Radio Domus: Safeguard, moderation built for languages global tools miss

Problem

Radio Domus is a community radio station in Kenya with growing digital conversations around its content on Facebook and YouTube. Off-the-shelf moderation kept failing in both directions. Global tools mis-flag benign local usage, such as an audience member commenting that a show "was a bomb," meaning it was excellent, and Facebook flagged it as terrorism. Meanwhile, clearly abusive Swahili and Sheng terms pass unflagged. Manual moderation ate limited staff time. The station wanted moderation that understands how its community actually speaks.

Solution

Safeguard is a locally built, AI-assisted moderation tool with:

• keyword filtering that flags harmful or misleading content using locally relevant terms

• post control, with review, approval and escalation before anything goes public

• capture of audience feedback, complaints and engagement signals

• ethical guardrails ensuring human oversight, data protection and accountability

Its core is a keyword library the station compiled itself of over 500 terms in English and Kiswahili, covering hate speech, ethnic discrimination, incitement, profanity and election manipulation, drawn from how the community uses language. Market tools alone would not work for a community station, so Radio Domus is building Safeguard on a paid Claude subscription while integrating third-party tools such as NapoleonCat and the Meta dashboard. Two external organisations were brought in for the front end, while the team learned the backend so the skills stay in-house. The human loop lives outside the tool, in Slack, where journalists submit drafts, an editor reviews, revisions follow, and approval precedes publication. The tool alerts whoever is online for post control, and a human makes the final call on everything.

Results

Safeguard is not yet public-facing. It is in phased internal adoption, tested by 11 users (eight staff, three interns) alongside existing tools. No formal impact analytics exist yet. The gains are meant to be at the workflow level, providing structured identification of harmful content, quicker internal review, clearer escalation decisions, more consistent moderation, and categorisation of comments to track patterns. In one documented example of success, abusive Swahili and Sheng terms that a generic English-trained tool would miss were flagged in its own Facebook comments.

Insights

The operating model is simple. AI suggests, a human reviews in Slack, the decision is logged. Core protections include data minimisation, role-based access to audience data, keeping sensitive data out of open AI tools, and reviewing false positives and negatives for bias. A moderation decision log builds newsroom-specific precedent over time. Safeguard is strongest as a content moderation and audience-intelligence prototype, according to the station. The fact-check archive and revenue potential are future extensions. Next steps include completing the dashboard, expanding the keyword library with election-risk terms ahead of Kenya's 2027 election, integrating the fact-check archive, and piloting with partner community radios.

Cost

Deliberately low, but the tech stack is wide, including subscriptions to Claude, ChatGPT, Slack, NapoleonCat and Trello, and a Microsoft 365 trial. A single AI subscription left the social media side uncovered, so complementary tools had to be budgeted.

Replicability

The method is transferable. Newsrooms need to run their own keyword-gathering exercise in their community's languages, choose build-and-integrate rather than build-or-buy, keep the human loop in cheap common tools, and log moderation decisions to build their own precedent. External technical help is expected, and budget for several tools rather than one subscription. It's recommended to start with the one or two platforms where a newsroom's audience actually is.