AI Hallucinations Have Already Disqualified a Tender Bid. South African Businesses Could Be Next.

In November 2025, a Polish contractor lost a contract worth approximately €3.7 million (PLN 15.5 million) because its bid submission contained AI-generated content that cited nonexistent documents. The ruling was an oversight finding, not fraud. Intent was not a defence. That case went largely unnoticed in South Africa. It should not have.

All court cases are verifiable on SAFLII (saflii.org). Government incidents are sourced from official statements on gov.za and sanews.gov.za.

The Pattern South Africa Has Spent Twelve Months Establishing

In the twelve months between January 2025 and May 2026, South Africa recorded five separate, documented incidents in which unverified AI-generated content found its way into official documents — two High Court judgments, a Cabinet-approved immigration white paper, a national AI policy, and an investigation into a sitting judge. In each case, the same pattern: AI produced a plausible output, a human accepted it without verification, and an official document was compromised.

1. Mavundla v MEC (January 2025)

A candidate attorney submitted a supplementary notice of appeal citing nine case authorities. Seven did not exist. Judge Bezuidenhout held that relying on AI without verification is "irresponsible and unprofessional" and referred the matter to the Legal Practice Council. (SAFLII: [2025] ZAKZPHC 2)

2. Northbound Processing v SA Diamond Regulator (June 2025)

Junior counsel used an AI tool called "Legal Genius" — described as exclusively trained on South African legal judgments — which produced two fictitious case citations. Acting Judge Smit referred the conduct to the Legal Practice Council and stated: "coherent and plausible outputs are not sufficient if they are false." (SAFLII: [2025] ZAGPJHC 661)

3. National AI Policy Withdrawn (April 2026)

Minister Malatsi withdrew South Africa's Draft National Artificial Intelligence Policy after News24 found at least 6 of 67 academic citations were hallucinated. The minister called it "an unacceptable lapse that compromised the integrity and credibility of the draft policy." (SAnews.gov.za, 26 April 2026)

4. Home Affairs Officials Suspended (April 2026)

Two senior officials were suspended after AI-generated references were found in a Cabinet-approved white paper — 102 of 148 references were unverifiable or fictitious. The department appointed two independent law firms and committed to reviewing all policy documents produced since 30 November 2022, the date ChatGPT was released. (Official DHA statement, gov.za, 30 April 2026)

5. Johannesburg Acting Judge Under Investigation (May 2026)

A Johannesburg High Court acting judge is under investigation for allegedly using AI to draft a judgment citing a non-existent case. The appellant's AI detection tool found the judgment carried "high risk in terms of citation integrity." (Legalbrief, 18 May 2026)

The Specific Legal Risk for Tender Submissions

When a company submits a government tender, it signs Standard Bidding Documents that are legal declarations. Since their consolidation by National Treasury in 2022, the key document is the SBD 4 (Bidder's Disclosure), which covers declarations of interest, past SCM practices, and independent bid determination. The information is sworn to be accurate.

If an AI tool generates a requirement that does not exist, a regulation that has been misquoted, or a capability claim that is inaccurate — and the bidder submits without verification — the legal exposure is significant. Immediate disqualification is the minimum consequence. In serious cases, a finding of misrepresentation can lead to criminal prosecution under PRECCA and, following conviction, listing on the national Register for Tender Defaulters — which restricts the company and its directors from public procurement for years.

Since 3 April 2024, Section 34A of PRECCA makes companies criminally liable for corrupt activities by persons associated with them if the company failed to have adequate prevention procedures. A company with no verification protocols for AI-generated tender content may find Section 34A relevant.

The Distinction That Matters

The argument is not that AI should be excluded from tender preparation. The argument is that general-purpose AI tools — ChatGPT, Gemini, Grok — were not built for South African procurement. They predict statistically likely language. They do not know your B-BBEE level, your CIDB grade, your CSD status, or the specific requirements of your tender document.

Purpose-built tools read the actual document, validate outputs against the requirements found there, flag low-confidence outputs for human review, and maintain an immutable audit trail. The human remains in the loop at every decision point. That is compliance-centric AI — and it is the only approach that does not carry the legal risks documented above.

⚠️ The Polish contractor did not set out to mislead anyone. Neither did the Home Affairs officials or the legal teams in Mavundla and Northbound. Intent has not been a defence in any of these matters. Verification is the only defence that has worked.