Source-linked AI summary
Automated Disinformation and Malicious AI Swarms: Risks for Democracy and Development in Africa
Daniel Thilo Schroeder, Philipp M. Lutscher, Samba Dialimpa Badji, Stefan Brenner, Wenchao Dong, Tsehaye Haidemariam, Saheed Bidemi Ibrahim, Amanuel Tesfaye Kebede, Jonas R. Kunst, Johannes Langguth, Pauline Lemaire, Mulatu Alemayehu Moges, Lukasz Olejnik, Kristin Skare Orgeret, Gerald Walulya
TL;DR
The paper examines how emerging AI-enabled influence could interact with fragile and contested information environments in Africa, where autonomous swarms are not yet established at scale. It develops an operational framework and considers Mali and Ethiopia, concluding that hybrid human–AI operations may be the more immediate risk, with linguistic asymmetries constraining both influence and defense.
Problem
Method
The paper defines swarm capabilities and thresholds, then uses a forward-looking risk analysis of African contexts, including Mali and Ethiopia.
Source: §4 · p. 9, §11 · p. 19Results
Hybrid operations combining AI-enabled scale and adaptation with human cultural knowledge, credibility, and judgement may pose the most plausible near-term threat.
Source: §11 · pp. 19–20Takeaways & Limitations
Proposed swarm influence could fabricate apparent consensus and weaken trust in public institutions, threatening democratic participation and governance.
Source: §7 · p. 15Fully autonomous swarms are not proven to be widely deployed, and uneven African-language performance may constrain offensive capabilities while weakening defenses.
Source: §4 · pp. 8–9
Abstract
from arXiv · showhide
Generative artificial intelligence is reshaping how information is produced, accessed, and circulated, while enabling disinformation campaigns of increasing scale and sophistication. There is currently no clear evidence that fully autonomous AI swarms conduct influence operations at scale, but their enabling capabilities are advancing. We define malicious AI swarms as coordinated, persistent, and adaptive multi-agent systems designed for influence operations, distinguishing them from AI-assisted content production and centrally managed synthetic personas. We examine their implications for hybrid regimes and conflict-affected states in Africa, where institutional constraints and fragile media environments may heighten vulnerability. Drawing on Mali and Ethiopia, we consider how automated influence could infiltrate communities, fabricate consensus, and erode trust in governance and development. The cases illustrate different configurations of state and non-state influence: competing actors in Mali's fragmented information environment, and more organized state-led strategies of narrative management in Ethiopia. African-language and training-data asymmetries may constrain influence capabilities while weakening defensive responses. Hybrid human-AI operations could combine automated scale and adaptation with local knowledge and credibility. This forward-looking risk analysis develops a scenario of increasingly accessible AI-driven coordination, rather than claiming that autonomous swarms are already operating at scale in Africa. We propose a layered governance approach linking technical safeguards to platform accountability, civic institutions, and regional coordination to protect democratic participation, peacebuilding, and development.
1 Introduction: From Bots to Swarms in the Global South
The paper frames malicious AI swarms as an emerging, not yet established-at-scale, risk and develops an Africa-focused analysis of how they could interact with contested information environments. Mali and Ethiopia illustrate distinct pathways, while the paper connects these risks to democratic participation, peacebuilding, and development.
1 Introduction: From Bots to Swarms in the Global South
- Fully autonomous swarms remain an emerging threat, although generative text, agentic planning, and multimodal synthesis already provide potential enabling components.The paper distinguishes adaptive multi-agent campaigns from AI-assisted content production and centrally managed synthetic personas.Source: §1 · p. 2
- Hybrid human–AI operations may be the nearer-term concern, combining human contextual knowledge, credibility, and judgement with AI-enabled scale, translation, and message variation.The use of coordinated real-world campaigns remains untested in the cited experimental work.Source: §1 · p. 4
- Automated disinformation could distort public deliberation, undermine peacebuilding, and weaken institutional foundations of democratic development.The paper links these risks to fragile institutions and uneven digital infrastructure, while noting that partial connectivity limits immediate direct reach.Source: §1 · pp. 3–4
- The paper addresses limited African coverage in English-language research by offering a forward-looking analysis of how malicious AI swarms could operate in African political contexts.It integrates computational social science, media studies, political science, ethnographic research, and AI governance.Source: §1 · pp. 4–5
- Mali and Ethiopia share conflict, contested legitimacy, polarized publics, diaspora participation, and constrained media, but differ in influence organization and state capacity.Mali represents a fragmented information environment with competing actors; Ethiopia illustrates more organized state-led narrative management.Source: §1 · pp. 3, 5
2 Digital Ecosystems and Democratic Fragility in Africa
African digital ecosystems combine expanding platform use with uneven connectivity, offline information channels, and limited oversight. These conditions shape where manipulated narratives may travel and how they can affect democratic deliberation.
2 Digital Ecosystems and Democratic Fragility in Africa
- Mobile internet and social platforms have become important information and debate channels, but radio and interpersonal networks continue to mediate circulation offline.Access remains uneven, particularly for rural and lower-income communities.Source: §2 · p. 5
- Uneven connectivity may concentrate fabricated consensus among connected agenda-setters, whose online narratives can reach less-connected people through offline intermediaries.This can lead observers to mistake synthetic consensus for national opinion, while online corrections may not follow falsehoods across the digital divide.Source: §2 · pp. 5–6
- Digital publics are shaped by diaspora participation and political alignments, which can support pluralism while also amplifying polarized narratives or external agendas.The cited Ethiopia analysis identifies hostile Facebook rhetoric that may reinforce ethnopolitical polarization.Source: §2 · p. 6
- Pressure on fact-checkers, unreliable-source access, and weak truth-telling institutions can make coordinated manipulation harder to identify and counter.The paper connects democratic deliberation to the integrity of digital communication ecosystems.Source: §2 · p. 6
3 Citizen Journalism and the Weaponization of Social Media
Citizen journalism broadens participation and reporting where professional media are constrained, but decentralized, emotionally charged communication can also weaken verification and facilitate manipulation. In conflict settings, these dynamics connect online discourse to mobilization and contested narratives.
3 Citizen Journalism and the Weaponization of Social Media
- Citizen journalism enables people to document events, verify claims, and circulate political narratives where press freedom, media institutions, or geographic access to reporting are limited.This bottom-up information production has been viewed as a democratizing force and a channel for marginalized participation.Source: §3 · p. 7
- In the Malian Twittersphere, authority is negotiated through institutional position, social connections, follower networks, and perceived authenticity rather than a strict journalist–audience divide.The network includes diverse producers interpreting and reframing political conflict, alongside concentrations of state-aligned influence.Source: §3 · p. 7
- Decentralized narrative production expands access but dilutes verification norms and editorial accountability, allowing emotionally resonant claims to take hold without institutional validation.Identity alignment and viral affordances can outweigh evidentiary credibility in what becomes socially accepted as truth.Source: §3 · p. 7
- Conflict-affected platforms can serve as networked-conflict infrastructure, where rumors and calls to action circulate before independent verification and hashtags can carry dehumanizing rhetoric.The cited Tigray-conflict research describes emotionally charged “hatetags.”Source: §3 · p. 7
- Political elites, armed groups, foreign actors, and commercial networks can seed narratives through trusted local voices or mimic grassroots reporting, making coordinated influence difficult to distinguish from civic participation.The same infrastructures that enable democratic expression can be weaponized to intensify polarization and undermine trust.Source: §3 · pp. 7–8
4 From Botnets to Malicious AI Swarms
The paper defines malicious AI swarms by coordinated, persistent, adaptive multi-agent behavior rather than content volume alone, and distinguishes them from related forms of AI-enabled influence. It argues that hybrid human–AI operations are a nearer-term concern, while language and data asymmetries constrain both offensive and defensive capabilities.
4 From Botnets to Malicious AI Swarms
- Fully autonomous, self-reinforcing swarm networks are not yet proven to be widely deployed, although component capabilities are emerging and barriers to experimentation are falling.Small locally run models can generate short persona-conditioned political messages, but the experiments do not establish the cost of a complete influence campaign.Source: §4 · pp. 8, 10, 11
- Malicious AI swarms are distinguished by multiplicity, coordination, persistence, adaptation, and collective agency, not by scale or generative content alone.The five dimensions can occur to different degrees rather than as binary properties.Source: §4 · p. 9
- High-volume AI content, fixed-schedule reposting accounts, and human-directed use of generative AI do not alone meet the swarm threshold.Those cases lack the adaptive, persistent, and collectively organized characteristics associated with swarm-like behavior.Source: §4 · p. 9
- The capability ladder ranges from human-operated AI content production and synthetic personas to adaptive multi-agent campaigns and goal-revising swarms.The upper levels involve reduced human supervision and, at the highest level, agents revising intermediate objectives and reallocating tasks.Source: §4 · p. 9
- African-language performance gaps may constrain autonomous campaigns while also weakening moderation, fact-checking, attribution, and coordinated-behavior detection.Code-switching, translation, multimodal content, or local human intermediaries may help attackers circumvent model limitations.Source: §4 · pp. 9–10
- Training-data imbalance and limited transparency can produce factual or cultural errors in synthetic personas and defensive systems alike.The paper notes that these deficiencies can make personas easier to identify while also generating false positives or missing locally meaningful expressions and narratives.Source: §4 · p. 10
- Hybrid or “cyborg” propaganda may be a more immediate risk than autonomous swarms, with AI generating and testing material while humans supply judgement, language competence, credibility, and access.Locally embedded networks remain important to effective operations.Source: §4 · pp. 10–11
- A plausible swarm scenario involves infiltrating salient communities, sustaining conversations, adapting strategies, and tailoring messages to social identities over time.The proposed risk is intensified where weak verification, decentralized authority, and fragile trust enable manufactured consensus to appear locally grounded.Source: §4 · p. 11
5 Case Study 1: Mali as a Microcosm of Emerging Threats
Mali’s conflict-affected information environment combines weak journalistic gatekeeping, citizen reporting, and competing political narratives. The case illustrates how future AI-enabled influence could exploit these conditions, although key mechanisms remain untested and Mali is not representative of Africa.
5 Case Study 1: Mali as a Microcosm of Emerging Threats
- Approximately 68% of surveyed journalists and media professionals reported censorship concerns or self-censorship, especially on security and justice issues.In this environment, claims may gain authority through circulation and group alignment before independent verification.Source: §5 · p. 12
- Citizen reporting offers local voice and transparency, but conflict updates often rely on unconfirmed accounts, unverifiable images, and politically shaped narratives.State forces, armed opponents, and foreign actors already use connected networks to advance competing accounts rooted in themes such as sovereignty and military legitimacy.Source: §5 · p. 12
- Potential AI swarm-like systems could use social networks, diaspora channels, and citizen-journalism styles, but concealment and ease of imitation remain untested.The authors expect synthetic participation to become harder to distinguish from grassroots discourse and easier to normalize.Source: §5 · p. 12
- Under automated manipulation, civic and community-reporting infrastructures could accelerate misinformation, factional identities, and political violence, weakening democratic reconstruction and peace negotiations.Mali is a most-likely case because of its conflict, digitally networked communication, and exposure to foreign influence; it is not representative of Africa.Source: §5 · p. 12
6 Case Study 2: Ethiopia as a Laboratory of Authoritarian Information Adaptation
Ethiopia illustrates how coordinated digital influence may operate through organized state and pro-government networks as well as in a multilingual, conflict-affected environment. The case highlights how emerging AI capabilities could augment existing narrative-management practices, while human actors remain important for cultural credibility and audience knowledge.
6 Case Study 2: Ethiopia as a Laboratory of Authoritarian Information Adaptation
- Ethiopia complements Mali by showing how influence can be embedded in state strategies of narrative management, regime legitimation, and political mobilization.State and pro-government actors have demonstrated capacity to organize domestic and diasporic participation in digital campaigns.Source: §6 · pp. 12–13
- During conflict, Ethiopian online discourse has included hostile, identity-based hashtags and rhetoric in English- and Amharic-language Facebook discussions.Orgeret et al. (2026) describe “hatetags” during the Tigray conflict; Gessesse and Moges (2026) identify sectarian and hostile rhetoric in Amhara conflict discourse in September–December 2023.Source: §6 · p. 13
- The government and supporters used Twitter during the 2020–2022 civil war to promote strategic narratives to domestic and international audiences.The campaign depicted the Tigray People’s Liberation Front as an existential threat, framed Western criticism as neo-colonial intervention, and mobilized diaspora supporters to amplify state-aligned narratives.Source: §6 · p. 13
- Information management has expanded beyond censorship and shutdowns to include a “media army,” paid influencers, coordinated messaging, and agenda-setting.These strategies seek to dominate attention and structure debate around state-established agendas.Source: §6 · p. 13
- Existing networks of officials, activists, influencers, diaspora supporters, and state media could provide an organizational base for hybrid AI influence operations.AI could lower costs, accelerate message variation, and segment communication, while human networks continue to provide role differentiation, audience knowledge, and adaptation to political events.Source: §6 · p. 13
- AI-generated images, video, voice, and music promoting Ethiopia’s Red Sea ambitions have circulated on X and TikTok, with some amplification by pro-government accounts.The material uses nationalist narratives; the authors suggest generative AI may lower the cost of emotionally evocative propaganda for domestic legitimation and foreign-policy objectives.Source: §6 · pp. 13–14
- Government communication has also co-opted fact-checking language, raising the possibility that future influence could mimic verification’s authority signals.Uneven performance across African languages may leave culturally persuasive communication dependent on human linguistic and contextual knowledge; models may supply scale, translation, synthetic media, and rapid variation.Source: §6 · p. 14
- Ethiopia illustrates a pathway in which AI-enabled coordination could strengthen regimes’ ability to manufacture support, crowd out dissent, and advance political objectives.The political effects depend not only on technical capability but also on who controls surrounding organizational and institutional infrastructure.Source: §6 · p. 14
7 Pathways of Harms to Democracy in Africa
The paper describes potential democratic harms arising from the interaction of automated influence with existing political, institutional, and communicative conditions. Proposed pathways include fabricated consensus, weakened institutional trust, polarization, distraction, and delegitimation of electoral or transitional processes.
7 Pathways of Harms to Democracy in Africa
- Converging synthetic accounts could fabricate apparent consensus and shift perceptions of which views are common, acceptable, or politically viable.Perceived agreement may persuade where political opinion is influenced by social-network and community-identity cues.Source: §7 · p. 15
- Disinformation targeting grievances could deepen mistrust in public institutions and weaken the authority of public decision-making.The text also notes that regimes may use social media and AI-generated materials to manufacture legitimacy and deter political engagement.Source: §7 · p. 15
- Amplifying narratives of betrayal, victimhood, or historical trauma may normalize hostile rhetoric and harden group boundaries in societies divided by identity.Under existing conflict, these dynamics could narrow deliberation or compromise and contribute to escalation.Source: §7 · p. 15
- Flooding channels with distracting claims or commentary can make reliable information harder to identify and drive participation decline through uncertainty and fatigue.The text distinguishes this pathway from direct silencing: people may withdraw when deciding what to trust becomes difficult.Source: §7 · p. 15
- During elections and transitions, adaptive messaging could mobilize some groups, discourage others, distort perceived fairness, or delegitimize outcomes before they occur.The authors warn that perceived inevitability of manipulation could undermine elections and constitutional negotiations.Source: §7 · p. 15
- Automated influence could exploit both political divisions and civic engagement, gradually reshaping everyday interpretations of truth, credibility, and public representation.The authors describe these shifts as potentially cumulative, with little room for recovery once entrenched in fragile contexts.Source: §7 · pp. 14–16
8 Governance and Safeguards
The proposed response links technical safeguards with platform accountability, civic resilience, and regional coordination because no single intervention can eliminate the threat. It aims to make manipulation harder while protecting political speech and authentic civic communication.
8 Governance and Safeguards
- Safeguards should span technical, institutional, civic, and transnational layers because influence campaigns exploit weaknesses across the information ecosystem.The authors caution that moderation must balance detection with political speech, transparency, accountability, and informational justice.Source: §8 · p. 16
- Platforms and researchers could detect coordinated inauthentic behavior through organization, timing, narrative alignment, and posting patterns rather than adjudicating truth alone.The proposal includes transparent audits, independent access to platform data, and regionally relevant standards, while warning against enabling repression or uneven enforcement.Source: §8 · p. 16
- Provenance and identity systems could constrain synthetic-account fleets, but should allow verified pseudonymity and protect speakers from unmasking.The text notes that anonymity can protect activists from identification and retaliation in repressive contexts.Source: §8 · p. 16
- Civic resilience can be supported through media literacy, local fact-checking, community radio, and verification resources translated into local languages.The aim is to sustain baseline epistemic trust and dialogic openness, not to make citizens detect every false statement.Source: §8 · p. 16
- Regional bodies could extend existing election, mediation, and cybersecurity coordination to shared monitoring and early warning for AI-enabled influence.The proposed networks would prioritize elections and political transitions while remaining transparent, rights-respecting, and accountable.Source: §8 · pp. 16–17
- Governance should shape automation rather than prohibit it, making manipulation costlier and easier to detect while making authentic civic communication safer and more durable.The paper frames democracy protection as reinforcing social and institutional infrastructures of trust before swarm capabilities mature.Source: §8 · p. 17
9 Towards Resilient Information Ecosystems for Democracy and Development
Resilience means sustaining democratic deliberation within plural, contested public spheres, not restoring a supposedly stable media past. It requires support for local information infrastructures and links between information integrity, accountable governance, and inclusive development.
9 Towards Resilient Information Ecosystems for Democracy and Development
- Resilience is the capacity to absorb disagreement, evaluate competing claims, and maintain trust in collective decision-making.The goal is to strengthen deliberation within plural public spheres, not suppress complexity or restore past media hierarchies.Source: §9 · p. 17
- Local journalists, community radio, vernacular media, fact-checkers, and civic educators need safety, legal protection, and viable independent funding.These actors provide continuity and memory that algorithmic content streams cannot yet replace.Source: §9 · p. 17
- Media-literacy programmes should reflect local conditions and existing community practices rather than treat communities as passive misinformation victims.Evidence from Kenya and Nigeria shows that accuracy prompts can reduce intentions to share COVID-19 misinformation.Source: §9 · p. 17
- Federated platforms and community-governed data infrastructures could give local institutions more power to set norms and moderation practices.These alternatives do not eliminate manipulation risks but may better align information systems with regional political cultures and development priorities.Source: §9 · p. 18
- Independent media and alternative platform governance may reduce incentives for African states to disrupt influence operations through shutdowns or platform blocks.The passage connects these measures to concerns about network disruptions, cyberattacks, and platform blocking.Source: §9 · p. 18
- Digital resilience depends on material governance, including responsive public services, accountable institutions, fair elections, and inclusive development planning.Information integrity is framed as a condition of sociopolitical life, not merely a technical problem.Source: §9 · p. 18
- The broader aim is to sustain coexistence and collective sense-making amid contestation, plurality, and emerging automated disinformation.The paper argues that these conditions should be shaped before automated disinformation capabilities fully mature.Source: §9 · p. 18
10 Research Agenda: Operationalising AI Swarms and Testing Their Effects
The research agenda treats malicious AI swarms as a graded, empirically testable phenomenon, assessed across accounts, networks, systems, and human–AI arrangements. It proposes simulations and multilingual validation to examine when coordinated synthetic participation becomes influential.
10 Research Agenda: Operationalising AI Swarms and Testing Their Effects
- Swarm-like influence should be measured across multiplicity, persistent identities or roles, shared objectives, feedback adaptation, and collective task allocation.These properties should be treated as dimensions rather than binary conditions because systems may display different degrees of organization.Source: §10 · p. 18
- Assessment should combine account-level continuity and interaction histories with network coordination and system-level adaptation, information sharing, and task redistribution.Revision of campaign goals would indicate more advanced autonomy but is not required.Source: §10 · pp. 18–19
- Researchers should trace how labour, information, and decisions are distributed across automated agents and human strategists, specialists, influencers, and operators.Human editing, shared narrative templates, and labour shifts after platform enforcement can help distinguish semi-autonomous swarms from human-controlled AI-assisted campaigns.Source: §10 · p. 19
- Simulations could compare centrally coordinated accounts, adaptive multi-agent systems, and hybrid campaigns across language competence, human involvement, institutional trust, corrective exposure, and platform design.Outcomes include perceived consensus, polarisation, correction reach, institutional trust, political participation, and transmission of narratives into offline communication.Source: §10 · p. 19
- Models and detection methods should be evaluated by language, dialect, and register, with local speakers assessing idiom, code-switching, historical references, and coded expression.Without such validation, estimates of influence capabilities and defensive performance may be unreliable.Source: §10 · p. 19
11 Conclusion
The conclusion frames malicious AI swarms as a plausible near-future risk shaped by political conditions, not technology alone. Its cases suggest distinct pathways of influence, while its proposed response joins technical and platform measures to civic institutions, peacebuilding, and development.
11 Conclusion
- The risk of sustained, adaptive synthetic influence arises through interaction between AI capabilities and contested legitimacy, uneven media infrastructures, polarized publics, and actor capacity.The paper presents malicious AI swarms as a plausible near-future escalation, not a technology-only threat.Source: §11 · p. 19
- Mali's fragmented information environment and Ethiopia's more organized state-led strategies illustrate distinct pathways for AI-enabled influence.The cases show potential risks both where information environments are weakly governed and where political actors can organize and direct digital influence.Source: §11 · p. 19
- Language and training-data asymmetries may constrain fully autonomous influence while also weakening moderation, detection, and fact-checking.The most plausible near-term threat may be hybrid, combining AI scale and adaptation with human cultural knowledge, credibility, and strategic judgment.Source: §11 · p. 20
- A layered response links provenance, platform transparency, independent local media, contextual media literacy, and regional monitoring while protecting expression and anonymity.The paper proposes governance across data, model, agent, platform, civic, and regional layers.Source: §11 · p. 20
- Resilience also depends on democratic institutions, peacebuilding, and development frameworks, while future research should operationalize persistence, coordination, adaptation, task allocation, and human–AI interaction.The paper argues that African governments and partners have an opportunity to shape digital ecosystems before swarm capabilities become widespread.Source: §11 · p. 20