In the unassuming floodplains of Gaibandha, a northern district of Bangladesh known for its rivers and seasonal floods, a single individual is alleged to have quietly built a propaganda machine that would have been unthinkable just a few years ago. According to a report published Thursday by Anthropic, the US-based artificial intelligence company behind the Claude model, this person used Claude to generate and distribute fabricated pro-Awami League news content for sixteen months. The scale is staggering: at least 1,500 headlines, 300 false narratives, and 1,500 prompts for creating images, all aimed at a rural Bangladeshi audience with limited literacy. The content was fed relentlessly into Facebook Live, YouTube, and TikTok, where it reached supporters of the Awami League in villages and towns across the country. The operation was not a sophisticated state-run intelligence program, but the work of one person with a custom script, a stable internet connection, and a deep understanding of Bangladesh’s political culture. Anthropic framed the report as part of an ongoing effort to track misuse of its technology, and the unusually detailed findings raise important questions about how easily AI can be transformed into a weapon of political manipulation. There is also a quieter, more human story here: a digital ghost embedded in Gaibandha, exploiting the trust and political sympathies of rural voters, the very people most often forgotten in conversations about technological change.
What makes the operation so striking is the care with which it was tailored to its intended audience. The prompts used to instruct Claude contained phrases that ring with the operator’s intimate knowledge of village life. The content was to be “hot and aggressive,” spoken in language simple enough “so even village people understand.” These are not neutral technical instructions; they are the words of a person who knew his audience firsthand, who understood that rural supporters of the Awami League did not want nuanced analysis but emotional reinforcement—something they could feel in their bones and share with their neighbors. The operator aimed at the heart of social media consumption in rural South Asia: short video clips, livestreamed commentaries, and audiovisual posts that can be absorbed quickly, without the burden of reading. The infrastructure was crude but effective. Claude generated batches of Bengali headlines, fabricated story details, and prompts for accompanying images. These were then deposited into cloud-storage folders, transformed into audio and video formats by separate programs, and scheduled for automated bulk upload to YouTube and other platforms. The operator masked his trail further by using a third-party automation service to hide his internet address. Once the system was humming, it required almost no direct human oversight. The entire production line—from raw idea to published video—became a self-sustaining generator of biased, false political content, designed to flow like water through the social feeds of village Bangladesh.
The targets of this operation reveal a clear political intent. Anthropic’s report identifies the Bangladesh Nationalist Party, Jamaat-e-Islami, the National Citizens Committee, the interim government, and student protest leaders as the primary subjects of attack. The fabricated narratives followed recognizable propaganda tropes: opponents were labeled foreign agents, accused of seeking Taliban-style rule, or implicated in invented assassination plots and secret political alliances. Some material claimed to expose financial corruption by rival leaders; other stories fabricated the existence of hit squads and underhanded schemes. These accusations were not random—they were crafted to align with local fears and existing political distrust, converting nuanced political opponents into existential threats in the minds of rural viewers. The operation unfolded against the backdrop of Bangladesh’s turbulent 2024, when mass student protests led to the ouster of longtime Prime Minister Sheikh Hasina and the installation of an interim government. For Awami League loyalists, the loss of power was humiliating and disorienting. The disinformation campaign can be understood as a continuation of that political contest by other means, a stubborn effort to keep the old party’s flame burning by discrediting its rivals and painting the new order as illegitimate. But the operator’s choice to anchor these stories in images of extremism and conspiracy betrays a darker strategy: to widen the chasm between Bangladesh’s political factions, making reconciliation ever more difficult and turning everyday conversations into battlegrounds for the soul of the nation.
The technical sophistication of the operation, combined with its simplicity, is one of the most disturbing aspects of the report. The conversation between the operator and Claude was mediated through a custom Python program with an almost comically honest name: “fake_news_3.py.” Each time it ran, it demanded a fixed package: fifteen Bengali headlines, three detailed fabricated stories, and fifteen English-language prompts for generating images. This assembly-line approach meant that a single operator could produce an endless stream of materials without rethinking artistic or journalistic choices. The integration of cloud storage folders allowed the content to be reviewed and processed before production, while a separate program managed the conversion of files into audio and video. These were then uploaded in bulk to YouTube, with publication scheduled days in advance, giving the impression of a newsroom carefully pacing its releases. The use of a third-party automation service to obscure the true internet address added another layer of insulation, making it harder for investigators to trace the source. Perhaps the most sobering detail is the claim that all of this required only minimal human supervision after it was set up. The operator had essentially handed Claude a blueprint and a to-do list, and the AI followed through systematically. Anthropic was able to detect the abuse, but not before it had produced months of highly visible disinformation. The cat-and-mouse game between AI creators and AI abusers is clearly only just beginning, and the Gaibandha case is a powerful reminder of how quickly the tools of content creation can be repurposed as tools of mass deception.
One of the most compelling sections of the report addresses the murky question of attribution. Anthropic states plainly that, in its investigation, it could not find proof linking this disinformation operation to either the Awami League or any Indian entity. This admission is refreshingly honest, but it leaves an uncomfortable riddle unsolved. The narratives were transparently pro-Awami League, and they aligned with Indian geopolitical interests, as India has traditionally had warmer relations with the Awami League than with its rivals. Yet the absence of evidence is not evidence of absence. The operator may simply have been a passionate citizen loyalist, hoping to defend the ousted party from the comfort of a Gaibandha home. It is also possible that he was a contractor working for shadowy interests, but carefully isolated enough to protect his patrons from liability. Or perhaps he was a victim of his own creation, a person who so deeply absorbed the worldview he was promoting that he continued out of conviction rather than obligation. Anthropic’s inability to trace the command chain is not a failure of investigation but a reflection of the reality of decentralized information warfare. In the digital age, a single person can be the tip of an opaque iceberg, carrying out tasks that serve greater political projects while remaining apparently disconnected from them. This ambiguity is precisely what makes disinformation so dangerous: even when the source is identified, the true orchestrator remains hidden, and public confusion persists.
The implications of the Gaibandha operation reach far beyond Bangladesh. This case marks a turning point in the democratization of disinformation. What once required a budget, a team, and editorial oversight is now achievable by anyone with access to an AI chatbot and modest coding skills. The consequences for democracies, transitional states, and fragile nations are profound. As AI continues to improve, the cost of producing credible-looking propaganda will fall even further, making it possible for small groups and even lone individuals to destabilize political processes. But the response must be human. Anthropic’s publication of its findings is an important step—transparency about misuse helps societies understand the threats they face—but technology alone cannot solve a problem that is fundamentally about human participation. Grassroots media literacy, community-based fact-checking, and local journalism that can respond quickly to false narratives are essential. Governments must resist the temptation to mimic disinformation tactics and instead invest in the institutions that build public trust. And platforms like Facebook, YouTube, and TikTok must take greater responsibility for detecting and labeling synthetic content, especially in languages and regions that receive less attention from moderation teams. Ultimately, the story of one person in Gaibandha is a story about vulnerability—the vulnerability of rural voters navigating a digital world they did not create, and the vulnerability of all of us in an age when our eyes and ears can no longer be trusted. But it is also a story about resilience, because naming the threat is the first step to overcoming it. The question that remains is whether the world will learn the lesson in time, or whether the next Gaibandha will already be quietly running its own script somewhere in the dim glow of a village computer screen.