Cyborg Propaganda: Hybrid Human–AI Influence

• Cyborg propaganda is a hybrid form of influence where authentic human identity is combined with algorithmic automation to disseminate strategic narratives.
• It employs architectures like hybrid accounts and distributed human relay models, interleaving automated broadcasting with genuine user engagement.
• Its dynamic oscillation between bot-like and human-like behavior challenges traditional detection systems and raises complex governance issues.

Searching arXiv for papers on cyborg propaganda, computational propaganda, and propaganda detection. Cyborg propaganda denotes propaganda conducted through hybrid human–machine systems in which human identity, judgment, or social legitimacy is combined with automation, algorithmic optimization, or AI-generated articulation. Across contemporary research, the term covers at least three related formations: social-media accounts that alternate between bot-like and human-like behavior for strategic communication (Ng et al., 2024); campaigns in which verified human users ratify and disseminate centrally generated, AI-personalized messages (Kunst et al., 13 Feb 2026); and broader human–AI assemblages that generate, optimize, detect, or modulate propaganda within socio-technical systems (Klincewicz et al., 3 Mar 2025). In this sense, cyborg propaganda is not reducible to classic botnets, nor to ordinary grassroots activism, nor to purely human rhetorical practice. It is characterized by hybridity of agency, time-varying automation, and the strategic exploitation of the ambiguity between authentic identity and synthetic articulation (Ng et al., 2024).

1. Conceptual scope and definitions

The concept of cyborg propaganda is anchored in a broader transition from classical propaganda to computational propaganda. A recent literature review defines computational propaganda as “the use of algorithms, automation and human curation to purposefully distribute misleading information over social media networks to manipulate public opinion, for political polarization etc.” (Pote, 2024). Within that broader class, cyborg formations occupy the space where human and automated behaviors are fused rather than cleanly separated.

One operationalization appears in the study of Twitter strategic communication, where a cyborg account is defined quantitatively as an account whose bot classification fluctuates across time windows: it is detected as a bot on some days and as a human on others (Ng et al., 2024). Let $s_i(t) \in [0,1]$ denote the bot probability assigned to account $i$ on day $t$, with label threshold $\tau = 0.70$. An account is labeled bot when $s_i(t)\ge\tau$ and human otherwise. If $F_i$ is the number of bot/human label flips within a month, and $\overline{D}_i$ is the mean absolute bot-score change across flips, then the implemented criterion is:

$$F_i \ge 3 \quad \text{and} \quad \overline{D}_i \ge 0.10$$

This definition captures frequent and substantial oscillation in bot-likelihood rather than mere noise around a decision threshold (Ng et al., 2024).

A distinct but complementary formulation defines cyborg propaganda as “the synchronized and semi-automated dissemination of algorithmically generated articulations of narratives via a large number of verified human accounts. … Identity is authentic; articulation is synthetic” (Kunst et al., 13 Feb 2026). Here the cyborg is not a temporally unstable account but a human participant whose speech is partly ceded to a central AI-mediated coordination system. The same paper distinguishes four quadrants by crossing identity type with articulation type: bot nets are synthetic identity plus algorithmic articulation; troll farms are synthetic identity plus organic human articulation; grassroots action is verified human identity plus organic articulation; and cyborg propaganda is verified human identity plus algorithmic articulation (Kunst et al., 13 Feb 2026).

Other adjacent literatures widen the term further. “Slopaganda” describes generative-AI-enabled propaganda that operates through flooding, microtargeting, and cognitive hijacking at scale, again as a human–AI assemblage rather than a purely automated system (Klincewicz et al., 3 Mar 2025). Work on social cybersecurity likewise groups “bots, cyborgs, trolls, sock-puppets, and deep fakes” as instruments of technologically mediated influence operations (Mulahuwaish et al., 6 Apr 2025). A plausible synthesis is that cyborg propaganda functions as a middle layer between classical propaganda and fully autonomous influence systems: humans remain in the loop, but machine systems increasingly shape expression, distribution, and adaptation.

2. Historical emergence from computational propaganda

The emergence of cyborg propaganda is inseparable from the digitization of information warfare and the limits of traditional propaganda concepts. Classical propaganda was organized around visible elites, mass media, and recognizable symbolic devices. Computational propaganda research argues that digital media has altered the relevant symbolic infrastructure: automation, recommendation systems, platform metrics, microtargeting, and identity cloaking now function as significant symbols and operational resources in their own right (Pote, 2024).

Earlier work on misinformation and propaganda emphasized the role of search engines, social media, web spam, fake accounts, and “Twitter bombs” in shaping visibility and public belief (Metaxas, 2018). That literature already described a hybrid epistemic condition in which human judgments are filtered through platform ranking and social metrics. Cyborg propaganda extends this condition by making the propagandist itself hybrid. Instead of merely exploiting algorithmic channels, actors integrate automation into the expressive unit—whether an account, a campaign app, or a coordinated swarm of verified users (Kunst et al., 13 Feb 2026).

The strategic contexts in which this evolution becomes visible are high-salience, polarized events. A large-scale longitudinal study analyzes over 3.1 million “consistent agents” in English-language Twitter data from the 2020 Coronavirus pandemic and the 2020 U.S. elections, two environments saturated with strategic communication over vaccination, public-health trust, voter fraud, QAnon, BLM, and related narratives (Ng et al., 2024). These settings are treated as archetypal propaganda battlegrounds where hybrid human–automation techniques are especially adaptive.

Recent theory further argues that the distinction between genuine grassroots activism and automated influence operations is collapsing because AI tools can now monitor sentiment, generate personalized content, and coordinate verified citizens through partisan coordination apps (Kunst et al., 13 Feb 2026). This creates a regulatory gray zone: legal and platform frameworks designed for synthetic botnets are less effective when dissemination is performed by real people, even if the articulation is centrally generated and optimized (Kunst et al., 13 Feb 2026). This suggests that cyborg propaganda is not an anomaly within digital propaganda but a likely successor to earlier bot-centric operations.

3. Architectures of cyborg propaganda

Research identifies at least two core architectures. The first is the hybrid account model: a single social-media profile alternates between automated amplification and human posting, thereby combining scale with authenticity cues (Ng et al., 2024). The second is the distributed human relay model: a central hub uses AI to generate and personalize narratives which verified users then approve and post under their own names (Kunst et al., 13 Feb 2026).

In the hybrid-account model, the technical signature is temporal instability in classifier outputs. The account exhibits regular, tool-driven behavior at some times and organic, conversational behavior at others. BotHunter, the detection system used in the Twitter study, is a supervised random-forest classifier using account metadata and temporal features such as posting periodicity and source client; this is important because third-party tools and regular posting rhythms are markers of automation and cyborg-like behavior (Ng et al., 2024). When users interleave scheduled broadcasting, retweet automation, or campaign-style hashtagging with personally written comments, they produce the oscillatory bot/human pattern formalized above.

In the distributed human relay model, the architecture has four layers: organizer directive, AI monitor, AI multiplier, and verified human users (Kunst et al., 13 Feb 2026). The organizer hub, often a partisan coordination app, ingests signals about public sentiment and issues missions. The AI monitor tracks topics, hashtags, counter-narratives, and performance metrics. The AI multiplier turns a central directive into thousands of personalized messages tailored to each participant’s style and network. Verified human users then ratify and publish these posts. The process is explicitly closed-loop: engagement data feeds back into strategic directives and into the generative system that crafts future messages (Kunst et al., 13 Feb 2026). This architecture preserves human authenticity at the level of account identity while centralizing narrative production.

A third architecture appears in work on generative-AI propaganda, where human ideologues, bureaucratic middle layers, foot soldiers, platform recommenders, and generative models form a cyborg propaganda assemblage (Klincewicz et al., 3 Mar 2025). Here, the machine component does not merely automate repetition. It enables mass personalization, near-zero marginal cost content generation, and rapid adaptation across demographic and psychographic segments. This suggests that cyborg propaganda is increasingly an optimization problem over hybrid systems rather than a simple blend of manual and automated posting.

4. Social-media cyborgs as strategic communicators

The most concrete empirical account of cyborg propaganda on social media comes from the Twitter study of COVID-19 and U.S. election discourse (Ng et al., 2024). The authors analyze two large datasets: 62,072,853 unique users and 355,743,163 tweets in the Coronavirus stream, and 23,933,084 unique users with 193,821,760 tweets in the election stream, from which they derive a longitudinal subset of 3.1 million consistent users (Ng et al., 2024).

The cyborg criterion identifies accounts that both frequently flip bot/human labels and do so with substantial shifts in bot-likelihood. The study reports that bot-to-human and human-to-bot flips are approximately symmetric for cyborgs, indicating continual oscillation rather than monotonic change. Cyborgs also display higher variance in bot scores than stable bots or stable humans (Ng et al., 2024). This suggests that a single-snapshot bot detector is structurally inadequate for hybrid strategic communicators.

Manual annotation of a 1% sample, with Cohen’s $\kappa = 0.87$, shows that 36% of cyborg accounts are activists, 27% are renowned people, 12% are other communication accounts such as marketers and financial promoters, and 25% had been suspended at follow-up (Ng et al., 2024). In aggregate, approximately 63% of cyborgs in that sample are activists or renowned persons, indicating that hybridization is disproportionately associated with overt strategic communication rather than random automation (Ng et al., 2024).

The observed behavior pattern is a mixture of automated broadcasting and human “personal touch.” Examples include retweeting official health guidance, campaign news, or issue-specific messaging using scheduling tools or third-party apps, then interleaving personal opinions, mundane life updates, or direct engagement with followers (Ng et al., 2024). This hybrid style supports three strategic objectives stated in the paper: establishing and maintaining a personal or organizational brand, amplifying messages and campaigns, and participating in polarized issue framing (Ng et al., 2024). The result is a propaganda instrument that remains socially legible as a person while enjoying some of the throughput advantages of automation.

5. Network position, discourse, and collective action

Cyborg propaganda is not merely a content phenomenon; it is also a network-structural one. In the Twitter study, the all-communication networks contain nodes as users and directed edges for retweets, quote tweets, or mentions. Degree centrality, betweenness centrality, eigenvector centrality, follower counts, friend counts, and retweets received are used to characterize cyborgs (Ng et al., 2024). Cyborg accounts are found to have higher degree centrality and betweenness centrality than non-cyborgs, but lower eigenvector centrality, with all differences statistically significant at $p<0.001$ (Ng et al., 2024). They also have more friends and somewhat more followers on average, while 0% of cyborg accounts are verified and 100% of verified accounts fall into the non-cyborg category (Ng et al., 2024).

This profile implies tactical bridging positions rather than top-level hub status. Cyborgs are well situated to broker information across communities, inject narratives into local neighborhoods, and amplify issues without being the most globally central actors (Ng et al., 2024). A plausible implication is that cyborg propaganda is structurally optimized for mid-level brokerage rather than overt prominence.

Content analysis reinforces the strategic-communication interpretation. The Twitter study uses manual hashtag labeling and a stance propagation algorithm on user–hashtag bipartite graphs to assign stances such as pro-vaccine vs anti-vaccine and conservative vs liberal (Ng et al., 2024). Cyborgs appear in all stances; they are not tied to a single ideology. Topic modeling with LDA shows that within each stance, cyborgs and non-cyborgs discuss broadly similar themes, but cyborgs often emphasize campaign-like slogans, branded hashtags, and consistent narratives (Ng et al., 2024). This is compatible with propaganda as organized message amplification rather than as a specific ideology.

At the collective-action level, theory proposes a “collective action paradox” of cyborg propaganda (Kunst et al., 13 Feb 2026). On one side, AI-assisted coordination can “unionize influence,” pooling the reach of ordinary citizens who would otherwise remain algorithmically invisible. On the other, it can reduce those citizens to “cognitive proxies,” participants who retain consent but surrender articulation to a central system (Kunst et al., 13 Feb 2026). This reframes cyborg propaganda as neither simply fake activism nor straightforwardly emancipatory mobilization. It is an architecture that can either lower coordination costs for marginalized actors or transform distributed citizens into human relays for top-down narrative logistics.

6. Detection, analysis, and mitigation

Cyborg propaganda is difficult to detect because most existing systems assume stable distinctions between bots and humans, or between propagandistic and non-propagandistic language. The Twitter cyborg study explicitly argues that standard bot-detection pipelines are poorly suited to hybrid accounts because they rely on static labeled datasets and single snapshots of account features (Ng et al., 2024). The proposed remedy is longitudinal analysis of bot-score trajectories, using $i$0, $i$1, and the variance of $i$2 as core signals (Ng et al., 2024).

The broader literature on social cybersecurity generalizes this problem. A survey notes that current detection systems tend to classify single accounts rather than coordinated campaigns, depend heavily on platform-specific data structures, and remain largely Twitter-centric (Pote, 2024). The same review emphasizes that hybridization between automated and human-driven behaviors weakens distinctions between bots and humans, and recommends greater attention to coordination detection, cross-platform activity, and image-based features (Pote, 2024). Social cybersecurity surveys similarly frame cyborgs as part of a wider landscape of attack vectors requiring dynamic network analysis, NLP, graph models, and agent-based simulation (Mulahuwaish et al., 6 Apr 2025).

Propaganda detection itself presents another layer of difficulty. Research on BERT for sentence-level propaganda classification shows that strong benchmark performance can still reflect narrow lexical biases. A fine-tuned BERT-large ensemble achieved F1 = 0.62 on the shared-task test set and 0.66 on the development set, ranking third of 25 teams, but error analysis showed that the model often misclassifies opinion pieces as propaganda and cannot distinguish quotations of propaganda from actual use of propaganda techniques (Hua, 2019). The model’s implicit prototype of propaganda is dominated by emotionally charged terms such as “devastating,” “horrific,” and “disgusting” (Hua, 2019). This indicates that automated propaganda detectors embedded in cyborg systems can themselves become sources of systematic bias.

More recent work on LLMs extends this concern from detection to generation. When prompted explicitly for propaganda-style messaging, GPT-4o and Mistral Small 3 produced outputs classified as propaganda at 99%, and Llama 3.1-Instruct 8B at 77% (Jose et al., 4 Mar 2026). These outputs relied heavily on Loaded Language, Exaggeration/Minimization, Flag-Waving, and Appeal to Fear (Jose et al., 4 Mar 2026). The same study shows that mitigation via Supervised Fine-Tuning, Direct Preference Optimization, and ORPO can substantially reduce propagandistic generation, with ORPO performing best: on a held-out development set, an unaligned Llama 3.1 model produced propaganda on 77% of prompts, while ORPO reduced that to 10% (Jose et al., 4 Mar 2026). This suggests that cyborg propaganda defense may require both aligned generation models and downstream detectors, rather than relying on either alone.

7. Longevity, governance, and contested implications

One of the most consequential empirical findings about cyborg propaganda concerns persistence. In the Twitter study, one year after data collection, suspension rates for bots, cyborgs, and humans differ sharply. In the Coronavirus dataset, 89.2% of bots were suspended, compared with 56.0% of cyborgs and 19.5% of humans; among surviving accounts, average ages were $i$3 days for bots, $i$4 days for cyborgs, and $i$5 days for humans, with ANOVA $i$6 (Ng et al., 2024). In the U.S. election dataset, 76.9% of bots, 49.2% of cyborgs, and 19.9% of humans were suspended; surviving cyborgs again had the longest account age at $i$7 days, compared with $i$8 for bots and $i$9 for humans, with ANOVA $t$0 (Ng et al., 2024). Cyborgs are thus less likely to be suspended than bots and are, among non-suspended accounts, the longest-lived category.

The paper offers two explanations: cyborgs evade bot detectors because their behavior is only intermittently bot-like, and platforms may exhibit “graciousness” or tolerance toward prominent activists or officials whose behavior appears legitimate or politically sensitive (Ng et al., 2024). A plausible implication is that platform governance regimes structurally favor hybrid propagandists over overt botnets, thereby encouraging a shift toward cyborg tactics.

The governance problem becomes sharper in the verified-human architecture of cyborg propaganda. Because the content is AI-authored but posted by humans, it occupies a “human-in-the-loop loophole” relative to regulations and moderation rules focused on fake or automated accounts (Kunst et al., 13 Feb 2026). Proposed governance responses include targeting coordination hubs rather than individual users, treating campaign apps as regulated intermediaries, mandating disclosure of AI-assisted or campaign-coordinated political content, and shifting the legal and analytical focus from bot-vs-human to coordinated-vs-organic activity (Kunst et al., 13 Feb 2026).

Normative evaluations remain contested. Some work presents AI-augmented subjectivity itself as a desirable horizon. The “c(ai)borg” literature advocates a transparent, literate, ethically reflective human–AI hybrid who moves “from guilt to empowerment through transparency and skill-building” (Aal et al., 30 May 2025). That paper is not about propaganda operations, but it demonstrates how cyborg discourse can normalize hybridization by reframing dependence on AI as growth and competence rather than loss (Aal et al., 30 May 2025). A plausible implication is that propaganda about cyborgization and propaganda by cyborg systems can reinforce each other: one normalizes the socio-technical subject, the other weaponizes it.

At the farthest edge of the literature, work on augmented humans, neuromorphic interfaces, and closed-loop neurocybernetic systems shows how future cyborg propaganda could extend beyond social-media articulation into direct modulation of perception, attention, and cognition (Kott et al., 2015, Talanov et al., 2022). The military information-warfare perspective already treats augmented humans, intelligent systems, and the information domain as deeply entangled vulnerabilities (Kott et al., 2015). The neuropunk perspective describes architectures for closed-loop integration of biological tissue, BCI, and neuromorphic middleware that could, in principle, support real-time cognitive steering (Talanov et al., 2022). These papers do not analyze propaganda in the ordinary communication-studies sense, but they suggest that the cyborg in cyborg propaganda may eventually cease to be a metaphor for mixed posting behavior and become a literal substrate of influence.

Cyborg propaganda therefore names a transition in propaganda’s unit of operation. The relevant actor is no longer simply the propagandist, the bot, or the media outlet, but a hybrid system in which authenticity, automation, narrative, and optimization are unevenly distributed across humans and machines. Empirical research shows that such systems are already tactically positioned, strategically useful, and more persistent than pure bots (Ng et al., 2024). Theoretical work suggests that they blur the boundary between mobilization and manipulation, exploit regulatory gray zones, and may become increasingly central to political communication as AI systems gain more adaptive and personalized expressive capacity (Kunst et al., 13 Feb 2026, Klincewicz et al., 3 Mar 2025).