The Policy That Cited Sources That Did Not Exist: AI Governance Lessons from South Africa

A policy document with footnotes under a magnifying glass

South Africa’s draft national artificial intelligence policy was approved by Cabinet in March 2026 and published for public comment in April. Weeks later, reports revealed that several of the academic references cited in the document appeared to be fictitious or attributed to journals that had not published them. In June, Cabinet approved the withdrawal of the draft. A revised policy is now expected to go to Cabinet by March 2027, months later than first promised, and TechCentral reported in late October that the policy team was racing a November 30 deadline for the redrafted text.

The setback raised questions about how generative AI had been used in preparing the document and whether adequate checks were in place before approval. Commentators have noted the irony of a national AI policy undermined by the kind of error that AI tools are known to produce.

A case study for every organization

We do not cite this episode to single out South Africa. Similar failures have embarrassed law firms, consultancies and corporations around the world. The value of the case lies in how clearly it shows the risks that arise when generative AI enters formal decision-making without a verification process.

Fabricated citations are a symptom of a deeper governance gap. Generative models produce fluent text that looks authoritative. If an organization has no rule requiring that every source be checked, no record of where AI was used, and no reviewer accountable for accuracy, errors will eventually reach official documents.

What governance looks like in practice

MBA programs should teach AI governance as an operational discipline, not an abstract ethics module. In practical terms, that includes:

  • Disclosure. Documents should record whether and how AI tools were used in drafting, research or analysis.
  • Verification. Every factual claim and citation in a document intended for decision-making should be checked against the original source by a named person.
  • Clear accountability. AI can assist, but responsibility for a policy, report or contract remains with the humans who sign it.
  • Training. Staff need to understand how generative models fail, including hallucinated facts and invented references, before they are given access to them.

Why this matters for public and private leaders alike

Meanwhile, South African businesses and public institutions continue to adopt AI tools without a formal national policy. This is common across much of the world. Organizations cannot wait for regulation to define good practice; they must govern themselves first.

At EUCLID, students are expected to verify their sources, and every course closes with an oral examination in which they must defend their work in person. We regard these practices as preparation for professional life. The managers and civil servants who will draft tomorrow’s strategies and policies must be fluent with AI and uncompromising about the truth of what they sign.

Sources

aiMBA Insight is published by the EUCLID and EULER MBA programs. Explore our MBA in Artificial Intelligence and Digital Transformation and our other specialized MBAs.

More from aiMBA Insight

Lead the AI transition

Specialized online MBAs from EUCLID and EULER, built for professionals in the public and private sectors.