Wack, M., & Prochaska, S. (2026). Making sense of AI-generated disinformation: How audience interpretations influence the impact of deepfakes in Kenya. Social Media + Society, 12. https://doi.org/10.1177/20563051261462092
Summary
This study investigates whether the visible interpretations of other audience members shape how people make sense of political deepfakes. Set in Kenya during the Ruto-Odinga electoral rivalry, the authors ran a large preregistered survey experiment (N=7,008) in which participants viewed an AI-generated audio or video clip of a presidential candidate confessing to a fabricated corruption scheme, accompanied by embedded comments of varying skepticism. The central argument is that deepfake impact is not fixed at the point of production but is socially negotiated through “collective sensemaking” as content circulates — with peer commentary and partisanship substantially moderating whether the fake damages or spares a candidate’s standing.
Key Contributions
- Experimental evidence that deepfake interpretation is participatory and social, not solely a matter of individual cognition.
- A novel in-survey instrument that embeds platform-style audience comments to improve ecological validity in deepfake experiments.
- Extension of collective sensemaking and motivated reasoning theory to AI-generated audio-visual disinformation.
- Centers an under-studied LMIC context (Kenya), foregrounding the interpretive agency of African audiences and non-Western misinformation scholarship.
- Practical insight into the cues that guide or mislead audiences in low-moderation platform environments.
Methods
- Preregistered (OSF #182038) between-subjects online survey experiment with 7,008 validated Kenyan participants recruited via Instagram, WhatsApp, and Facebook ads, stratified by region, age, and gender.
- Random assignment to video (N=2,820), audio (N=2,733), or control (N=1,455) conditions; depicted candidate (Ruto or Odinga) also randomized to test aligned vs. misaligned partisanship.
- Deepfakes built with open-source DeepFaceLab and Wav2Lip (video) and ElevenLabs voice cloning (audio), refined with local feedback for plausibility; no AI labels or watermarks were used to preserve ecological validity.
- Three embedded comment conditions — credulous (“not surprising”), incredulous (“clearly fake”), and mixed.
- Outcomes: 0-100 candidate feeling thermometer, perceived media legitimacy, institutional trust, and corruption perceptions; analyzed via robust regression clustering standard errors at the respondent level, with extensive debriefing and comprehension checks.
Findings
- Merely viewing the video modestly reduced candidate support (β=-2.25, p=.04) relative to baseline.
- Credulous “not surprising” comments produced the sharpest decline (β=-4.90 vs. baseline; β=-2.56 vs. no-comment video).
- Incredulous “clearly fake” comments had no reliable effect versus baseline but yielded a non-significant uptick versus the no-comment condition.
- Mixed comments produced an intermediate drop, ending statistically similar to control — suggesting induced uncertainty.
- Partisanship moderated effects: among co-partisans support fell across all conditions with little movement from cues; among cross-partisans, incredulous comments raised support (~5 points, p=.006), partially reversing the deepfake’s damage.
- Perceived authenticity mediated the effect: seeing the clip as real lowered support, while seeing it as fake raised support above baseline.
Connections
This work is a direct application of participatory disinformation theory to AI-generated media, closely tied to Starbird’s crisis-informatics program Starbird2025-jj and related sensemaking work by Prochaska Prochaska2025-ef. It sits within the broader empirical literature on deepfake persuasion and detection — connecting to studies of AI-generated multimodal disinformation and its perceived authenticity such as Hameleers2026-mc and Di-Domenico2026-zq. Its LMIC focus and platform-context emphasis resonate with non-Western disinformation scholarship including Waight2026-ts and Waight2025-al.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify