Scaling Disinformation with AI: A Threat to the Chemical Weapons Prohibition?

The impact of disinformation on society and politics is growing as a result of advances in AI technology.¹ This technological evolution affects all potential targets of disinformation campaigns. For example, it can undermine trust in states and international institutions, thereby increasing societal instability and polarization. It can also influence political decisions and votes, reinforce particular narratives or interests, and damage the reputations of individuals or institutions. This development is occurring in a social media environment in which false news often spreads further than fact-based information. New technological developments, particularly the emergence of LLM-based AI agents, are further increasing the scalability of disinformation campaigns.² Such AI agents can generate and adapt large volumes of text, images, and conversational interactions in real time. They can also maintain plausible synthetic identities and coordinate activity across multiple platforms. Consequently, they reduce the cost of producing and tailoring disinformation, while being considerably harder to detect than the bot networks used in the past. The development of agent-based AI is evolving so rapidly that its implications cannot yet be predicted. However, potentially disruptive scenarios for democracy are being outlined,³ leading to the pertinent question: “Why invade a country to change its regime when you can just use social media?”—a question that calls for a political response.

While this development poses a threat to all potential targets of disinformation campaigns, its impact on the prohibition of chemical weapons deserves particular attention. The OPCW has repeatedly been the target of disinformation campaigns, and such efforts continue to this day. One of the reasons for this is that the prohibition of chemical weapons is underpinned by strong societal and political norms. Their use therefore creates a need to conceal, deny, reframe, or shift responsibility in order to avoid the high political, legal, and reputational costs associated with violating the prohibition. A closer examination of the opportunities offered by new technologies in this field is therefore essential for the prohibition to be successful.

The Spread of Disinformation on Social Media

Due to a combination of user psychology and bot networks, disinformation can spread more rapidly and widely on social media than fact-based news. Platform algorithms play a significant role in this process by amplifying posts that attract an increased amount of initial attention. Bots can support this early amplification and sustained dissemination, while human users often provide the broader social reach and credibility that make narratives ⁴ Initial reports of an incident also tend to circulate more widely than subsequent reports, making them particularly effective vehicles for disinformation because they require no research or fact-checking. Fact-based information always has a speed disadvantage in this respect.

Recent analyses suggest that bots now generate more than half of all global web traffic.⁵ Despite the enormous volume of bot activity, it remains relatively manageable and is only of major relevance in the early stages of disinformation campaigns. Bot networks are comparatively easy to monitor as they exhibit limited linguistic and narrative diversity, repeatedly using the same talking points. They can also be identified through network and metadata analysis (e.g., unnatural posting rhythms and repost patterns). However, the number of fake accounts has grown to such an extent that manually monitoring them seems almost impossible, helping misinformation to gain increasing traction across many areas of society. For example, Meta reported that it had detected and removed 4.3 billion fake accounts in 2024 alone.⁶

In later stages of disinformation, false information is often unintentionally spread by people who fail to check the facts. The use of emotions such as surprise or anger encourages people to share such non-factual content more readily, whereas facts are usually presented in a neutral tone. These dynamics and amplifying factors pose a direct challenge to democratic processes, as algorithmically amplified false reports systematically distort the political information landscape, undermining public trust in institutions and promoting polarization.

Disinformation and AI Agents

While traditional bots are comparably easy to identify, it is becoming increasingly difficult to detect fake accounts that use AI. For example, AI-generated synthetic images can replace easily traceable stolen profile pictures. The expanding capabilities of AI could lead to a multimodal escalation that would take disinformation campaigns to a whole new level. In addition to creating authentic video and audio data for synthetic identities, agent-based LLM systems pose a particular threat to the online information landscape.⁷ These systems use LLMs as a control center to perform autonomous actions on a computer and can be operated through simple prompts in everyday language. Prompts such as “Build a social media presence and post about current topics (weather, news, typical social activities) from different perspectives, without political bias. Include spelling mistakes, keep it brief, and imitate genuine human behavior” can generate synthetic identities that bear no resemblance to traditional bots. A large number of such synthetic identities can then be mobilized on command to disseminate a political narrative or AI-generated disinformation.

Artificial intelligence agents appear more credible, are harder to detect, and can be more effective than bots. Experimental platforms such as Moltbook, a social network designed for AI agents to post, comment, and interact with one another, illustrate the feasibility of networked agent interaction at scale.⁸ While this does not necessarily reflect actual prevalence, it highlights how AI is rapidly transforming automated data traffic on the “real” internet. By 2025, the number of AI-driven bot cyberattacks, not necessarily related to disinformation, had increased more than tenfold compared to the previous year, bringing the daily average to 25 million. Meanwhile, the number of unreported cases is disproportionately higher, as AI-driven attacks are so difficult to detect.⁹

The OPCW Case

The prohibition of chemical weapons is supported by robust international and societal norms that reject their use. These norms ensure that no state can use chemical weapons without facing significant political consequences. To mitigate these costs, disinformation campaigns have been conducted alongside chemical weapons attacks.¹⁰ The aim of these campaigns was not to question the prohibition of chemical weapons itself, but rather to target three central functions of the chemical weapons ban regime: first, the clarification of responsibility for the use of chemical weapons;¹¹ second, the institutional legitimacy of the OPCW, which such campaigns sought to undermine by portraying it as incompetent or manipulated;¹² and third, its governance structures, which were depicted as biased and subject to the control of individual states.¹³

Thus, disinformation is more than just a theoretical risk; it is a recurring practice in the context of chemical weapons use, as demonstrated by incidents in both Syria and Ukraine. One prominent example was a cross-platform disinformation campaign targeting the Syrian Civil Defense (SCD), commonly known as the White Helmets. In this case, evidence of the use of chemical weapons in Syria recorded by the SCD was portrayed as fabricated. The campaign involved several actors, with official authorities providing political legitimacy and originating the claims, state-affiliated media producing citable material, and non-state actors lending it the appearance of independent validation. The use of social media platforms ensured the reach and repetition of the claims, exploiting disinformation’s speed advantage by circulating the narrative before official investigations were completed. Although this particular campaign occurred several years ago, ongoing disinformation campaigns against the OPCW remain a pressing issue.¹⁴

The enhanced capabilities of disinformation campaigns increase the likelihood that actors can conceal chemical weapons use or mitigate its political consequences, thereby necessitating new institutional responses. New AI models used to disseminate disinformation surpass the capabilities of traditional bots and the methods used to detect them. Detection requires forensic LLM and style analyses, which are still under development and carry the risk of false positives. Technical censorship is not a viable solution, as agent-based LLMs are typically dual-use technologies with legitimate applications across society and the economy. Instead, institutional structures must become more resilient to disinformation campaigns by enhancing transparency around their activities and reporting on emerging methods that could be used to discredit them.

Internally, the OPCW could establish a permanent monitoring, analysis, and response framework to track and address disinformation operations targeting the organization and its verification activities. This framework could also be used to develop guidelines for dealing with disinformation activities aimed at undermining the organization’s credibility. The need for this predates AI, with earlier discussions centering on politically sensitive challenge inspections that carry the risk of discrediting the OPCW, its procedures, and its personnel. The OPCW could also respond with a proactive communication strategy, providing clear and repeated explanations of its methods, mandate, and limitations. At the same time, structures specifically focusing on AI-assisted disinformation would enable the organization to systematically gain operational experience with AI-supported tools, rather than only engaging with AI reactively in the context of disinformation. This would also help to narrow the current “execution gap”—the difference between the opportunities that AI offers for arms control and its still very limited practical application within the regime.

Given the OPCW’s limited resources, addressing these tasks will require multilateral cooperation. One way of achieving this would be to issue states parties with recommendations on developing their capacity to analyze disinformation. This could involve, for example, promoting research into the detection of coordinated AI-enabled account networks, as well as methods for assessing their likely origin, intent, and links to state or non-state actors. These capabilities also offer a valuable opportunity to respond to attacks on other international organizations, thereby strengthening international relations in general. As the situation in Syria has demonstrated, it is crucial to develop an information strategy encompassing all relevant organizations (e.g., the OPCW, the UN and the WHO), and to ensure effective coordination among them. Social media platforms can also contribute positively by introducing mechanisms to label or detect AI “swarms”, particularly in the context of sensitive events (e.g., chemical weapons deployments or OPCW reports), as long as such activity can be reliably identified, for instance with the help of AI-agent providers.

The Chemical Weapons Convention is a treaty that explicitly allows for adaptation in response to new developments. The OPCW is responsible for the scientific analysis of such developments and proposes procedures for harnessing the benefits and addressing the challenges of AI for the Convention, although AI-based disinformation is not yet a focus of its activities.¹⁵ In the face of the rapid evolution of new disinformation capabilities, the Convention and its implementing body are under pressure to demonstrate how disarmament treaties can adapt to new technological developments—and could take a leading role in doing so.