A political operator in Kenya used artificial intelligence to manufacture fake grassroots support for government leaders and push narratives aimed at weakening the opposition ahead of the 2027 General Election, according to a new report by Anthropic.
The company said the operator used its Claude AI models to produce large batches of social media posts designed to appear as genuine public reactions rather than part of an organised political campaign.
Anthropic detailed the activity in its September 2026 report on artificial intelligence misuse, released on Thursday, September 10, covering operations it detected and disrupted between December 2025 and August 2026.
The Kenyan operation included posts praising Energy Cabinet Secretary Opiyo Wandayi over the government’s handling of a proposed Kenya Power electricity tariff increase. The messages used hashtags such as #PowerReliefKE and #PoweringTheNewKenya in an attempt to give the content greater visibility online.
Part of Anthropic’s report, published on Thursday, September 10, 2026, on page 77. PHOTO/Screengrab
The same campaign pushed claims that the United Opposition was falling apart ahead of the 2027 election, with some of the messaging aimed at former Deputy President Rigathi Gachagua and former President Uhuru Kenyatta.
According to Anthropic, the operator repeatedly asked Claude to generate batches of exactly 50 posts at a time. The instructions were designed to make the messages resemble ordinary social media conversations and hide the fact that they were being produced as part of a coordinated operation.
The company, however, said the political themes and messages were not created by the AI. The actors had already developed the narratives before using Claude to produce them on a larger scale.
Anthropic said it found no evidence linking the operation to the Kenyan government. It assessed the activity as being consistent with a domestic Kenyan operation but said it was unable to establish which organisation was behind it.
The operation was placed in Category One on Anthropic’s Breakout Scale, which has six levels for measuring the spread and reach of influence campaigns. Category One is the lowest level and covers activity that remains limited to one platform or community, without signs that it has spread widely.
Anthropic said the fake accounts did not achieve meaningful engagement with genuine users. It found that Claude’s main value to the operator was the ability to quickly produce large quantities of political content that was made to look authentic.
The company also found another operation linked to the same actor involving Kenyan retail brands.
In that campaign, the operator used personas called “SHANKI” and “Elkins Marketer” to promote commercial products. Advertising material was rewritten and presented as if it were ordinary social media content, with links placed in every fifth post.
The findings come as Kenya heads towards the 2027 General Election, with political activity and public debate increasingly taking place on social media platforms.
The Independent Electoral and Boundaries Commission has previously identified the digital space as a major area of concern. In procurement documents for social media monitoring services, the commission said Kenya had more than 15 million active social media users and warned that online platforms could be used to quickly spread electoral misinformation, disinformation, hate speech and incitement.
The potential reach of such content may now be higher. DataReportal estimated that Kenya had 23.4 million internet users by the end of 2025, accounting for 40.5 percent of the population, while social media user identities stood at 18.4 million, equivalent to 31.8 percent of the population.
The data showed Facebook had a potential advertising reach of about 17 million users in Kenya, while X had about 2.09 million users.
Kenya has previously faced organised online campaigns involving political misinformation and hate speech.
Research by the Mozilla Foundation before the 2022 General Election found 130 TikTok videos containing hate speech, incitement and political disinformation had attracted more than four million views across 33 accounts.
After the election, Mozilla reported that Facebook, TikTok and Twitter had faced difficulties in containing misinformation, including false claims concerning political candidates and election results.
The use of generative artificial intelligence introduces another layer to the problem by allowing operators to create large amounts of seemingly genuine political content in a shorter time and at lower cost.
While the Kenyan operation identified by Anthropic did not achieve a wide reach, the case shows how AI tools can be added to existing political influence and online marketing methods as parties and political actors prepare for the 2027 election.
Anthropic said detecting such activity early is important because AI-generated material can be produced at scale before it reaches social media platforms and genuine users.