimation and analyzed separately in a compliance-sensitive model. This rule is conservative by design and preserves the distinction between assignment and actual opportunity to process the disclosure.
An illustrative mathematical definition of disclosure salience may be used for design auditing, not for causal inference. Let S = w1P* + w2T* + w3C* + w4V*, where each component is normalized to the unit interval before aggregation and the weights are fixed in advance and sum to 1. P* denotes positional prominence on a 0-1 scale defined relative to a preregistered screen map, with values closer to 1 for locations nearer the principal focal region of the ad. T* denotes temporal availability before the first evaluative response and is computed as min(d_visible_before_response, cap)/cap, where d_visible_before_response is the accumulated duration of intervals during which the disclosure is visible before the first evaluative response, where cap is a preregistered upper bound. C* denotes visual contrast rescaled to 0-1 using a preregistered minimum and maximum contrast range applied uniformly across stimuli. V* denotes the proportion of the disclosure area visible in the default device viewport, bounded between 0 and 1. The primary rule is equal weighting, w1 = w2 = w3 = w4 = 0.25. A sensitivity analysis may vary the weights within a preregistered range while preserving their sum, but the substantive analysis should not redefine salience after outcomes are known. This normalization rule makes S comparable across stimuli and prevents scale artifacts from dominating the audit score.
5. Decision sequence for field implementation
The protocol uses a fixed decision sequence. Step 1: select one ad domain and one platform context. Step 2: create a single base ad that can plausibly be encountered in that context. Step 3: generate disclosure variants that differ only in the assigned disclosure treatment. Step 4: pilot the stimuli for recognition, readability, and equivalence of non-disclosure features. Step 5: randomize eligible participants to condition. Step 6: record exposure order, dwell time, scrolling behavior, and response latency. Step 7: administer outcome measures in a pre-specified sequence, beginning with comprehension and ending with policy attitudes. Step 8: estimate treatment effects with models that adjust only for pre-treatment covariates and exposure fidelity indicators.
The sequence matters because disclosure can be missed, misread, or processed only after other ad features have already shaped evaluation. The social-media pharmaceutical experiment showed that source cues can alter trust through perceived affiliation. The price-disclosure experiment showed that informational content alone does not guarantee behavioral impact. The alcohol studies showed that brief exposure can affect proximal outcomes and that message content linked to sponsorship can also alter brand-related judgments and intentions. These patterns imply that a disclosure test should not wait until the end of the survey to ask whether the message was noticed. Instead, the protocol requires an early check on disclosure perception, followed by a pre-specified battery of outcomes. This ordering reduces retrospective rationalization and helps separate failed exposure from null persuasion effects. [1] [2] [3] [4]
The comparator condition should be a no-disclosure control matched for every non-disclosure attribute. If the platform ordinarily displays sponsorship or provenance tags, the control should mimic the platform's native visual grammar while withholding explicit AI attribution. If policy practice already requires a generic label, the experimental contrast should be between generic and AI-specific disclosure rather than between disclosure and no label. This boundary condition is important because the policy question may concern incremental disclosure content rather than disclosure in the abstract.
6. Outcome strategy, mediation, and heterogeneity
The outcome strategy should privilege proximal responses first and behavioral proxies second. Immediate proximal outcomes include perceived transparency, trust, and perceived manipulation. Intermediate outcomes include message sharing, seeking further information, and willingness to interact with the source account. Distal outcomes, if the design allows, include actual click-through, time on linked content, or opt-in for further communication. The methodological reason for this hierarchy is that prior studies show effects may appear in informational judgments, affect, attitudes, or intentions without immediate movement in later behaviors. This protocol therefore does not treat a null behavioral effect as evidence that disclosure is irrelevant. It instead treats such a pattern as consistent with a mismatch between the observation window and the mechanism being tested. [2] [3] [4]
Mediation analysis should be limited to pre-specified pathways. The primary path is disclosure assignment -> perceived transparency or source authenticity -> trust or sharing intention. A second path is disclosure assignment -> perceived manipulation -> avoidance or policy support. The cited social-media advertising study shows one way to combine content characteristics, psychological mediation, and perceived platform environment within a single model, while the emoji review indicates that adjacent digital-advertising literatures often disperse attention across many variables with limited replication. Accordingly, this protocol permits no more than two primary mediators in any deployment wave. Additional mediators may be exploratory, but they should not be redesignated as primary after the data are observed. This restriction is especially important in field tests, where measurement burden can itself degrade response quality. [6] [5]
Heterogeneity should be tested only where a plausible boundary condition exists. The protocol specifies platform type, ad domain, and prior familiarity with AI tools as candidate moderators. A health or finance ad may trigger more scrutiny than a low-stakes entertainment ad; a feed platform may invite shorter attention than a search interface; and users already familiar with generative systems may interpret disclosure less negatively because the label confirms an expected production method. These moderators are not asserted as universal; they are treated as hypotheses requiring direct testing. The design should not search indiscriminately across subgroups because that would inflate the risk of spurious moderation claims.
7. Validation design, failure cases, and limits of generalization
Validation has three layers. First, stimulus validation checks whether the disclosure is noticed and correctly interpreted. Second, construct validation checks whether the trust, transparency, and manipulation items load coherently and distinguish disclosure-relevant reactions from general ad dislike. Third, field validity checks whether the same treatment pattern appears across at least two platform-ad combinations or at two time points with different user cohorts. The last layer is essential because a disclosure effect that appears only in one artificial interface may be too context-bound to inform policy. The protocol does not assume cross-platform equivalence; it tests it.
Several failure cases are foreseeable. If the disclosure is too subtle, treatment compliance fails and the study estimates weak visibility rather than disclosure effects. If the disclosure is too forceful, the treatment may alter ad aesthetics so strongly that the manipulation no longer represents ordinary practice. If the ad is highly familiar or highly polarizing, pre-existing beliefs may dominate disclosure processing. These cases are not reasons to abandon the design; they are reasons to document stimulus properties and to report negative or attenuated findings precisely. The price-disclosure experiment illustrates the importance of distinguishing informational change from behavioral change, and the social-media pharmaceutical experiment shows that some source-related cues can shift trust and recommendation or sharing outcomes through perceived affiliation or control. On that basis, null or mixed results are analytically meaningful rather than merely inconclusive. [2] [1]
Generalization is limited in four ways. First, findings from one platform grammar may not transfer to another, especially when user expectations about sponsorship or automation differ. Second, disclosure effects may vary by product category because health, finance, and entertainment ads invite different standards of credibility. Third, a short field test captures immediate responses better than delayed behavior. Fourth, a study based on a single creative unit cannot settle whether all AI-generated ads should be disclosed in the same way. The manuscript therefore proposes a transferable methodology, not a universal disclosure rule. Its contribution is bounded: it defines how to test disclosure in a way that separates manipulation salience, content equivalence, and platform context.
Clock semantics and confirmatory contrast
Visibility intervals are measured on the same monotonic clock as the first response and clipped to the response time. Their union, not their sum with double-counted overlaps, defines visible duration. Negative or conflicting timestamps make visibility unverified. The cap and duration use the same registered unit, ordinarily milliseconds. Mere rendering before response does not establish continuous visibility. All assigned participants remain in the intention-to-treat analysis; verified-viewer comparisons are secondary and may be selected by attention or device behavior. In the proposed factorial design, the primary contrast is disclosure assignment averaged over the randomized salience and format factors, on provenance comprehension. The other endpoints and interactions belong to prespecified secondary families with Holm adjustment. Mediation analyses require additional assumptions and are exploratory; a salience audit score does not itself identify a causal mediation pathway.
8. Conclusion
This protocol establishes a conceptual and methodological basis for field-testing disclosure of AI-generated advertising. The proposed design specifies what should be manipulated, which outcomes should be prioritized, how exposure should be operationalized, and how mediators and moderators should be bounded to avoid overextended inference. Prior experimental work on pharmaceutical affiliation disclosure, pharmaceutical price disclosure, alcohol warning advertising, and alcohol counter-advertising indicates that transparency cues and related message interventions can influence trust, perceived information, affect, and some intentions or brand evaluations, but that effects are conditional on context and outcome selection. Additional methodological guidance comes from a systematic review that identified fragmentation and limited field experimentation in adjacent digital-advertising research and from an observational social-media advertising study that modeled content characteristics, psychological mediation, and perceived platform environment while treating its design implications as inferential. What is established here is a testable framework for disclosure experiments. What remains untested is whether AI-authorship disclosure produces consistent effects across domains, platforms, and audience segments, and whether any observed effects are strong enough to support a particular regulatory format. The next step is empirical implementation under the decision rules specified above. [1] [2] [3] [4] [5] [6]
Contribution and availability
This manuscript proposes a bounded, field-testable protocol for studying disclosure of AI-generated advertising. It defines experimentally manipulable disclosure dimensions, a compliant exposure rule, a salience metric, prioritized outcomes, a mediation structure, validation layers, and boundary conditions. The design is explicitly transferable across ad domains while remaining strict about content equivalence and platform context.
Funding: Roman Vlasov’s own funds. Editorial responsibility and competing interests: Roman Vlasov is the founder and Editor-in-Chief of Smart World Journal and the creator of the BILDRY construction-planning platform. Other competing interests have not been declared.
This conceptual study reports no newly collected empirical data. Its proposed procedures require the prospective validation described above.
References
- DeAndrea, David Christopher; Vendemia, Megan Ashley (2016). How Affiliation Disclosure and Control Over User-Generated Comments Affects Consumer Health Knowledge and Behavior: A Randomized Controlled Experiment of Pharmaceutical Direct-to-Consumer Advertising on Social Media. Journal of Medical Internet Research. 18(7), e189. doi:10.2196/jmir.5972.
- Hollin, Ilene L.; Ball, Jennifer Gerard (2022). Does price disclosure in pharmaceutical advertising result in price transparency? Evidence from a randomized experiment. Exploratory Research in Clinical and Social Pharmacy. 8, 100180. doi:10.1016/j.rcsop.2022.100180.
- Stautz, Kaidy; Marteau, Theresa M (2016). Viewing alcohol warning advertising reduces urges to drink in young adults: an online experiment. BMC Public Health. 16(1), 530. doi:10.1186/s12889-016-3192-9.
- Dixon, Helen; Scully, Maree; Niederdeppe, Jeff; Brennan, Emily; O'Brien, Kerry; Vandenberg, Brian; Pettigrew, Simone; Wakefield, Melanie (2023). Can counter‐advertising dilute marketing effects of alcohol sponsorship of elite sport: A field experiment. Addiction. 118(12), 2360-2373. doi:10.1111/add.16317.
- Vardikou, Chrysopigi; Konidaris, Agisilaos; Koustoumpardi, Erato; Kavoura, Androniki (2025). Emojis in Marketing and Advertising: A Systematic Literature Review. Behavioral Sciences. 15(11), 1490. doi:10.3390/bs15111490.
- Han, Yu (2026). The relationship between the social media advertising content characteristics and the advertising effectiveness of female users: a moderated mediation model based on basic psychological needs. Frontiers in Psychology. 17, 1831184. doi:10.3389/fpsyg.2026.1831184.