AI promise versus newsroom reality gap
The European Union’s Artificial Intelligence Act, the first comprehensive law of its kind, came into force this month. It does not ban AI in newsrooms. It demands something more uncomfortable: that media organisations explain what the technology actually does, rather than what it claims to do. That distinction matters far beyond Europe, and it cuts to the heart of a problem South African journalists are only beginning to name out loud.
The gap between promise and practice is where deception lives. Not the dramatic deception of deepfakes designed to fool voters, though that exists too. The quieter deception is the one happening inside newsrooms every day, where an AI tool drafts a paragraph, translates a quote, or summarises a court ruling, and the journalist on the other end cannot always tell where the tool’s confidence ends and its competence begins.
The Illusion of Effortless Accuracy
AI presents itself as a tireless assistant. It never sleeps, never gets bored, never skims a document because the coffee has gone cold. That presentation is the first layer of the deception. The technology appears to have done the work, and appearing to have done the work is often enough to pass a tired editor’s eye.
The reality is messier. Training data for most AI models is drawn overwhelmingly from outside South Africa, which means the tools carry a quiet bias toward contexts they understand and a quiet blindness to contexts they do not. A phrase that makes perfect sense in a Johannesburg political rally can become something else entirely when processed through a model shaped by American or European corpora. The tool does not announce its confusion. It simply produces output with the same confident tone it uses for everything else.
That confidence is the deception. It is not malicious. It is structural. An AI system cannot tell you what it does not know, because it does not know that it does not know. It presents every answer with the same flat certainty, and the journalist who has not been trained to interrogate that certainty is left with a choice: trust the output or verify it manually, which defeats the purpose of using the tool in the first place.
The Secret
Users and the Silent Workflow
A recent study from the Centre for Information Integrity in Africa (CINIA) found that AI use in South African newsrooms is often driven by whoever happens to have a personal interest in the technology. [1] There is no formal policy. There is no training programme. There is an individual who figured out that ChatGPT can draft a press release summary in thirty seconds, and that individual starts showing colleagues, and soon the tool is embedded in the workflow without anyone having made a deliberate decision about whether that is a good idea.
Some journalists told researchers they suspected colleagues were using AI “secretly.” Others admitted they were unclear about what they were allowed to do and what they were not. The result is a newsroom where the technology is everywhere and nowhere at once. It shapes the work, but no one has agreed on the rules. It influences what gets published, but no one has decided who is accountable when it goes wrong.
This is the second layer of deception. The AI is not hiding. The humans are. They are hiding because they know the tool is useful, and they suspect that formal approval would come with restrictions. So they use it quietly, which means they also hide their mistakes quietly. A journalist who uses AI to translate a political term and gets it wrong is unlikely to raise their hand and announce the error if they were never supposed to use the tool in the first place.
When the Machine Hallucinates Authority
The most public failure of AI in South Africa this year did not happen in a newsroom. It happened in a courtroom. An acting judge in Johannesburg produced a ruling that cited case law that did not exist. The AI had hallucinated the citations. The judge had apparently trusted the output without verification.
That case is a warning to journalism because the same failure mode applies. An AI tool does not distinguish between a real precedent and a plausible one. It generates text that looks like a citation, sounds like a citation, and carries the same grammatical weight as a real citation. The only difference is that the underlying case was never decided, the ruling was never written, the court never sat.
The deception here is not that the AI lied. It is that the AI produced something indistinguishable from the truth, and the human on the other end did not have a reliable way to tell the difference. In a courtroom, that leads to embarrassing reversals. In a newsroom, it leads to published errors that erode the one asset journalism cannot afford to lose: credibility.
The government’s own draft AI policy suffered the same failure. The Department of Communications and Digital Technologies produced a document with fake citations, apparently generated by AI and not checked. If the people writing the rules for AI cannot follow the basic discipline of verifying what the technology produces, that is a signal to every newsroom in the country about how easily the gap between promise and practice widens.
The European Standard as a Mirror
The EU AI
Act does not solve these problems. What it does is force a conversation. Under the new law, media companies can be held liable for harmful AI-generated content. They must demonstrate accountability. They must show that a human was in the loop, that the output was checked, that the publication can explain why the tool was used at all.
Hendrik Sittig, director of the Konrad-Adenauer-Stiftung’s Media Programme for Sub-Saharan Africa, puts it plainly: the Act does not force newsrooms to stop using AI. It requires transparency, labelling, and editorial oversight. News organisations that use AI will increasingly have to explain how it is being used and who is ultimately responsible for what they publish.
That is the point where the gap between what AI claims and what it does becomes visible. When a newsroom has to explain, in writing, why an AI tool was used for a particular story, the explanation itself reveals the limits. The tool did the first draft. The tool translated the quote. The tool generated the image. And then a human checked it, or did not, and that distinction becomes part of the public record.
South Africa does not have an equivalent law. The Press Council of South Africa issued a guidance note in 2023 urging editors to be thoughtful about AI tools, but the note is not binding. [2] The Council’s executive director, Phathiswa Magopeni, says the industry discussions come down to media ethics: verification, accuracy, and editorial due diligence. [2] She wants to see disclosure requirements that match the reality of AI being used at different stages of production, from research to writing to image generation.
The gap between the European approach and the South African approach is not just legal. It is cultural. Europe has decided that AI transparency is a right, something audiences are owed. South Africa is still having the conversation about whether that right exists and how it would be enforced. In the meantime, the tools are already in the newsroom, already producing output, already shaping what audiences read and hear.
The Trust Calculus No One Wants to Do
Editors in South Africa are under pressure. Staff cuts are common. Automation is seen as a way to compete with online platforms that do not carry the same professional obligations as legacy media. The temptation to let AI fill the gaps is real, and it is growing.
The risk is that the audience notices before the newsroom does. Audiences are not stupid. They can tell when a story has a certain generic quality, when the language does not quite match the publication’s voice, when the details do not quite add up. And once trust is gone, it is very hard to rebuild.
The same CINIA study found that most instances of “inauthentic” AI-generated content in South Africa have appeared on social media rather than broadcast news. [1] Deepfakes endorsing political parties during elections were called out by a fact-checking community that is energetic but underfunded. That community is doing the work that newsrooms should be doing themselves, and they are doing it with fewer resources.
The deception is not always visible. Sometimes it is a mistranslated political term that changes the meaning of a quote. Sometimes it is a summarised court ruling that leaves out a crucial qualification. Sometimes it is an image that looks like a photograph but was generated from a prompt. Each of these is a small gap between what the AI appears to have delivered and what it actually delivered. Each gap is a small erosion of trust.
The Future Is Not a Policy Document
CINIA’s Rejoice Malisa-van der Walt argues that AI regulations for the country need to be dynamic and inclusive, warning that static rules will quickly become obsolete. [1] The technology is advancing faster than any static rule can track. In a news context, that means consulting audiences about their tolerance for AI use, identifying the red lines, and being honest about where the technology is being used and where it is not.
The problem is that policy documents are slow and AI is fast. By the time a code of conduct is drafted, consulted on, revised, and adopted, the technology has moved on. The tools that seemed exotic when the drafting began are now standard. The risks that seemed theoretical are now concrete.
This is why the gap between what AI claims and what it does cannot be closed by regulation alone. It has to be closed by practice. It has to be closed by journalists who develop the habit of asking what the tool actually did, not what it appears to have done. It has to be closed by editors who demand explanations, not just output. It has to be closed by newsrooms that treat AI as something to be managed with the same rigour they apply to any other reporting tool.
The EU AI Act is a start. It creates a framework that forces the conversation. But the conversation will only be useful if it is honest, and honesty is exactly what the technology makes difficult. The AI does not lie. It just produces text that looks like the truth, and the humans who use it have to decide, every single time, whether they are going to check or whether they are going to trust.
That decision is the real battleground. Not the policy. Not the code of conduct. Not the guidance note. The decision, made thousands of times a day in newsrooms across the country, about whether to verify what the machine produced or to let it run. Every time a journalist chooses to verify, the gap narrows. Every time they choose to trust, the gap widens. And the gap is where the deception lives.
Professional news journalism is vital for democracy. Credibility is its currency. And credibility is built on a simple promise: that what you are reading has been checked by human eyes and human hands, and that the people who published it can explain why they did what they did. AI does not break that promise by existing. It breaks it when the humans who use it stop asking the hard questions about what the technology actually delivered, rather than what it appeared to deliver.
The gap is not going to close on its own. It is going to close because journalists decide, one story at a time, that the machine’s confidence is not the same as their own verification. That is the only way the deception ends — not with a law, not with a code, not with a policy, but with the quiet, unglamorous, absolutely essential work of checking the work.