Making Sense of AI-Generated Disinformation: How Audience Interpretations Influence the Impact of Deepfakes in Kenya

Participatory Disinformation and Generative AI

The strategic dissemination of false and misleading information has contributed to diverse negative impacts around the world. Some of the more apparent domains where these phenomena are visible include democratic backsliding (Benkler et al., 2018, 2020; Bennett & Livingston, 2018, 2020; Ong & Cabañes, 2018; Batista Pereira et al., 2022), climate science denialism (Oreskes & Conway, 2010), and health misinformation, especially surrounding vaccine skepticism (Beers, Nguyen, et al., 2023; Moran et al., 2022). Recent reductions in the resources dedicated by social media platforms to moderation have exacerbated these challenges at the same time as the popularization of AI tools has made it easier to produce strategic content at scale.

Generally speaking, disinformation is defined as false or misleading information that is intentionally disseminated to further larger goals, including political or financial gain or simply the disruption of a targeted population’s ability to understand a situation (Jack, 2017). Disinformation campaigns are often conceptualized as including a mix of true and false information, at times obfuscating the nature of the campaign and its apparent goals while making it difficult for the public to make sense of reality. Critically, disinformation, particularly disinformation on social media, is heavily participatory (Starbird et al., 2019). As messages and information proliferate, targeted populations pick up the narratives and amplify the content unwittingly as they begin to believe the messages (Starbird et al., 2019). The processes by which this occurs are diverse, widely studied, and adversarial, meaning that as deceptive tactics are illuminated, propagandists adapt their methods to further avoid detection. Researchers have highlighted numerous tactics utilized by propagandists, including hack and leak campaigns (e.g., Rid, 2020), “astroturfing” which describes the wide use of inauthentic, often bot, accounts that pretend to be real users supporting a message (Keller et al., 2020), and efforts to prompt users to interpret digital affordances, perceived pathways for acting in a digital environment (e.g., a “like” button), in a specific misleading way, such as pushing an audience to misinterpret the significance of a highly cited source on Google Scholar (Tripodi et al., 2023).

Recent work has also highlighted how disinformation targeting U.S. elections often relies on the network structure of social media and its attendant attention economy (Beers, Schafer, et al., 2023) to promote particular false or misleading information. This work emphasizes the adaptive performances of influencers and political elites who strategically amplify certain content in ways that resonate with cultural narratives, capturing the attention of audiences while simultaneously promoting a specific ideological position that is often contrary to existing evidence (Prochaska et al., 2025; Starbird et al., 2023). Modern campaigns are broad, and utilize many tactics simultaneously, making detection and mitigation challenging, especially given that the participation of unwitting audiences makes ascertaining the primary drivers of a campaign difficult.

Context also matters. The impacts of disinformation are expressed differently depending on the unique cultural and socio-political contexts of targeted populations. Previous research has documented disinformation campaigns and their unique expressions across a diverse set of cultural and geographic regions. ² In addition, an anti-elite, populist tone has been documented in India (Gonawela et al., 2018; Kyle & Gultchin, 2018), Argentina (Centenera & Criales, 2023), and the United States (Gonawela et al., 2018; Hameleers, 2020), suggesting that while disinformation may be regionally specific, there are themes that appear to be widespread globally. This literature highlights the breadth of the problem and demonstrates how mitigation efforts must be relevant to the specific populations targeted while also considering intersections in international discourse, such as populist and anti-media sentiments.

Recent advances in generative AI further complicate mitigation, as propagandists have access to more sophisticated tools to advance their goals. In addition to the potential for AI-generated content to mislead on its own, scholars have discussed challenges like the “liar’s dividend” which describes the potential for the mere existence of AI content to be used by bad actors to claim real content is fabricated (Chesney & Citron, 2019). Recent work has begun to compile an evidentiary base documenting how AI-generated content is being integrated into ongoing disinformation campaigns, highlighting the rapid adoption of novel technologies by propagandists. For example, Wack, Ehrett, et al. (2025) examine how these concerns have already begun to manifest as real-world threats. The authors found that the quantity and breadth of disinformation produced by a known disinformation outlet increased through access to AI tools, and that articles remained persuasive despite the outlet’s use of AI. In light of this evidence, it is more critical than ever that global societies understand how publics make sense of deceptive AI-generated content. The current study provides some initial insights toward this end.

Interpreting AI-Generated Content

The growing prevalence of deepfakes has prompted considerable scholarly attention, including investigations of technical detection methods and a growing body of work examining the psychological mechanisms that guide audience responses despite an initial dearth of studies (Hancock & Bailenson, 2021). A central thread of inquiry has examined whether deepfakes are more credible or persuasive than other forms of disinformation. Early work found that viewers of a viral deepfake of President Obama became more distrustful of news on social media even though they did not believe the video was real (Vaccari & Chadwick, 2020). Barari et al. (2021) conducted a study showing that deepfakes could mislead the public about fabricated scandals, but that their persuasive impact was similar to that of equivalent claims in text or audio. In contrast, Sundar et al. (2021) found that Indian WhatsApp users were more likely to believe deepfake videos than audio or text, and other studies similarly found deepfakes to be more credible than alternative formats (Hwang et al., 2021; Lee & Shin, 2022), particularly in terms of source vividness, which in turn enhanced credibility and engagement intentions (Lee & Shin, 2022). However, Hameleers et al. (2022) found minimal differences in how audiences interpreted deepfakes versus text-based disinformation, and, relevant for our study, observed that credibility and perceptions of politicians were influenced more by social media source cues than by the format of the disinformation itself. In addition, a meta-analysis of 24 experimental studies across 10 countries found that while deepfakes elevated emotional responses, their impact on credibility was not uniform, and that media literacy moderated effects on both credibility judgments and sharing intention (Kang & Valadez, 2025).

Beyond the question of whether deepfakes deceive, a growing body of work has examined the cognitive mechanisms underlying audience judgments. Shin (2026) found that users who view deepfakes as realistic are more likely to consider sharing them, but that reliance on cognitive shortcuts often overrides analytical reasoning, with detection improving when users are exposed to relevant cues and explanatory information. Similarly, Hu and Huang (2025) found that both perceived technical quality and content familiarity function as heuristic cues that increase perceptions of video realness and trust in the source, with content familiarity exerting a stronger effect on trust for synthetic videos than for real ones. Moreover, deepfakes containing obvious lies tend to be perceived as less credible, while subtle distortions are often seen as more credible, especially if they contain figures with whom viewers identify (Hameleers et al., 2024).

Importantly for our current study, credibility judgments are also shaped by system-generated cues such as follower count and popularity, emphasizing the importance of contextual cues for shaping interpretations (Jin et al., 2025). Similarly, a study conducted across eight different countries found that participants who relied on social media for news increased the likelihood that they believed deepfakes after repeated exposure (Ahmed et al., 2024). In addition to context cues, the domain of a deepfake plays a role in how misperceptions are expressed. For example, health-related deepfakes are associated with significantly increased misperceptions about topics such as drinking cold water and cancer (Lee & Hameleers, 2025) and flu vaccines (Lee et al., 2024).

Despite this accumulating evidence, several important gaps remain. The existing literature has largely treated deepfake interpretation as an individual cognitive act, underexploring the social and platform contexts in which audiences actually encounter this content. In addition, cultural differences in the studies reviewed, such as between the Indian WhatsApp users in Sundar et al.’s (2021) study and the Dutch participants in Hameleers et al.’s (2022) investigation, may explain some variation in results, yet relatively few studies account for cross-cultural differences. Most importantly for our study, the majority of studies expose participants to deepfakes in controlled, decontextualized settings, which limits their ecological validity. Some studies have attempted to investigate specific cues (e.g., Jin et al., 2025), but in general, experimental designs fail to fully grapple with how audiences collectively make sense of deepfakes in the social media environments where they are actually encountered, a limitation we seek to fill in our work.

Collective Sensemaking and Rumors on Social Media

Much of the previous work examining how deepfakes are interpreted builds on psychology research seeking to understand how individuals interpret information, with much of this work focused on micro-level processing of discrete pieces of information by individual participants. For example, literature on motivated reasoning (e.g., Xu et al., 2025; see also Kunda, 1990; Pennycook & Rand, 2019) is useful for understanding individualized cognitive effects, but doesn’t always account for the social, discursive processing of information that many people utilize when they make sense of novel or ambiguous information. To fill some of this gap, collective sensemaking theory provides one useful framework for understanding meso-level processing in social settings. We note that other disciplines, including Science and Technology Studies (STS), offer complementary perspectives on how technological artifacts are socially interpreted and contested (e.g., Jasanoff, 2015). However, we adopt collective sensemaking as our primary framework because it most directly addresses the real-time, discursive process by which groups interpret uncertain information in social settings, which is the dynamic that our experimental design is intended to capture.

At its core, collective sensemaking describes the process by which groups make sense of uncertain situations through the interaction of both individual and social processing (Weick, 1995). Much of the theorizing surrounding sensemaking emerges from organizational psychology, seeking to improve the efficiency and accuracy of decision-making in rapidly evolving environments, including in scenarios where human or man-made disasters require a rapid response to minimize loss of life and/or infrastructure (e.g., Maitlis & Sonenshein, 2010; Weick, 1993). Based in part on this intellectual tradition, work from crisis informatics has applied this theory to social media spaces where mass audiences frequently attempt to collectively make sense of emerging events as they happen (Arif et al., 2017; Maddock et al., 2015; Zhou et al., 2023). Much of this work uses qualitative interpretation and computational tools to analyze observational data as opposed to the experimental methods utilized in much of the psychology literature. At times these methodological differences can complicate our understanding of the precise, individualized effects which result from instances of collective sensemaking. However, we believe that it is useful to integrate the theories emerging from experimental work with the mixed-method framework of collective sensemaking to more fully inform the interpretation of our findings by situating them within the digitally mediated, discursive processes documented by crisis informatics work.

Studies that utilize collective sensemaking theory to understand how and why misinformation spreads online often do so by highlighting the relationship between collective sensemaking and rumors (Arif et al., 2017; Chandra & Pal, 2019; Maddock et al., 2015). Rumors are unofficial stories defined by their inherent uncertainty and spread through informal channels (Kapferer, 2013; Shibutani, 1966). Shibutani (1966) proposed that rumors are similar to hypotheses that emerge in reaction to uncertain events, such as elections (Wack, Schafer, et al., 2025), as collectives try to make sense of novel or ambiguous circumstances. Importantly, rumors are not defined by facticity; they can be either true, false, or somewhere in between, as their facticity is often unknown when they begin to spread. Viewed in this light, rumors often emerge as a byproduct of collective sensemaking as groups propose or discard potential explanations for the event in question (Shibutani, 1966).

More recent work applying the concept of rumors to social media discussions has highlighted how sensemaking processes can naturally go wrong (e.g., Arif et al., 2017; Chandra & Pal, 2019), as well as how those processes can be intentionally disrupted and manipulated by disinformation campaigns and propagandists (Prochaska et al., 2023; Starbird et al., 2023). The advances of AI further exacerbate this problem (e.g., Wack, Ehrett, et al., 2025), making it even easier to run disinformation campaigns by easily creating large amounts of strategically useful content. This makes uninterrupted collective sensemaking ever more difficult in social media spaces that don’t dedicate resources to minimizing this kind of manipulation. It is within this context that our study sits, providing insight into how individuals interpret information in the presence of a small number of comments that suggest particular interpretations for participants. Importantly, while our study design does not emulate the full process of collective sensemaking, as that requires iterative discourse, it simulates a critical step of the process by examining how participants’ interpretations are informed by audience commentary.

Case: Kenya’s Political Landscape

Kenya, the middle-income country that serves as the base for the study, boasts one of the most active and well-connected social media user bases in Africa. Despite the recency of the country’s transition to democracy, Kenyan society is characterized by its robust civil society and active participation in democratic politics. We leverage the competitiveness of the country’s politics and the digital participation of the national public to conduct an experiment within the context of a region which has been identified as uniquely susceptible to both misinformation and electoral falsehoods (Badrinathan & Chauchard, 2024). Specifically, we draw on the electoral contest held between the eventual victor and current Kenyan President William Ruto and democratic activist and opposition party candidate Raila Odinga.

Reflective of electoral conditions in other first-past-the-post presidential democracies, the election became a two-party contest pitting the two candidates in direct competition. The narrow margins between the candidates drew increased attention to the contest from both genuine and malicious actors. The level of competition and stakes of the contest were reflected in efforts aimed at distorting the result, including the presence of disinformation campaigns and ubiquitous electoral misinformation (Miyandazi & Thuo, 2025) and the use of deceptive deepfakes and related forms of propaganda developed using Generative AI tools (Owino, 2022; Wack, Mudavadi, & Jalbert, 2025). In using this contest as the context for our study, we were able to ground our experiment in a rapidly digitizing LMIC currently struggling to manage the consequences of widespread AI-augmented electoral disinformation.

Importantly, Kenya’s relationship with AI technologies extends beyond their use as tools for disinformation. Kenyan workers have been at the center of global AI labor dynamics, including the documented exploitation of content moderators and data labelers working for major technology companies under poor conditions and low wages (Perrigo, 2023). This context underscores the importance of ensuring that research on AI in African contexts is not limited to documenting exploitation but also examines how these technologies affect citizens as audiences and political actors. By studying how Kenyan audiences interpret AI-generated content, we contribute to an understanding of AI’s impacts that centers the agency and interpretive practices of those most directly affected by these technologies.

Finally, we use the term “partisan” to describe alignment with one of the two major political coalitions, Kenya Kwanza (led by William Ruto) and Azimio la Umoja (led by Raila Odinga), rather than as a reference to bifurcated ideological partisanship. Political alignment in Kenya is substantially shaped by ethnic and regional affiliations, with Ruto drawing in supporters from the Kalenjin community and Odinga from the Luo community, though both coalitions in this most recent election represented broader multi-ethnic alliances. Our operationalization of partisanship, based on participants’ self-reported candidate support, captures this coalition-based alignment.

Comment Conditions

As discussed, our treatment revolved around the audience’s interpretation of the multimedia evidence. To accomplish this, we developed the accompanying comments to reflect levels of consensus. Specifically, participants were further stratified into three comment-consensus categories: (a) credulous messaging, (b) incredulous messaging, and (c) mixed messages. Figure 2 presents an example of each treatment comment:

OPEN IN VIEWER

Figure 2.

Comment conditions.

As the focus of the study was dependent on the external validity of participant assessments of participant comments, we attempted to improve the verisimilitude of the comments by basing the comment treatment on the internal commenting features present on online platforms, including YouTube. To do so, we developed a novel survey instrument that integrated the comments into the survey flow, which aimed to make the comments appear as natural as possible. This included the integration of an additional option for participants to add their own comment as well as through modeling the comment features on previous surveys conducted on the platform. In addition, participants were informed that they would be given the opportunity to leave comments in an attempt to add to the validity of the experimental design (see Figure 1). While prior attempts to improve the validity of comment conditions have relied on screenshots of social media posts to simulate interactions with stimuli, our design intended to separate the comments from the stimuli for further differentiation.

Immediately after viewing the treatment and reading these related comment(s), participants were asked a series of questions about the media they encountered as well as their perceptions of the wider community’s assessment of the policies discussed in the treatment. To operationalize personal support, participants were asked questions about the candidate they viewed. Specifically, participants were prompted to respond on a sliding scale from 0 to 100 to the following prompt: “Rate how warmly you feel towards the political candidate (0 – cold; 100 – warm)?” ⁵ In addition, participants were asked to respond to elicitations of accuracy (i.e., “How likely is it that this media source was legitimate?”) as well as questions on trust in institutions and political corruption. Specifically, participants were asked “How much of the time do you think you can trust each of the following groups to do what is right?” Respondents responded to the query with regard to the national government, the media, and ordinary people. Participants were also asked “How many of the politicians running the government do you believe are corrupt?” Responses ranged from “all of them” to “none of them” (full survey questions and responses can be found in the Supplemental Appendix). Finally, participant demographics were recorded. This included participants’ ethnicity, gender, age, and educational attainment.

We acknowledge that the experimental setting may introduce demand characteristics that could affect participants’ self-reporting. Several design features were designed to help mitigate this concern. First, participants were not informed that the stimuli were AI-generated prior to viewing, preserving ambiguity about the content’s authenticity. Second, participants were told they would be viewing recordings from political figures and social media (framing consistent with real-world encounters rather than experimental expectations). Third, the comment conditions were integrated into the survey flow using a novel instrument designed to make them appear as natural social media commentary rather than experimental prompts. Fourth, the 95% correct identification rate during the post-treatment debrief check (reported in Supplemental Appendix A1) suggests participants engaged with the stimuli as intended. Nevertheless, we note this as a limitation in the “Discussion” section.

Effects of Partisanship

Our findings also show that partisan alignment moderates support for the candidates depicted in the deepfake (Figure 4). Exploratory partisan-subgroup models reveal that the comment effects are conditioned by whether the respondent is politically aligned with the candidate featured in the clips. Among participants who saw politicians they supported, simply viewing the video without any comment depressed candidate support by a little more than three points relative to the baseline group that never saw the clip (β = −3.11, SE = 1.08, t = −2.88, p = .004). Both the incredulous (“clearly fake”) and the credulous (“not surprising”) replies produced similar reductions (β = −2.97, SE = 1.09, t = −2.72, p < .01; β = −4.14, SE = 1.09, t = −3.81, p < .001, respectively), while the mixed set of comments had no reliable impact (β = −1.24, SE = 1.12, t = −1.11, p = .27). Re‑estimating the model with the no‑comment video as the reference confirms that none of the comment types further shift evaluations among co‑partisans (all β ⩽ 1.89, all p-values > .16).

OPEN IN VIEWER

Figure 4.

Estimated personal support by condition.

For respondents whose party affiliation did not match that of the politician depicted, the pattern partly reverses. The incredulous dismissal that the clip is “clearly fake” improves candidate evaluations, raising support by 2.78 points relative to the baseline (β = 2.78, SE = 1.49, t = 1.87, p = .06) and by nearly five points compared with the treatment seen without comments (β = 4.77, SE = 1.75, t = 2.72, p = .006). In contrast, the credulous “not surprising” remark reduces support by about three points versus baseline (β = −3.14, SE = 1.45, t = −2.17, p = .03) but does not differ from the no‑comment condition (β = −1.15, SE = 1.72, t = −0.67, p = .50). The mixed‑comment treatment remains indistinguishable from either comparison group (all p-values ⩾ .33).

Taken together, these subgroup analyses indicate that co‑partisans are generally susceptible to the negative tone of the video itself, with little additional movement from comment cues. Cross‑partisans, however, respond positively when another viewer explicitly questions the clip’s authenticity, suggesting that an incredulous counter‑narrative can blunt, and possibly even reverse, the persuasive power of potentially damaging political content among opponents.

Encouragingly, estimates of support are moderated by participants’ perceived accuracy of the video in question. Figure 5 shows that when participants view the video as real, their support for the video’s subject decreases compared to the baseline, while when they view the video as fake, their estimated support for the video’s subject actually increases above the baseline, although this does not reach the standard significance threshold of p < .05.

OPEN IN VIEWER

Figure 5.

Estimated support for the candidate depicted in the deepfake by whether participants perceived the content to be real or fake.