Anthropic, the US company behind the Claude AI model, has revealed that a Kenyan operator used its technology to mass-produce political content designed to look like genuine public opinion ahead of the 2027 General Election.
The company published the findings in a threat intelligence report covering several types of Claude misuse, including cyberattacks, surveillance, and coordinated influence operations.
One case involved what Anthropic called a domestic Kenyan actor running a political astroturfing campaign (organized effort that pretends to be a spontaneous grassroots movement), tracked internally under the name GTG-54004.
How the Campaign Worked
According to Anthropic, a single person used Claude across multiple sessions to generate tweets in batches of exactly 50 at a time.
The person gave Claude specific political topics, talking points, and hashtags, and told the model to write the posts so they would look like spontaneous grassroots commentary rather than a coordinated campaign.
Much of the content praised Energy Cabinet Secretary Opiyo Wandayi for halting a planned Kenya Power electricity tariff increase. These posts used the hashtags #PowerReliefKE and #PoweringTheNewKenya to boost their reach on X.
The same operation also pushed narratives claiming that Kenya’s United Opposition coalition was falling apart before the 2027 election.
These posts specifically targeted former Deputy President Rigathi Gachagua and former President Uhuru Kenyatta, figures who have grown increasingly critical of President William Ruto‘s government.
The timing lines up with real political tension the country is currently experiencing. Gachagua and other opposition figures have stepped up criticism of the Ruto administration, while Kenyatta has faced attacks from Ruto allies over claims he remains politically active despite being retired.
Nandi Senator Samson Cherargei even threatened to introduce a bill to strip Kenyatta of his retirement benefits over these allegations. Kenyatta has also been accused, without confirmed evidence, of funding the Linda Mwananchi faction of the opposition led by Nairobi Senator Edwin Sifuna.
Same Operator, Different Job
Anthropic found that the person running the political campaign used the identical AI workflow for commercial work, marketing Kenyan retail brands under the names “SHANKI” and “Elkins Marketer.”
This showed the operation wasn’t purely ideological. It was also a business, and the same tools and techniques were reused for paid client work.
How Anthropic Caught It
The investigation began after a tip from OpenAI about repeat inauthentic behavior on its own platform. That lead prompted Anthropic to look into similar activity on Claude, which led investigators to this operation.
Anthropic says it removed the account and the organization behind it, then built detection systems based on the behavior pattern to catch similar attempts in the future.
How Much Damage Did It Do?
Despite how organized the operation was, Anthropic rated its real-world impact as low. Using its own Breakout Scale, the company classified the campaign as Category One, its lowest tier.
The fake posts stayed contained within a network of fake accounts and paid local influencers on a single platform and never appeared to reach or sway actual voters.
Anthropic also said it found no evidence that the Kenyan government was directly involved.
Investigators noted the campaign’s pro-government tone and hashtag choices suggested alignment with the ruling coalition, but they could not identify a specific organization or confirm official backing.
The Government’s Response
Energy CS Opiyo Wandayi addressed the report directly, saying it did not prove that he, his ministry, or anyone working under him authorized or took part in the campaign.
He argued that anonymous, coordinated social media activity happens constantly online and that having a coordinated network push a narrative is not the same as proving government involvement.
Wandayi called the report “unverified” and said what mattered was actual evidence tying the accounts to his office, not just the existence of the posts themselves.
