Lake Region Bulletin: A fact-checking agent built by non-technical editors
"Coding using words" — a custom ChatGPT agent trained on the newsroom's own debunks, with two editors as the gate.
Problem statement
Lake Region Bulletin (LRB) is a small digital newsroom in Kisumu, western Kenya — around ten staff, with high turnover as journalists leave for better-paying jobs — founded in 2022 to cover community stories dropped by shrinking mainstream print. It runs a fact-checking desk, and the manual work of collecting suspect content, running it through verification tools and producing debunks was slow and labour-intensive. The question the newsroom set itself was simple: how do we make this work easier — getting more out of the existing team without hiring?
The solution
LRB built a custom agent on top of ChatGPT — not a from-scratch tool and not commissioned work, since developing from scratch was judged too technical and too expensive (cloud hosting, maintenance, monitoring) for their scale. Two editors did the work themselves after in-person training; neither has a technical background. The unlock, in their words, was learning they could "code using words" — configuring the agent purely through natural-language instructions. The recipe: explain to the agent who LRB is and what it wants to achieve; load it with instructions on how to fact-check, plus reference resources; upload several of LRB's own published fact-checks so it learns the house format. In use, the agent traces whether a photo or video has been published before and where — most of the mis/disinformation they see is recycled, such as a Nigerian photo passed off as a Kenyan event — letting them locate originals and debunk. Findings are fed back in and the tool drafts a fact-check story in LRB's format; a human reviews before anything is published. When outputs failed the "human test", they iterated on the prompt rather than rebuilding — adding resources and more specific instructions.
Results
The agent reduced fact-checking time, freeing journalists for other tasks, and is now part of the regular workflow. No AI-caused mistake has ever gone to press — every error was caught in human review. The limits are equally instructive: persistent hallucinations (in one illustrative case, the tool asserted Barack Obama is the current US president), which is exactly why the human cannot be removed from the loop; some websites block AI access, so needed information is sometimes unreachable; and imprecise prompts sent the tool "fishing" wrong information from the internet.
Insights
Access is deliberately restricted to the two editors "to regulate its access for proper management", with wider onboarding only "once we are convinced it is safe". The desk covers on average two fact-check stories a week, plus general verification of figures in other reporting — development, climate and science stories included.
Cost
A grant covered expenses; the main cost is the paid subscription, around US$25–30 a month — a figure the newsroom itself was still to confirm — after testing packages from the smallest tier upward and finding a modest one sufficient. A secondary cost: supporting staff home Wi-Fi, since the team works remotely.
Replicability potential
This is a prompt-configured agent on a consumer AI subscription, built by non-technical editors and trained on the newsroom's own published stories — no coding, minimal budget. LRB's advice to peers: define a specific task first; research which tool is strong for that task; expect free tiers to be limiting; and expect the tools to keep changing. What another newsroom must supply itself: its own corpus of house-format fact-checks, and its own local threat picture — LRB's agent is tuned to recycled foreign content re-captioned for Kenya, and yours would need tuning to whatever circulates where you are.