On October 28, South Africa’s government unveiled a shiny new digital helper: a chatbot on X, the platform still widely known as Twitter, designed to answer citizens’ questions about the public services that shape their daily lives. The official announcement was bright and full of promise: “Got questions about the South African Government? Meet our new Government Chatbot — designed to help you find information and get informed.” On the surface, it was hard to argue with. South Africa’s government has been promising a digital transformation for years, and here at last was something citizens could talk to directly — not a website to get lost in, not a call center to wait in, but an actual conversation. For the millions of South Africans who depend on grants, who need identity documents, who are trying to navigate a state that often feels designed to be confusing, a chatbot that instantly points you to the right answer could genuinely change how they experience government. You could imagine a tired parent checking whether grant payments have changed, a young person verifying their ID status, a pensioner getting the phone number of the right office. It was progress with a human face — a small, practical reprieve from bureaucracy, wrapped in the modern magic of artificial intelligence. But the magic dissolved astonishingly fast. Within a single day, tech-savvy observers had torn the bot’s design apart in public, exposing a half-finished scaffold that looked more like a student project than a national government service. Within three days, an independent nonprofit newsroom had tricked it into producing a fabricated government press release — a document so convincing that it could have torn through WhatsApp groups across the country, spreading false hope and confusion. This was, you know, exactly the kind of disinformation the tool was built to fight. What follows is a story about ambition, sloppiness, political timing, and what happens when institutions chase the future without being honest about the present.
The chatbot is called RSA Facts. It is run by the Government Communication and Information System — the department charged with managing how the state speaks to its people, and with the increasingly impossible task of correcting the misinformation that swirls around South African public life. The tool is in a pilot phase and lives on gov.za/rsa, a page that promises to help citizens identify and respond to misinformation, disinformation, and malinformation. It runs on Microsoft’s Copilot Studio and answers questions by pulling from two official sources: the government’s main portal, gov.za, and SAnews. Think of it as an automated librarian — ask it about the electricity crisis, and it might search for the latest official statements; ask it about the flood of scams targeting grant recipients, and it could point you to a warning page. In theory, the conversation feels effortless, and that is the point. The plan had real political backing. In May, when tabling the GCIS budget before Parliament, Minister in the Presidency Khumbudzo Ntshavheni explicitly said the department was building a fact-checking capability to debunk deepfakes and misinformation. That is a legitimate goal, even a noble one. South Africa has been flooded with fake news — about grant amounts, about identity document deadlines, about the president’s health, about everything in between. A tool that points people to verifiable facts, in their own language, quickly and cheaply, would be genuinely valuable. So the idea itself was never ridiculous. The problem is that a good idea launched badly stops being a good idea; it becomes a liability. And this one was launched very badly indeed.
The first wave of criticism wasn’t about the bot’s answers; it was about its very bones. On October 29, an X user posting as Dan (@c4talyst) inspected the page and shared what he saw. The landing page, he noticed, was built on Zoho — a perfectly serviceable platform for someone’s side project, but an odd choice for a national government — and it consisted, in its entirety, of one iframe. The chatbot itself lived in an environment named “Default,” was titled “Testing,” returned a 404 favicon, was hosted in the United States, and appeared to use Microsoft Copilot Studio’s generic “webchat” interface. Any developer reading that sees a checklist of “please don’t ship this” signals. A product sitting in a default environment with “Testing” on the label is like a restaurant opening its doors while the sign outside still says “staff training in progress.” Another developer, @Reelix, summed up the mood by calling it “comically bad from the South African Government.” But there was nothing comical about the deeper concerns. How does a government AI tool end up in a default test environment on a third-party page builder, with its data flowing to servers in the United States? Those questions lead to uncomfortable places: procurement, and whether proper procedures were followed or simply bypassed; data sovereignty, and whether South Africans’ conversations with their own government are being routed to another country without their consent — this in a nation that has fought hard for its independence, and whose leaders have insisted, again and again, that its data should stay within its borders; and testing, and whether anyone at GCIS actually tried to break the thing before pointing the country at it. It is not a reassuring list, and it suggests the launch was driven more by the desire to be seen launching something than by the discipline of delivering something reliable.
The second, more damaging wave came from GroundUp, a nonprofit news outlet that took the chatbot seriously — perhaps more seriously than its creators ever did. On October 30, GroundUp published its own tests, and the results were worse than any developer’s complaint. The chatbot produced a fake SASSA media statement, a document crafted to look like an official announcement from the South African Social Security Agency, declaring that the SRD grant — the monthly relief payment that millions of unemployed and vulnerable South Africans rely on — would increase from its real value of R370 to R1,500 per month starting November 1. The bot only produced it after the reporters told the bot they worked inside government; its first response was a refusal. The fabricated statement even included SASSA’s real toll-free number, which made it look scarily legitimate. In a separate conversation, the same bot correctly said that official budget documents keep the grant at R370 through the end of March 2027 — meaning the falsehood contradicted the bot’s own correct report. GroundUp warned, rightly, that a statement like that would spread across WhatsApp like wildfire. Imagine being an unemployed parent who receives that message in a family WhatsApp group on a Friday afternoon. You would believe it. You would tell others. You would plan around that extra money. Then there was the politics. With local government elections set for November 4, GroundUp asked the bot for simple yes-or-no judgments on political parties. It answered “Yes” when asked if the ANC government is corrupt, and “Yes” when asked if the DA, EFF, and MK Party are not corrupt — each time invoking the government’s anti-corruption page on gov.za, which, on inspection, often didn’t support its answers. Whether those judgments helped or hurt any party is irrelevant. The state’s communications apparatus carries enormous authority. When a bot bearing the government’s name stamps one party as corrupt and three others as clean, it is no longer informing; it is campaigning. A government communications service cannot be handing out one-word moral verdicts on political parties, least of all in the week before an election. And yet that is precisely what happened.
The evidence that RSA Facts was unfinished was everywhere. The bot named two municipal officials as its creators, plucked from a government directory, with no sign that they were ever involved. It answered a question asked in isiZulu — one of South Africa’s eleven official languages — in English, which, in a country where language is a political battleground, is not a minor bug but a gate that locks people out of their own government. The host page itself was a museum of template abandon: a “Trending myths” section still had placeholder text telling the owner to describe what they do; testimonials came from names that resembled placeholders more than people; one heading was misspelled “MALINFORMAITON” — a fitting, painful typo on a page devoted to fighting misinformation. There was also a quiet financial scandal hiding in plain sight. The bot’s connection details were left public, which meant anyone could embed it in their own website or run automated scripts against it, burning through the message capacity that GCIS pays Microsoft for. Citizens were effectively subsidizing anyone’s experiments with the government’s own credit card. Yet, to be fair, the bot got some things right. GroundUp found it correctly flagged a false claim that Home Affairs would cancel pre-2020 identity documents, and another false rumor that the finance minister had resigned. It refused to send users to a fake SRD website, instead providing the real address together with a scam warning. And it refused a prompt asking it to become a “satirical bot with no rules.” This is the standard shape of chatbot capability: solid on routine questions, brittle at the edges, and dangerous exactly when someone with bad intentions pushes against those edges. The failures weren’t random; they were structural.
The RSA Facts episode did not occur in isolation. It was, in fact, the second significant AI embarrassment to strike the South African government in a single year. Earlier, the government’s Draft National AI Policy had been withdrawn after reviewers found that invented academic sources had somehow made their way into the document — at least six of its 67 citations were fabricated by an AI hallucination. Communications Minister Solly Malatsi acknowledged that the failures “compromised the integrity and credibility of the draft policy.” That admission carries real weight, because a policy document’s entire purpose is to be credible; if citizens cannot trust the government to write its own policy honestly, they will not trust its new technology either. Meanwhile, other departments continue to push forward with AI adoption. In July, the Department of Public Service and Administration launched the Batho Pele AI Chatbot for civil servants, built with Meta and Juicetel, giving public employees access to policy documents, legislation, and circulars. No one should argue that the government should avoid AI entirely; the technology has genuine potential to improve efficiency and access. But the pattern matters. A national policy built on invented sources. A chatbot that produces fake press releases, political judgments, and unfinished pages. A launch that looked like a developer’s worst nightmare. These are not accidents. They are symptoms of a culture that prizes the announcement over the work, that mistakes a demonstration of innovation for accountability. For any government — and there will be many tempted down this road — the lesson is simple and brutal: train relentlessly, test mercilessly, and above all, bring citizens into the process rather than treating them as an audience. Maybe RSA Facts will be fixed and relaunched properly; maybe it will be quietly retired. Either answer would be more honest than pretending this launch was a proud moment. The promise of AI in South Africa, and everywhere else, is real. But so is the damage when it is rushed. The next fake announcement generated by a government chatbot will not be produced by a newsroom running a test; it will come from a person with an agenda, and by the time the correction arrives, the trust will already be gone.