This Mini Review employs a narrative synthesis approach to examine the intersection of AI-generated visual disinformation and social equity. Literature was identified through systematic searches of academic databases (Scopus, Web of Science, Google Scholar) and specialized repositories (arXiv, SSRN) conducted between October and December 2025. Search terms included combinations of “generative AI,” “deepfakes,” “visual disinformation,” “algorithmic bias,” “marginalized communities,” “digital equity,” and “intersectionality.” The review prioritized peer-reviewed articles, institutional reports, and policy documents published between 2023 and 2025 to capture recent developments in generative AI technologies, with foundational theoretical works included for conceptual grounding.

Inclusion criteria focused on: (1) empirical studies or theoretical analyses addressing AI-generated visual content; (2) research examining differential impacts across social identities or communities; (3) scholarship on platform governance, algorithmic systems, or digital equity; and (4) English-language sources with accessible full text. The final corpus comprised approximately 50 sources spanning communication studies, critical algorithm studies, computer science, sociology, and policy research, with representation from North American, European, and Global South contexts. Thematic analysis identified four primary analytical domains: AI production mechanisms, intersectional vulnerability frameworks, platform algorithmic governance, and equity-centered interventions. The vulnerabilities matrix (Table 1) represents a conceptual synthesis derived from patterns across the reviewed literature rather than original empirical data.

Recent advances in generative AI have democratized access to sophisticated image manipulation and generation tools. These technologies operate through complex training processes on massive datasets, learning to generate novel visual content that mimics authentic photographs and videos. The global market for AI-generated deepfakes is projected to reach $79.1 million by the end of 2024, with a compound annual growth rate of 37.6% (Perkins and Verevis, 2015). In 2023, deepfake fraud attempts accounted for 6.5% of total fraud incidents, marking a 2,137% increase over three years (Perkins and Verevis, 2015). While these innovations offer legitimate creative applications, they simultaneously create unprecedented opportunities for malicious actors to produce convincing disinformation at scale (Carr and Köhler, 2024).

Critically, the training data underlying these AI systems reflects and often amplifies existing societal biases. Studies of major image generation models reveal systematic underrepresentation of non-Western contexts, stereotypical portrayals of racial and ethnic minorities, and skewed gender representations (Bonil et al., 2025; Lendvai and Gosztonyi, 2025). When prompted with neutral descriptors, these systems disproportionately generate images aligned with dominant cultural stereotypes, effectively encoding discrimination into the visual output. These biased outputs become particularly harmful when deployed in disinformation campaigns targeting specific communities.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Beyond content generation, AI systems play a crucial role in amplifying and distributing visual disinformation through platform recommendation algorithms. Social media platforms employ sophisticated machine learning models to maximise engagement, often prioritising emotionally charged and controversial content regardless of veracity. Research demonstrates that false visual information spreads significantly faster than actual content, with platform algorithms inadvertently accelerating this dissemination (Hootsuite, 2025). Marginalised communities face heightened exposure to such content due to algorithmic targeting practices that exploit demographic characteristics and online behaviours (Arora et al., 2023). Platforms’ optimisation for engagement creates feedback loops in which sensationalised disinformation receives preferential distribution to vulnerable populations, compounding informational inequities.

Communities of colour experience distinctive vulnerabilities to AI-generated visual disinformation rooted in historical patterns of discrimination and contemporary algorithmic bias. Facial recognition systems, which often serve as components of verification tools, demonstrate significantly higher error rates for individuals with darker skin tones, particularly for minority groups (Franklin et al., 2024). This ‘technology performance gap’ means that minority communities cannot rely equally on AI-powered tools to detect synthetic media, leaving them with asymmetric protection against deepfakes and manipulated images.

Moreover, AI-generated disinformation frequently exploits racial and ethnic stereotypes, with documented cases of synthetic images depicting people of colour in criminal contexts or crises circulating widely during social movements and political campaigns (Insikt Group, 2024; Carr and Köhler, 2024). The cultural context of information evaluation further compounds these vulnerabilities. Communities with collectivist cultural orientations exhibit distinct trust heuristics and information-sharing patterns compared to individualist Western contexts. Yet, platform architectures and detection systems are predominantly designed based on Western user behaviours and values (Lendvai and Gosztonyi, 2025).

Socioeconomic disparities create fundamental inequities in both exposure to and protection from AI-driven visual disinformation. Low-income populations typically access the internet through mobile devices with limited screen size and computational capacity, constraining their ability to scrutinise image authenticity through zooming, reverse image search, or metadata analysis (Hollimon et al., 2025; Raihan et al., 2024). These communities also face restricted access to premium fact-checking services, verification tools, and high-quality media literacy education.

Research examining information behaviours across socioeconomic strata reveals that individuals with lower educational attainment and income demonstrate reduced confidence in identifying manipulated images and greater reliance on social proof as a credibility heuristic (Obed and Boateng, 2025). When AI-amplified disinformation reaches virality through engagement-optimizing algorithms, these social proof signals become particularly misleading for economically marginalised populations. Furthermore, communities experiencing economic precarity often lack the temporal resources for extensive fact-checking, creating time-based vulnerabilities that algorithmic systems can exploit through rapid-fire dissemination of synthetic content.

The Global South faces compounded vulnerabilities to AI-generated visual disinformation, including infrastructure limitations, linguistic barriers, and cultural distance from dominant fact-checking ecosystems. Major fact-checking organisations and verification platforms predominantly operate in English and serve Western audiences, leaving non-English-speaking communities with minimal access to authoritative debunking resources (Lendvai and Gosztonyi, 2025; Poggi, 2025). As of 2024, 2.6 billion people remained offline globally, with internet penetration in developing countries significantly lower than in developed nations (Poggi, 2025). AI-generated disinformation can be rapidly localised and culturally adapted to specific regions, whereas fact-checking responses lag significantly. Rural and remote communities additionally contend with limited bandwidth that restricts access to high-resolution image analysis tools while simultaneously experiencing targeted disinformation campaigns that exploit local socio-political tensions (Hollimon et al., 2025). The geographic concentration of AI development in wealthy nations further ensures that detection tools and platform safeguards are optimised to protect users in developed markets rather than to address the distinctive vulnerabilities of Global South populations.

Platform recommendation algorithms employ sophisticated demographic profiling to optimise content delivery, creating differential information environments across social groups. These systems analyse user interactions, demographic attributes, and inferred characteristics to predict engagement likelihood, often amplifying emotionally resonant content regardless of accuracy. Research demonstrates that algorithms disproportionately expose certain demographic groups to misleading visual content, particularly during politically charged periods or health crises (Liu et al., 2025; Hootsuite, 2025).

Critically, the engagement metrics that guide algorithmic distribution encode existing social inequalities. Content that reinforces dominant narratives and stereotypes typically generates higher engagement among majority populations, creating algorithmic feedback loops that marginalise counter-narratives and authentic representations from minority communities (O'Neil, 2016). When AI-generated disinformation exploits these stereotypes, platform algorithms may inadvertently prioritise its distribution to the very communities most harmed by it. This creates ‘algorithmic redlining,’ wherein marginalised groups receive inferior information quality through systematically biased content curation (Humber, 2023).

Content moderation systems, increasingly powered by AI, demonstrate systematic disparities in protecting different demographic groups from harmful disinformation. Automated moderation tools frequently misclassify content from minority communities due to biases in training data, leading to both over-moderation of legitimate speech and under-moderation of harmful disinformation (Bonil et al., 2025; Franklin et al., 2024). Analysis of platform enforcement reveals that AI-generated disinformation targeting marginalised communities receives slower removal and less comprehensive action compared to similar content affecting majority populations. These disparities reflect underlying inequities in how platforms perceive valuable users worthy of protection, with moderation resources concentrated on serving dominant demographic groups. The result is a two-tiered information ecosystem in which marginalised communities experience both heightened exposure to AI-driven disinformation and diminished platform accountability for addressing these harms.

The following matrix (Table 1) presents a conceptual synthesis of vulnerability patterns identified across the reviewed literature. These categories are analytical constructs that organize insights from multiple empirical studies, theoretical frameworks, and policy analyses rather than representing original empirical findings. The matrix illustrates how different axes of marginalization intersect to create compounded exposure to AI-generated visual disinformation.

Addressing the inequitable impacts of AI-generated visual disinformation requires a fundamental reconceptualisation of platform governance and algorithmic accountability. Current regulatory frameworks and industry self-regulation efforts predominantly centre on the experiences and needs of privileged users in developed nations, neglecting the distinctive vulnerabilities of marginalised communities (Lendvai and Gosztonyi, 2025). An equity-centred approach demands that platform policies, detection systems, and moderation practices be designed explicitly to protect the most vulnerable populations rather than to optimise for the majority user experience.

This reorientation requires several critical interventions. First, mandatory algorithmic impact assessments should evaluate how AI systems affect different demographic groups, with particular attention to compounding disadvantages faced by individuals at the intersection of multiple marginalised identities. Second, platform accountability mechanisms must incorporate meaningful participation from affected communities in the design, deployment, and oversight of content moderation systems (Eslami et al., 2025; Alon-Barkat et al., 2025). Research demonstrates that communities possess the capacity to critique algorithmic impacts when provided with participatory governance opportunities, challenging the assumption that technical complexity deters meaningful engagement. Third, regulatory frameworks should mandate equitable resource allocation for protecting diverse populations, preventing the current pattern wherein moderation capacity concentrates on serving dominant user groups (Negreiro, 2025).

Traditional media literacy interventions demonstrate limited effectiveness for addressing AI-driven disinformation among marginalised communities, as these programs typically assume Western cultural contexts, high educational attainment, and extensive digital access (Merod, 2025; DeMio, 2024). Culturally responsive approaches recognise that information evaluation practices are fundamentally shaped by community values, historical experiences with institutional trust, and culturally specific communication norms.

Effective interventions must be co-designed with target communities, incorporating local knowledge systems and leveraging existing community networks for peer education (Jackson et al., 2024). Programs should address AI-specific literacy needs, including understanding synthetic media generation, recognising algorithmic amplification patterns, and awareness of demographic targeting practices. Critically, such initiatives must confront the structural inequities that create differential vulnerabilities rather than placing responsibility solely on individual users to navigate hostile information environments (Naffi, 2025).

Despite growing recognition of equity concerns in AI systems, significant research gaps persist. First, empirical studies examining the specific mechanisms through which algorithmic systems disadvantage particular demographic groups remain limited, particularly for intersectional identities and Global South contexts (Pasipamire and Muroyiwa, 2024; Poggi, 2025). Second, longitudinal research tracking the evolution of vulnerabilities as AI technologies advance is urgently needed. Third, intervention effectiveness must be rigorously evaluated across diverse populations to identify which approaches successfully protect marginalised communities. Fourth, scholarship should examine how communities develop grassroots resistance strategies and vernacular expertise for navigating AI-driven disinformation ecosystems. Finally, critical attention to the political economy of AI development can illuminate how existing power structures shape whose interests algorithmic systems serve and whose harms receive acknowledgement and remediation (Crawford, 2021; Zuboff, 2019). Advancing digital equity in the age of AI-generated disinformation requires sustained interdisciplinary collaboration among computer scientists, communication scholars, social scientists, and, most importantly, the communities experiencing these algorithmic injustices.