Anthropic has removed an account that someone was using to mass-produce Kenyan political content built to look like ordinary public opinion. The operator asked Claude, Anthropic’s AI assistant, to write the posts so they read as spontaneous grassroots commentary rather than as a coordinated campaign.
The case is on pages 75 to 77 of Detecting and countering misuse of AI: September 2026, a 154-page report Anthropic published on 10th September 2026 covering misuse of Claude that it caught and disrupted between December 2025 and August 2026. Anthropic files the Kenyan operation as GTG-54004.
Nine influence operations are documented in the report. Kenya is the only purely domestic one in Africa, and it sits alongside a couple of more issues including:
- a Russian-funded radio operation in the Central African Republic and
- Russian state media writing about Moldova’s president before a national vote.
What the tweets said
A large share of the AI-generated posts praised Energy Cabinet Secretary Opiyo Wandayi for stopping a Kenya Power electricity tariff increase, under the hashtags #PowerReliefKE and #PoweringTheNewKenya.
Kenya Power applied on 31 March 2026 to raise retail electricity rates, and on Wednesday 3 June 2026 Wandayi announced at a press briefing that the application had been withdrawn. “The current retail electricity tariffs shall remain in force and unchanged, unless otherwise lawfully reviewed in accordance with the Energy Act,” he said. We covered that story at the time, including the part most of the celebration skipped: the energy charge is less than half your bill, and the fuel and forex adjustments that make up the rest kept rising anyway.
The second half of the operation pushed stories claiming Kenya’s United Opposition coalition was falling apart ahead of the 2027 general election. It named Rigathi Gachagua and former president Uhuru Kenyatta directly.
Anthropic’s understanding of the situation was this: the playbook is to boost the incumbent administration while undermining the opposition’s viability. That makes it “consistent with a single, coordinated electoral communications effort.”
What Claude actually did
Claude didn’t do the thinking. It was used to push the volumes.
Whoever handled the account supplied the topic, the talking points and the hashtags. Claude turned them into 50 themed posts per batch, character-counted and formatted for posting across platforms. In several sessions the output came back packaged as an interactive copy-all widget, ready to deploy.
Anthropic is clear about where the model was best fit for such an operation. It was used “to humanize and refine batches of 50 pre-drafted topics and tweets, showing that the model’s main value to the network was volume and the appearance of authenticity. The core ideological messages were completely decided by the actors before Claude was used.”
So nobody asked an AI what to think and say about Wandayi. Somebody already knew what they wanted said, and hired a machine to say it 50 times in 50 different voices. Better safety filters would not have stopped that.
The same file that attacked Gachagua also sold retail brands
The operator of the account was doing commercial marketing work too. This was under a persona called “SHANKI” or “Elkins Marketer”. Anthropic says the political template was reused verbatim for Kenyan retail brands, and that one promotional broadcast came back as organic-looking tweets with a link inserted into every fifth post.
Anthropic reads that as fitting “a commonly-observed pattern in Kenya where agencies pay local influencers to manipulate narratives.”
The political posts came off the same production line as the brand work, at the same batch size. Whoever wanted Wandayi praised did not have to build anything. They hired a firm or person who does this for retail clients on ordinary weeks.
Nobody has named who paid for it
Anthropic did not identify the people behind the operation. It found no evidence of government involvement and assesses the activity as entirely domestic. It says the pro-administration tone and the specific hashtags “suggest it was plausibly aligned with the ruling coalition, though we have not identified the exact organization responsible.”
So yeah, Wandayi, his ministry, or anyone in government isn’t particularly being accused of doing anything here. However, they’re the subject of the prasie.
Somebody paid for this. Fifty tweets at a time, across several sessions, with a brand-marketing business running off the same template, is paid work.
Anthropic banned “the account and the organization behind it”. That tells you it ran under some entity. Whoever commissioned it wanted a Cabinet Secretary to look good and the opposition to look finished. That work does not get done for free, and in Kenya it is rarely done by accident.
How this was caught
Claude did not refuse to write the tweets.
However, what stopped the operation was housekeeping. Anthropic says it acted “based on a tip shared by OpenAI about recidivist activity on their platform”. OpenAI flagged a repeat offender it had caught, Anthropic went looking for coordinated inauthentic behaviour in Kenya, and found this. It then removed the account and the organisation, and built detections around the behavioural signature so the same pattern gets blocked next time.
The operator had already been caught once on ChatGPT, moved to Claude, and was only caught again because two competitors talked to each other.
Anthropic points out that a social platform usually sees an operation once the content is already circulating, while an AI company sees it being written. The propaganda gets drafted somewhere before it gets posted, and that somewhere keeps logs. Anthropic is honest about where the visibility stops: “Our visibility into these operations ends once it’s live.”
None of this is new in Kenya. It is just faster now
Paid disinformation on Kenyan social media has been documented for years. Mozilla Foundation research published in 2021 by fellows Odanga Madung and Brian Obilo found at least 11 coordinated campaigns on Twitter, involving more than 23,000 tweets and 3,700 accounts, targeting Kenyan journalists, judges and civil society. People who took part told the researchers they were paid roughly USD 10 to USD 15 a day, sent over M-Pesa, to push hashtags into the trends.
A follow-up Mozilla investigation found the Spain-based organisation CitizenGO paying Kenyan influencers at similar rates to flood Twitter with misleading posts and graphics about reproductive health legislation before parliament.
What changed is the cost. In 2021 you needed 3,700 accounts and a WhatsApp group full of people willing to tweet for about KES 1,300 a day. In 2026 one person and a subscription produce 50 unique, character-counted, ready-to-paste posts per batch, and the humans are only needed at the end, to press post.
And, it’s campaign season
This will get worse, and it will come from every direction. Government does it. The opposition does it. Party-aligned bloggers, big X accounts, comedians, TikTokers and YouTubers get put on retainer by both sides, and a lot of the time the audience cannot tell a paid narrative from a genuine opinion, because the people delivering it are the same people they follow for entertainment.
Anthropic’s report frames this case as “a pro-government operator in Kenya” preparing fake grassroots posts “ahead of Kenya’s 2027 general election”. It is barely a year to the general elections, so the time is nigh.
The surveillance side is moving too. In February we reported on a Citizen Lab forensic investigation into the DCI’s use of Cellebrite tools to unlock activists’ phones ahead of 2027. Shaping opinion online and monitoring the people who object are both election spending.
What to do?
Every time you see what appears to be coordinated praise, treat that as a signal that it is paid for, or automated with AI.
Join the discussion