Since 2 August 2026, companies deploying generative AI systems to publish AI-generated or manipulated text that’s used to inform the public on matters of public interest have been subject to the EU AI Act’s Article 50 transparency requirements.
Such text must be disclosed as AI-generated or manipulated unless it has received substantive human review or editorial control and a natural or legal person holds ultimate editorial responsibility for publication, writes Tamsin Mackay, AI-fluent content, strategy and PR.
The regulation is moving AI writing from a question of which tool a business uses to one of governance and it can be relevant to non-EU organisations, including South African firms, where their AI-related activities or outputs fall within the Act’s territorial scope.
Translated, the regulation is saying that any organisation using AI under its authority (not just OpenAI or Anthropic), be it a bank, retailer, consultancy or publisher, is considered the deployer of the AI-generated content.
Anthropic confirmed that text generated by supported Claude models now carries an invisible, machine-readable watermark that is being applied globally. It doesn’t change the wording that the reader sees and can’t be traced back to them, but it does travel with the text if it is copied and pasted, even if it has been edited. Google’s SynthID and the C2PA provenance standard come at this problem from other directions, but across the board, AI is gaining scrutiny. If a piece has been substantively drafted and published without human oversight, it will increasingly trigger disclosure requirements.
The problem is that none of these mechanisms can establish authorship. Claude’s watermark only shows that it was involved at some point and it also can’t distinguish between the text written by the model or text that person has heavily edited. Currently, these tools just show that the content was processed by a machine not the human thinking behind it. And this opens up a gap in governance that companies need to address for several increasingly important reasons.
The first is verification. Gen AI produces content with the same fluency and confidence whether or not it is accurate. It can invent a statistic, misattribute a quote or cite a study that never existed and present these in relatively well-structured prose as if they are a given. People read and believe the data because, until AI’s fake information pervasively took over the internet, the written word was largely based on fact. A person with subject knowledge tends to stop at the point where the information doesn’t hold together, but the AI doesn’t have this ability. Removing human review doesn’t speed things up, it simply removes a step that tests whether or not the work is true so when the error is published, the consequence sits with the business.
Published output depends on a human who will check the sources and ensure that the facts shared by the AI are, in fact, real.
The second part of the work is judgement. AI-assisted writing tends to be fluently generic, producing work in the shape of an argument with very little behind it. The equivalent of eating candy floss when you’re hungry – the AI kind of fills you up for a while, but then you’re left feeling empty.
The models are trained on the collective output of other companies so the writing converges on a shared register and material produced this way sounds like everybody else’s writing. Audiences respond by reading it less closely and the result is a slower and more expensive form of damage. The limp quality of the writing wears away at the credibility that gave an organisation’s communication its value.
Seen in this light, provenance and watermarking are necessary but their limitations are where governance needs to step in. Does your company want to build and maintain a reputation for producing content where judgement was applied and it is worth reading? Have you put processes in place that ensure the content is accurate, trusted and distinct from everything else in the market?
While governance around the use of AI in South Africa is very likely to lag behind the EU for some time, companies that stand on clearly defined governance foundations are moving themselves out of the generative firing line. Ultimately, the value of content isn’t how much of it you can throw onto the internet with the hope of making it stick, but how well your audience responds to its value. Are you feeding readers with candyfloss or a roast dinner?
Learning how to use the AI more effectively and mitigate the damage of fake facts, generic slop and poorly structured content is both essential and easy. Employees can learn how to detect when a claim doesn’t hold together or when the writing has slipped into generic filler or how much of a piece is filled with slop-like tells.
Prompting is a linguistic skill and it can be taught alongside AI literacy through hands-on practice and training. Done well, it equips people at every level to work with AI safely inside the organisation’s own secure environment and builds governance today, at the edge of the Wild AI frontier.