Imagine scrolling through X and suddenly seeing dozens of Kenyan voices praising Cabinet Secretary for Energy Opiyo Wandayi for stopping a scheduled electricity tariff increase. Hashtags like #PowerReliefKE and #PoweringTheNewKenya are everywhere. It feels like a spontaneous wave of public relief — ordinary citizens, happy about lower bills, grateful for a politician who listened. But the artificial intelligence company Anthropic has revealed a very different story. Behind that wave was not a movement but a machine: a single person using Anthropic’s Claude chatbot to mass-produce fake political posts, in batches of exactly 50 tweets, designed to look like organic grassroots commentary. The operation also attacked Kenya’s opposition, spreading claims that the United Opposition coalition was breaking apart ahead of the 2027 elections, and it specifically targeted political figures such as former Deputy President Rigathi Gachagua and former President Uhuru Kenyatta. Anthropic described the effort as astroturfing — an engineered campaign made to look like a natural, spontaneous public movement while actually being driven by hidden interests. The company said it identified and removed the account, and while it could not establish who was behind the operation, the pro-government tone and the specific hashtags suggested that it was plausibly aligned with supporters of President William Ruto. The full report, covering misuse between December 2025 and August 2026, gives a rare glimpse into how generative AI is becoming a new tool for political influence in Kenya, even before the official campaign season for 2027 begins in earnest.
Anthropic’s investigation did not happen in a vacuum. The company said it began probing Kenya after receiving a tip from rival US firm OpenAI about undisclosed criminal activity on its ChatGPT platform. That cross-industry alert points to a growing reality: AI companies are becoming frontline monitors of political manipulation, sometimes spotting operations before social media platforms ever see them. In this case, the user behind the fake grassroots campaign used Claude across several sessions, carefully instructing the chatbot that the posts should look like organic, individual commentary rather than a coordinated effort. The repeated use of exact batch sizes and campaign hashtags suggests a methodical system — someone turning a large language model into a political influence engine. Yet despite all its precision, the operation did not ultimately reach real people. Anthropic said its activity remained isolated within a network of fake accounts and local influencers on a single platform. That may sound reassuring, but it also shows how easily a lone actor can build a whole parallel world of political debate — a world that mimics the shape of public opinion without ever touching it. And the next operation might not be so contained. The fact that a chatbot could be told to produce fake public support with such ease raises uncomfortable questions about authenticity online. When every tweet could have been written by an algorithm, genuine expressions of support become easier to dismiss, and real political conversations can be drowned out by manufactured noise. This is what researchers call the “liar’s dividend”: the more realistic generated content becomes, the easier it is for bad actors to deny that their own words are fake, and the easier it is for ordinary citizens to give up trying to discern truth from falsehood.
For ordinary Kenyans, this story is more than a tech-industry case study. It goes to the heart of political participation in a digital age. Social media has become the town square, the market, the church gate, the place where people argue about politics and share news with friends. When that space is flooded with generated posts, the boundary between real civic engagement and manufactured consent begins to blur. A politician can point to fake online support as if it were a public mandate. An opposition figure can be attacked by a barrage of fabricated accusations, each one generated in seconds. And a public that cannot tell the difference may become more cynical, more anxious, or more easily manipulated. The African context makes this particularly urgent. Kenya has a vibrant digital culture, but it is also a country where elections carry real tension and where social media has often been used as a weapon. The 2027 election is still some time away, but political operators are already testing new tools and learning what generative AI can do. The anonymous use of Claude in this case mirrors a broader global trend: political influence is no longer only about buying ads or hiring publicists. It is about using algorithms to manufacture the appearance of a social movement. But the same technology that enables cheap manipulation also makes it cheaper to expose. Anthropic was able to detect the operation, trace its structure, and remove the account. Detection, however, depends on the willingness of private companies to look, and on rival firms to share information with one another. That is a fragile foundation for protecting democracy. Every company that builds powerful language models has an interest in safety, but they also compete for market share. When profit and safety collide, there is always a temptation to look the other way. The fact that OpenAI tipped off Anthropic is encouraging, but it also reveals how much depends on private backchannel relationships rather than public regulation.
The broader backdrop is Kenya’s rapidly shifting information ecosystem. A recent study by the Kenya ICT Action Network (KICTANet), based on monitoring between January and March 2026, found that alternative media actors on social media are shaping political narratives often with far less editorial accountability than traditional media. Kenyans now get news from Facebook, X, Instagram, TikTok, YouTube, and WhatsApp as much as from television, newspapers, or radio. These platforms are not neutral pipelines; they are emotional delivery systems, built to keep eyes on screens and feelings running high. Algorithms tend to amplify content that sparks strong reactions — anger, fear, outrage, humour, or political loyalty. When a false narrative begins as a fabricated newspaper headline on Facebook, it can be screenshotted, passed around WhatsApp, turned into a TikTok clip, and then discussed for hours on podcasts and YouTube channels. By the time a mainstream media outlet fact-checks it, the narrative has already travelled far beyond the reach of any correction. KICTANet specifically flagged Facebook as a common starting point, where fake headlines, doctored political graphics, and impersonated editorial content first appear. TikTok then accelerates those narratives through emotionally charged and selectively edited clips, while YouTube and podcast channels extend their life through commentary and speculation. The study noted that platform algorithms play a major role in this amplification: sensational content often receives more visibility than balanced, evidence-based reporting. This is not unique to Kenya, but it matters deeply in a democracy where many people rely on social media for both news and a sense of belonging. When generative AI is added to the mix, the result is what researchers call narrative laundering: a false claim enters through an obscure corner, gets cleaned up by fake accounts and meme pages, and then emerges in the mainstream as if it were public opinion. That was exactly what the Claude-powered operation was trying to achieve.
KICTANet’s warnings about the growing use of generative AI make the situation even more sobering. Deepfakes and manipulated content will become more sophisticated ahead of the 2027 elections, the study says. False content may become more convincing and harder to detect; manipulated media may spread rapidly across platforms; verification processes may struggle to keep pace. This is not a distant prophecy. It is already happening in fragments: a fake video of a politician, an altered audio clip, a doctored photograph of a rally crowd, a headline that was never written. Each forgery is inexpensive to make and expensive to debunk. The asymmetry is the real problem. A single person with an account can generate hundreds of images, thousands of posts, and dozens of videos in the time it takes a journalist to verify one. And once a false narrative is out in the world, a correction rarely travels as fast or leaves the same emotional mark as the original lie. Kenyan society is already familiar with disinformation, but generative AI changes the scale and speed of the threat. During earlier election cycles, political manipulation required human labour: copywriters, graphic designers, social media managers, and entire teams of fake account operators. Today, a chatbot can do the work of a whole troll farm. In the Anthropic case, the account was used to generate batches of exactly 50 tweets, but there is no technical reason an operator could not generate thousands or millions. The limit is not the model; it is the patience of the person using it. This is why Anthropic’s report matters beyond Kenya. It gives the public a detailed, behind-the-scenes look at how an AI tool is being weaponised for political astroturfing in a real African democracy. It also exposes the limits of the platform-response model. Anthropic removed one account. The same actor could create another account tomorrow. The same technique could be used by dozens of actors in dozens of countries without breaking a single law. There is no global treaty that can stop a person from typing prompts into a chatbot. There is only a patchwork of corporate policies, platform rules, and civil-society monitoring — all of which can be outmanoeuvred by a determined operator.
Kenya has seen this movie before, in a different form. During the 2017 General Election, the British political consulting firm Cambridge Analytica became infamous over allegations that it used data analytics, psychological profiling, social media advertising, and grassroots networks to influence voters. The company was accused of harvesting Facebook user data through a third-party application and using it to target users with fake-news campaigns. That scandal became a global symbol of digital manipulation in elections. But the technology then was still primitive by today’s standards. Cambridge Analytica needed massive datasets, complex psychographic models, and a network of human operatives. Generative AI changes the equation completely. The Anthropic case shows that a single actor can now simulate public sentiment using nothing more than a chatbot and an internet connection. There is no need to steal data from millions of users when you can simply ask a model to imagine what those millions would say. The implications for electoral integrity are profound. If political actors can generate fake support and fake opposition at zero marginal cost, public opinion begins to lose its meaning as a signal to policymakers. An election campaign that appears to be surging online could be a mirage. A political coalition that seems to be falling apart may simply be the victim of a clever prompt. The erosion of trust is perhaps the greatest danger. The answer cannot be only better detection technology. It must also be a more informed and resilient public. Media literacy programmes, independent fact-checkers, transparent platform policies, and enforceable electoral regulations are all necessary. Kenyans, with their history of lively civic engagement, already have the tools to push back. But they need to understand that AI-generated content is now a permanent part of the political landscape. They need to ask questions about what they read, what they share, and what they believe. The machines may generate the lies, but only human beings can decide whether to spread them. The story of Opiyo Wandayi’s fake praise is not just a story about artificial intelligence. It is a story about power, influence, and the fragile trust that holds democratic societies together. As 2027 approaches, every Kenyan should remember that behind any hashtag, any viral post, any wave of online anger or approval, there may be a person, a group, or a chatbot with an agenda. The challenge is to keep asking one simple question: who is really speaking, and what do they want us to believe? Answering that question may be the most important form of civic participation in the age of AI.