EU AI Act transparency rules take effect as South African newsrooms scramble for policies

Transparency obligations under the European Union's Artificial Intelligence Act entered force this month, marking the first time a major jurisdiction has legally compelled media organisations to disclose when and how they use AI in content production. The regulation, which the European Union adopted after years of negotiation, does not prohibit algorithmic assistance in journalism. It does, however, impose labelling requirements, liability for harmful output, and a duty to demonstrate editorial oversight, provisions that are already reverberating in newsrooms far beyond the EU's borders.

What the EU law actually requires of media

The AI Act classifies certain media applications as high-risk, particularly those involving content distribution and audience engagement at scale. Providers must implement risk management systems, maintain technical documentation, and ensure human oversight throughout the lifecycle of any AI system they deploy. For news organisations, this translates into a practical obligation: every article, image, or broadcast segment shaped by generative tools must be traceable to a responsible editor, and audiences must be informed when synthetic content appears. Failure to comply carries fines of up to 35 million euro or seven percent of global annual turnover, whichever is higher.

Hendrik Sittig, who directs the Konrad-Adenauer-Stiftung's media programme for sub-Saharan Africa, emphasised that the legislation is not a ban. "The EU AI Act does not force newsrooms to stop using AI. Instead, it requires greater transparency, clear labelling and accountable editorial oversight," he said. The distinction matters because many South African editors initially interpreted the European move as a restriction rather than a framework for responsible use.

South Africa's regulatory patchwork

South Africa lacks a single AI statute. The Independent Communications Authority of South Africa regulates broadcasting spectrum and electronic communications, while the Broadcasting Complaints Commission and the Press Council of South Africa adjudicate ethical disputes. The Press Council, which oversees the Press Code of Ethics and Conduct for print and online media, has been monitoring AI's impact since generative tools entered newsrooms. Its 2023 guidance note urged editors to be "thoughtful when they deploy new AI tools" and insisted that AI-generated material be checked by "human eyes and hands" before publication. That note carries no binding force.

Phathiswa Magopeni, the Press Council's executive director, frames the challenge in ethical rather than technological terms. "It's basically about verification, accuracy and editorial due diligence," she said. For her, the two priorities are disclosure and explainability. Images sourced from agencies routinely carry credits, but AI's role can be invisible when embedded in research, translation, or image generation workflows. "We need to determine levels of disclosure because AI may be used at various stages of the production process, for example writing, research and the generation of images," Magopeni added.

Adoption without policy: the CINIA findings

A study published this month by the Centre for Information Integrity in Africa (CINIA) surveyed South African newsrooms and found what its director, Professor Herman Wasserman, described as a "fragmented and nascent policy landscape." Few organisations have written AI policies. Adoption is driven by individual journalists with a personal interest in the technology, who then informally teach colleagues. Some respondents told researchers they suspected peers were using AI "secretly." Others simply did not know what was permitted.

Wasserman acknowledged "great enthusiasm among South African journalists about the potential and possibilities of AI for journalism" but warned that guidelines for ethical and responsible use are absent. Training is scarce. The result is a newsroom culture where experimentation happens in the open but accountability structures have not caught up.

The language gap that European models cannot fill

A recurring complaint in the CINIA study concerned linguistic competence. Journalists reported that generative tools "mistranslate political or cultural terms" because the large language models underpinning them are trained predominantly on English-language data from the global north. South Africa has eleven official languages; most commercial models perform poorly on isiZulu, isiXhosa, Afrikaans, and others. This is not a minor inconvenience. In a country where language carries political weight, a mistranslated quote or a hallucinated cultural reference can inflame tensions or distort public understanding.

African developers are responding by building large language models trained on local corpora. These efforts remain early-stage and under-resourced compared to the compute budgets of OpenAI, Google, or Meta. Until they mature, South African newsrooms using off-the-shelf tools are effectively outsourcing editorial judgment to systems that do not understand the context in which they operate.

Professional standards versus platform speed

Legacy broadcasters and publishers in South Africa face the same commercial pressures as their European counterparts: declining advertising revenue, audience migration to social platforms, and newsroom headcount reductions. Automation promises speed and scale. But unlike social media platforms, which operate with minimal content liability in most jurisdictions, professional broadcasters and publishers are bound by broadcasting codes and press ethics. The EU AI Act widens that asymmetry by codifying transparency obligations for professional media while platform governance remains fragmented.

To date, the most damaging AI-generated falsehoods in South Africa have circulated on social media, deepfakes endorsing political candidates during election periods, rather than on regulated broadcast networks. An energetic but underfunded fact-checking community has contained the fallout. Editors know that luck will not hold if AI integration proceeds without guardrails.

When institutions hallucinate

Two recent episodes underscore the risk. The Department of Communications and Digital Technologies published a draft National AI Policy that contained fabricated citations generated by an AI tool. The document was withdrawn after experts identified the fake references. More recently, an acting judge in Johannesburg issued a ruling that appeared to rely on non-existent case law, prompting scrutiny of whether generative AI had been used in drafting the judgment. Both cases illustrate what happens when institutions adopt the technology without verification protocols.

Towards a binding AI press code

The Press Council is now consulting on a dedicated AI press code that would replace the 2023 guidance note with enforceable rules. Magopeni said the industry must decide what level of disclosure audiences expect at each stage of production, research, drafting, editing, illustration, distribution, and what constitutes adequate human oversight. The council also plans to consult audiences directly on their tolerance thresholds, an approach endorsed by CINIA's Rejoice Malisa-van der Walt, who argues that regulation must be "dynamic and inclusive" as the technology advances.