REACT Trust Scaffolding in Real-Time GenAI Advertising

Benjamin George, Bartlomiej Hanus, Steven Wallace

Generative AI now synthesizes personalized video advertisements in real time from live behavioral signals, with industry forecasts predicting that 40 percent of digital advertising will be generative by 2026. This introduces governance conditions that neither the algorithmic personalization literature nor the AI-generated content literature has previously addressed: the creative artifact is synthesized in the moment of consumption, delivered only to a single consumer, and may never be reviewed by any human. We propose REACT, a five-layer governance architecture for this context (Real-time risk assessment, Ethical content scoring, Accountability mechanisms, Compliance assurance, Trust scaffolding), and develop the consumer-facing Trust Scaffolding layer for empirical investigation. Three distinctive properties of real-time generative advertising alter how established consumer-response theories operate. First, “audit opacity”: because no human reviews the specific variant delivered to any given consumer, transparency disclosure becomes the only accessible legitimacy cue under the Persuasion Knowledge Model (Friestad & Wright, 1994). Second, “psychological proximity”: because data use is immediate and specific rather than abstract and future-oriented, consent signals operate on a more concrete object than the Internet Users’ Information Privacy Concerns framework (Malhotra et al., 2004) has previously addressed. Third, “provenance ambiguity”: because consumers cannot infer whether outputs were produced under human oversight, oversight indicators resolve the source identification that trust assessment (McKnight et al., 2002) otherwise requires. Each property maps to one consumer-facing governance signal: transparency to audit opacity, consent to psychological proximity, and oversight to provenance ambiguity. Our in-progress agenda tests these mappings in a pre-registered 2×2×2 between-subjects factorial experiment, with real-time salience established through a cover-story vignette and in-stimulus personalization cue, and validated through manipulation checks. We test a signal-coherence prediction (full-bundle superadditivity) against a competing cognitive-load prediction (multiple simultaneous disclosures producing subadditive effects). This work proposes the first empirical test of REACT’s Trust Scaffolding layer and invites audience discussion of the framing.