ANTHONY BUTLER | What if AI trains using Mbeki’s speeches?
Argument in a report by former president’s foundation ignores the real issues
First published in Business Day and BusinessLive
1 October 2026
Amid all the brouhaha about the threat that Artificial Intelligence (AI) poses to the future of human life, one even more frightening prospect has hitherto been overlooked. Rather than killing us, AI may instead subject us in perpetuity to the eternal guidance of large language models trained on the collected speeches and writings of former president Thabo Mbeki.
This terrifying spectre has been raised by the Thabo Mbeki Foundation’s release of a new report entitled The United States of America and the Future of Afrikaners in South Africa.
A pithy 123 pages long, and so relatively condensed for a contribution to the Mbeki oeuvre, the paper reinterprets the current disputes between the two countries as a manifestation of the “national question” in both countries.
In Mbeki’s usage, the national question refers to the resolution of colonial conquest’s legacy by freeing the dispossessed indigenous majority and building one united, non-racial nation where race no longer determines wealth or opportunity. The basic parallel the report draws is that the US has faced its own national question exhibited in slavery, segregation and the continuing socioeconomic marginalisation of Black Americans.
From this perspective, the Foundation argues that it is historically inconsistent for Washington to demand that SA abandon racial redress only three decades after 1994 when the US continued to use affirmative-action and other remedial policies long after emancipation.
This is not a very good argument, of course, which is one reason we can speculate that Mbeki played a personal role in the preparation of the report.
There is further evidence of Mbeki’s hand in accounts of transition era negotiations with Constand Viljoen which seem to be presented from his personal perspective, and the text slips into the first person singular when describing Mbeki’s interactions with Inkatha Freedom Party leader Mangosuthu Buthlezi.
To its credit, the foundation assembles empirical material to contest American misconceptions about SA, deploying the analyses of notable Afrikaners Frans Cronje and Herman Giliomee to argue that Afrikaner living standards rose faster between 1994 and 2004 than during any comparable 20-year period since 1945.
This certainly suggests that a genocide or persecution narrative cannot plausibly describe the socioeconomic condition of Afrikaners as a whole (but then we all knew that already).
The report then veers rather alarmingly in an Mbekian direction by equating the building of Afrikaner community institutions with a revival of apartheid-style “separate development”. It also speculates that the current US administration has aligned itself with a “white supremacist faction” in SA, citing as evidence US connections of Solidarity and AfriForum, the refugee programme, and opposition to land reform. There is evidence for each element of this narrative, it is true, but this does not amount to proof of such an alignment.
For now, Mbeki Foundation’s work still clearly draws on Mbeki’s intellect rather than on AI. A sophisticated language model would usefully have placed more weight on Trump’s domestic political coalition, US polarisation over racial preferences, immigration politics, evangelical and conservative networks, SA foreign policy with regard to China and Iran, US commercial interests, and SA’s involvement in the International Court of Justice against the government of Israel – all matters elided by Mbeki’s abstract and characteristically paranoid theoretical generalisations.
It is therefore essential that any future global AI safety plan prohibits the training of language models on archival materials gathered by the Mbeki Foundation. In the centuries ahead, when human beings send robots between the stars in fleets of interstellar rockets, we cannot take the risk that their interactions with alien civilisations will be mediated by Mbeki-trained embodied AI. Such robots will surely antagonise Daleks, Vulcans and Klingons by presenting reports about how to avert threats to the national democratic revolution, and by ruminating at great length on the application of the national question to human-alien relations
Butler teaches public policy at the University of Cape Town
