1 Introduction

Disinformation has become one of the most pressing global security challenges of the twenty-first century. Authoritarian regimes, in particular, have embraced the power of digital technologies to distort reality, suppress dissent, and manipulate domestic and international opinion. Among these regimes, the Iranian regime has stood out for its coordinated, large-scale deployment of what many analysts call a “cyber army.” This network of state-linked accounts, troll farms, and digital operatives operates across major social platforms, actively spreading propaganda narratives, harassing opposition groups, and attempting to influence global discourse.

Recent advances in artificial intelligence (AI) have added a new layer of complexity to this phenomenon. On one hand, AI enables researchers and platforms to predict and mitigate malicious behavior. On the other hand, AI-powered tools, particularly large language models (LLMs) have been co-opted to make disinformation more persuasive and harder to detect. A close examination of three IEEE Spectrum papers on reinforcement learning from human feedback (RLHF), malicious behavior prediction, and AI missteps in global security helps illuminate how Iran’s disinformation strategies intersect with emerging AI challenges.

2 Iran’s Cyber Army on Social Media

The Iranian regime has invested heavily in digital propaganda as a way to control the narrative at home and abroad. Twitter (now X) disclosures and subsequent academic studies have revealed thousands of accounts linked to Iranian state entities. These accounts push pro-regime messaging, amplify official positions, and attack opposition voices such as exiled activists, women’s rights advocates, and democratic reformers.

According to an IEEE Spectrum report on malicious behavior prediction, one dataset examined included 1,666 accounts tied to the Iranian government in 2019, which consistently produced content aligned with Iran’s diplomatic and strategic perspectives on world events (Hampson, 2025). Iran has been observed to employ a mix of positive and negative tones alternating between portraying the regime as a stabilising force and vilifying its adversaries. This hybrid strategy makes the propaganda appear less overtly hostile, while still undermining opposition groups and shaping perceptions in regime’s favour.

3 The Role of Reinforcement Learning From Human Feedback (RLHF)

AI systems themselves are not immune to manipulation. The paper by Edd Gent in IEEE Spectrum underscores how RLHF a widely used method to fine-tune language models can inadvertently make AI outputs more aligned with user satisfaction than with factual accuracy (Gent, 2025).

For authoritarian regimes like Iran, this creates fertile ground. If LLMs optimised through RLHF are more likely to generate “pleasing” but less truthful content, propagandists can exploit them to mass-produce narratives that sound credible, emotionally engaging, and persuasive, even if they are misleading. The Princeton researchers Jaime Fernández Fisac and Kaiqu Liang demonstrated that RLHF nearly doubles the “bullshit index” a measure of a model’s indifference toward truth while simultaneously boosting user satisfaction by 48 percent.

In practice, this means Iranian cyber operatives can lean on AI tools to craft propaganda that resonates with audiences, all while eroding the boundary between fact and fiction. This not only complicates the work of activists who try to debunk falsehoods, but also places democracies at a disadvantage in the information battlefield.

4 Predicting and Countering Malicious Behavior

Fortunately, AI can also be wielded against disinformation. The European research team profiled by Michelle Hampson developed a model that incorporates the timing and dynamics of user interactions to predict which accounts are likely to become malicious in the future (Hampson, 2025). Tested against datasets from China, Iran, and Russia, the model achieved notable success: in the Iranian dataset, it was able to predict 75 percent of malicious accounts after only 40 percent of their interactions.

This suggests that, despite Iran’s sophisticated tactics, patterns of behavior can still reveal malicious intent. If widely deployed, such predictive systems could allow platforms to preempt disinformation campaigns disabling or flagging coordinated Iranian accounts before they reach critical mass. However, the uneven results across datasets also highlight the challenge: disinformation strategies are adaptive, context-specific, and constantly evolving.

5 Disinformation as a Global Security Threat

The broader implications are explored in the guest essay “AI Missteps Could Unravel Global Peace and Security.” Here, the authors argue that AI misuses such as chatbots driving disinformation threaten not only social cohesion but also international stability (Areios & co., 2025). They stress that the AI community must broaden its engagement with policymakers, civil society, and international institutions to mitigate these threats.

For Iran, whose cyber army already weaponises disinformation, AI represents both an amplifier and a shield. The regime can employ AI to refine its propaganda strategies, making them harder to detect and more persuasive. At the same time, global initiatives like the UN High-Level Advisory Body on AI or the G7 Hiroshima AI Process can provide frameworks to hold such regimes accountable and promote responsible use of technology.

Central to this vision is education. As the authors argue, sustainable governance comes not only from regulation but also from equipping developers with the interdisciplinary knowledge to recognise and mitigate AI’s societal risks. For countering Iranian propaganda, this means training technologists to understand not just technical vulnerabilities, but also the geopolitical contexts in which disinformation thrives.

6 Conclusion

The Iranian regime’s cyber army demonstrates how authoritarian states leverage social media to suppress opposition and promote state narratives. With the advent of AI, these tactics are becoming more sophisticated. RLHF-fine-tuned models, while designed to enhance user satisfaction, can inadvertently enable the production of persuasive yet untruthful propaganda. At the same time, predictive AI systems show promise in detecting malicious behavior early, offering hope that disinformation campaigns can be disrupted before causing harm.

Yet, technology alone is insufficient. Countering disinformation requires a broader societal effort: interdisciplinary education for AI practitioners, collaboration between technologists and policymakers, and global governance initiatives. Only through such collective action can the international community safeguard truth in the digital era and resist the corrosive influence of disinformation campaigns driven by regimes like Iran.

7 References

Gent, E. (2025, August 12). “‘Bullshit Index’ Tracks AI Misinformation.” IEEE Spectrum. https://spectrum.ieee.org/ai-misinformation-llm-bullshit

Hampson, M. (2025, July 25). “Predicting Malicious Behavior on X Before It Happens.” IEEE Spectrum. https://spectrum.ieee.org/fight-misinformation

Areios, K. et al. (2025, July 16). “AI Missteps Could Unravel Global Peace and Security.” IEEE Spectrum. https://spectrum.ieee.org/ai-missteps-unravel-global-security