How does AI-origin disclosure change advertising response?

Persuasion knowledge, credibility and the distinction between visible labels and technical provenance

This retrospective review considers source versions available by 14 September 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: AI disclosure; advertising trust; persuasion knowledge; content provenance

1. The disclosure question

A statement that an advertisement used AI changes the information available to the audience. It can tell viewers something about the production process, but its meaning depends on what they understand AI to have done. An image generated entirely from a text prompt, a translated voice-over and a minor retouch are different uses. A single undifferentiated label can leave that distinction unresolved. The effect of disclosure must therefore be specified in terms of wording, prominence, timing and the underlying asset.

This review asks how AI-origin disclosure affects advertising evaluation and what the evidence can establish about the mechanisms involved. The focus is empirical response: credibility, trust, ad attitudes and behavioural intentions. Regulatory and technical sources are included to separate those outcomes from legal disclosure categories and machine-readable provenance. The presence of a required notice does not itself predict whether an audience will trust the ad [7] [8] [9].

Two comparisons need particular separation. Producing an ad with AI can change the material. Telling viewers that an unchanged ad was produced with AI can change their beliefs about it. A design that changes both at once estimates their combined effect. If the objective is to isolate origin information, the creative material needs to remain constant or its variation must be explicitly accounted for. Research about origin perception is related, but an inferred belief is not the same intervention as an assigned disclosure.

2. Review method and source boundaries

Targeted searches on 21 October 2026 combined AI advertising disclosure, trust, credibility, role transparency and persuasion knowledge, followed by author and title searches. Publisher records and author-institution abstracts were consulted for the central advertising experiments. Official European Commission, CAP and C2PA material was read for the distinct regulatory and technical concepts. Twelve sources form the selected evidence base. The selection is purposive and cannot establish the prevalence of an effect across every disclosure study.

Publication dates and source versions are distinguished. Baek and colleagues’ prosocial-advertising article was online in 2024 before its 2026 issue assignment [3]. This snapshot uses the Exner MSI manuscript distributed by January 2026 [10]. Its sibling-ad comparison is not combined with sample descriptions or figures from later records. A finding remains tied to the version in which its design, population and outcome are reported.

Several central studies were available only as abstracts or indexed publisher text. Their findings are described at that level, without numerical mediation estimates, unreported subgroup effects or claims about robustness that require full-text appraisal. The review introduces no experiment or new causal theory. Conceptual distinctions in the figures organise the existing questions; they do not represent measured causal coefficients.

3. Persuasion knowledge as a theoretical foundation

Friestad and Wright’s persuasion knowledge model describes how people use knowledge about persuasion agents’ goals and tactics when interpreting attempts to influence them. It provides a framework for understanding why production information could change an advertising response [1]. An AI label may alter beliefs about the advertiser’s effort, competence or intention, but the framework does not imply one fixed direction of effect. Those beliefs require measurement in the particular context.

A useful distinction is between knowing that a technique has been used and judging whether its use is appropriate. A viewer can recognise generated content and approve of it, or fail to recognise it while disliking the execution. Likewise, trust in a particular claim differs from trust in the organisation that published it. A study that measures one construct cannot stand in for all of them. The constructs may correlate without being interchangeable.

Theoretical language can also outrun identification. If disclosure is randomly assigned and credibility is measured afterwards, a difference in credibility can be attributed to assignment under the design’s assumptions. Claiming that credibility fully causes the later change in trust requires stronger assumptions about the mediator and other variables. Statistical mediation can be informative, but a post-treatment association does not automatically exclude alternative pathways.

The persuasion-knowledge framework predates generative advertising. That history makes it relevant to interpreting source cues, while leaving the current technology-specific mechanisms open to test. It should be used to formulate and distinguish measured constructs rather than to relabel every negative response as resistance. A disclosure might also improve understanding or make a production process seem more accountable, depending on the message and audience.

4. Direct experimental findings

Koning and Voorveld compare an AI-generated advertisement with and without disclosure in a between-subjects experiment involving 304 participants. The author-institution abstract reports increases in conceptual AI knowledge and attitudinal persuasion knowledge, with reduced trust in the advertisement and organisation through reported pathways. The publisher record also describes a positive pathway involving persuasion knowledge. The coexistence of these paths makes a single unqualified disclosure penalty an incomplete description [2].

The tested asset and notice matter. An experiment can identify a response to that combination, but its estimate need not apply to every product category, disclosure phrase or degree of human involvement. A short social-media ad encountered once may elicit a different response from repeated exposure to a familiar brand. The abstract-level appraisal here does not establish how those conditions would alter the result.

Baek and colleagues examine disclosure in prosocial advertising. Their author-university abstract reports that initial disclosure produced less favourable ad attitudes, with perceived credibility relevant to the relationship. A further study finds a reduced negative effect among participants who view AI as more human-like, while another examines donation intention [3]. These are context-specific responses to charitable advertising. Donation intention is a reported intention rather than an observed transfer of money.

The prosocial setting adds meaning to the stimulus. An image of a beneficiary can be interpreted as depicting a real person, an illustration or a symbolic representation. A notice can change that interpretation as well as the inferred production method. Consequently, an observed credibility effect may concern what the image is taken to represent. It cannot be assumed to describe a generated background texture or a clearly fantastical scene, where the audience is asked to believe something different.

Le and colleagues investigate AI role disclosure transparency: how clearly the communication explains AI’s role in creating an ad. Their publisher and institutional abstracts connect that perception to ad-creation-process credibility and ad evaluation [4]. This shifts the question from the mere presence of a label to the information it supplies. The material consulted does not support a universal ranking of wording formats or a numerical estimate for every disclosure strategy.

5. Aversion, appreciation and task expectations

Dietvorst, Simmons and Massey study algorithm aversion after people observe algorithms make errors. The result concerns a defined forecasting context and comparison with human judgment [5]. It cannot establish that all audiences dislike all algorithmic contributions. The mechanism also involves exposure to performance, which differs from seeing an origin label on an ad without observing how the tool works.

Logg, Minson and Moore document algorithm appreciation in other judgment settings, finding that people can prefer algorithmic to human advice [6]. Read together, the studies show that algorithm origin is interpreted within a task and comparison. Neither supplies a general consumer coefficient that can be attached to generated advertising. Their value here is to prevent a blanket assumption about audiences’ preferences for human or algorithmic sources.

Task expectations can change what a viewer learns from a disclosure. If an audience expects precise calculation, algorithmic involvement may convey a different cue from involvement in a personal testimonial. A generated voice in an obviously fictional animation differs from a generated spokesperson who appears to report lived experience. This is a conceptual distinction about the claim being communicated, not a new empirical moderator established by this review.

Expectations also operate at the brand level. Prior familiarity may shape how a notice is interpreted, while unfamiliarity may leave the notice carrying more of the audience’s available information. Establishing such an interaction requires a design that measures or varies familiarity. It should not be inferred merely because two separate studies, conducted with different brands and samples, report different outcomes. Between-study differences can reflect many changes at once.

6. Information clarity, attention and comprehension

A disclosure can be present without being noticed. It can be noticed without being understood, and understood without changing a decision. These stages require different measurements. A manipulation check asking whether a participant recalls a notice provides information about reception, while a comprehension question can show what the notice was taken to mean. An outcome measured only among people who remembered the label can introduce selection relative to the original randomised assignment.

Wording such as AI-generated can cover a wide range of processes. Explaining a specific role may give the audience more relevant information, but a longer explanation also changes reading effort and prominence. A comparison that holds total length and presentation constant can address a more precise question than one that replaces a short label with a paragraph. The role-transparency research makes that information dimension explicit [4].

Timing introduces another distinction. A label shown before exposure can affect expectations during processing; one shown afterwards can change retrospective evaluation. Those interventions need not have the same effect. A research design must say whether the audience encounters the disclosure with the asset, after answering evaluation questions or in a separate explanation. A single label-versus-no-label result cannot answer every timing question.

The relevant denominator is the assigned audience when estimating the effect of a disclosure policy. Excluding people who miss the notice may describe responses among attentive viewers but changes the question. Both analyses can be useful if labelled correctly. The resulting estimates should not be treated as interchangeable measures of what would happen if a platform introduced the notice for all users.

7. Regulatory transparency and empirical persuasion

Article 50 of the 2024 AI Act distinguishes provider obligations concerning machine-readable marking from deployer obligations concerning disclosure of deep-fake content. It also addresses other specified AI interactions and text published to inform the public on matters of public interest [7]. The provisions define categories of obligation; they do not constitute evidence that a notice increases or reduces trust. This review confines its legal discussion to that conceptual separation.

CAP’s May 2025 guidance states that existing advertising rules apply regardless of how content is produced. It describes no blanket UK requirement to disclose every AI use and asks whether the audience would be misled without disclosure. The guidance also distinguishes clarifying an ad’s message from contradicting it [8]. Its reasoning concerns the meaning communicated to consumers, which can differ from the mere presence of an AI production step.

These official materials should not be used to convert an effectiveness finding into permission to omit a required notice. Equally, a requirement to disclose cannot establish that every notice provides adequate comprehension. Normative or legal requirements and empirical response are separate questions with different sources of authority. A consumer experiment can inform understanding of a format, while its lawful application depends on the relevant rules and circumstances.

The current review does not map every jurisdiction, enforcement interpretation or platform policy. Such a map would need continuing updates and a different scope. Its narrower point is that a legal transparency category, a visible label and the underlying claim can each vary independently. Combining them under a single binary AI disclosure variable obscures the information an audience actually receives.

8. Technical provenance and the truth of an advertisement

C2PA specifies a system for binding signed provenance assertions to digital content. Its technical structure concerns records of origin and modification, signatures and validation [9]. Those mechanisms can help establish whether a record is intact and who asserted it. They do not independently establish that a product claim is true or that the depicted event occurred as a viewer imagines.

A content record can accurately identify a generated image while the associated headline remains unsupported. Conversely, the absence of a provenance record does not by itself demonstrate that an asset is fabricated. The scope of the record matters: what is asserted, who signed it and which changes are represented. A validation result therefore needs an interpretation at the level of the assertion, rather than a general stamp of truth.

A visible label and a machine-readable record also have different audiences. The former is designed for people encountering the ad. The latter can be processed by systems and may support an interface that displays provenance information. A technical standard does not guarantee that all platforms preserve the record or present it in the same way. The specification describes an interoperable mechanism; consumer comprehension remains an empirical question.

This distinction is relevant to experimental design. If participants see a provenance interface, the treatment includes its wording, visual hierarchy and source cues. A result from a bare AI-generated label cannot establish the response to that richer interface. The test needs to name the information displayed and the claims participants believe it verifies.

9. Why creative quality and disclosure effects can coexist

The Exner MSI manuscript compares 4,633 sibling ads under matched advertiser and campaign settings. Its abstract reports a click-through advantage for generated images conditional on their not looking generated [10]. This concerns visual origin inference. Assigned disclosure changes information given to a viewer, whereas perceived origin can reflect the visual material. The comparisons cannot be combined into a causal claim that withholding a label improves performance.

Meguellati and colleagues’ preprint reports a preference advantage for selected generated ads in one of its advertising comparisons [11]. Hartmann and colleagues report strong visual evaluations for generated marketing images [12]. Neither finding eliminates the possibility that origin information changes evaluation of the same material. The value of a creative execution and a response to its stated production process are separable outcomes, even when both influence an overall judgment.

The converse inference is equally weak. A negative response to a notice does not prove that generated content is visually inferior. An experiment can hold the asset constant and change trust by changing the description. A complete interpretation needs the stimulus, disclosure and endpoint together. This is particularly relevant where reported attitudes, clicks and purchases are discussed as if they were the same measure of advertising effectiveness.

10. Synthesis, uncertainty and conclusion

The selected evidence shows that AI disclosure can alter advertising response, with credibility, persuasion knowledge and role clarity relevant to the reported relationships. It does not support a single penalty applicable to every asset or audience. The direct experiments address specific notices and contexts, while broader algorithm-attitude research demonstrates that preferences for human and algorithmic sources vary with the task.

Source access limits the depth of appraisal. Several central reports were consulted through author or publisher abstracts, so the review cannot independently judge all exclusions, manipulation checks or mediation assumptions. No pooled estimate is offered. Studies using prosocial ads, fitness advertising and live display inventory also differ in audience, exposure and outcome, preventing a clean attribution of between-study differences to any one design feature.

The research question is best answered at the level of information and response. A disclosure communicates a particular account of how an asset was made; its effect depends on what that account changes in the audience’s understanding. Empirical studies can measure those changes. Technical records can document production assertions, and official rules can specify disclosure obligations. Keeping those functions distinct makes the evidence interpretable without assuming either universal distrust or universal acceptance of AI-created advertising.

Source access and appraisal register

The register describes the material consulted for this narrative review. Access to a manuscript does not imply replication, independent extraction or a formal risk-of-bias assessment. Source dates identify the underlying research.

References