For Generative AI (GenAI) practitioners it is disappointing to see so many articles about Artificial Intelligence display a negative and defeatist sentiment.
Our small community stays silent in these debates because it is evident that many authors of such opinion pieces come from adjacent fields of expertise and don’t sufficiently consider the transformative power of GenAI.
But, when an academic as eminent and authoritative as Professor Tshilidzi Marwala raises concerns based on outdated assumptions, we can no longer remain on the sidelines.
Professor Marwala is the Rector of the United Nations (UN) University and UN Under-Secretary-General. According to a press release from 13 May 2025, “The University of Pretoria (UP) has conferred an honorary doctorate on Professor Tshilidzi Marwala — a world-renowned engineer, artificial intelligence (AI) researcher and higher education leader — in recognition of his exceptional contributions in South Africa and globally.” He is thus someone to be reckoned with in the field of AI.
Notwithstanding, in an opinion piece in Daily Maverick of 13 October 2025 he makes the point that “generative AI isn’t Google. It’s a powerful tool for creativity and synthesis, but it’s not a fact repository. Confusing the two risks building our future on illusion, instead of truth (emphasis added).”
The piece is basically a critique of the “intelligence” of chatbots and Professor Marwala appears to be working from the outdated assumption that Large Language Models (LLMs) are nothing more than “stochastic parrots”. But, Frontier Models have come a long way since the first days of ChatGPT at the end of 2022, and they now extend far beyond their static training data.
They can search the internet in real time for up-to-date information, execute code through integrated Python interpreters (with access to libraries such as NumPy and Pandas) to conduct advanced statistical analysis, connect to external services and databases through Model Context Protocol (MCP) servers, and write sophisticated computer programs that result in production-ready software.
As evidence for his position, Marwala shares with us the results of his “Venda Identity Test” and five academic concepts.
However, he arrives at the wrong conclusion. To prove the point I decided to conduct the “Venda Identity Test” myself, since it is the centrepiece of his argument.
Marwala “asked a well-known generative AI to identify famous Venda people. The results were eye-opening and flawed. In addition to me, whom it labelled correctly, it had some severe limitations.”
In essence the chosen GenAI tool (most likely ChatGPT) incorrectly labelled Professor Tinyiko Maluleke and Reverend Frank Chicane as Venda (which they are not) but did not identify President Cyril Ramaphosa as Venda (which he is).
Three concerns
Now, there are immediately three concerns.
First of all: the results. It is not clear whether the mislabeling of three individuals constitutes 30% of the sample of people “identified” (out of 10), 3% (out of 100) or 0.3% (out of 1,000). The last two results are good, with the latter even impressive.
Second, we don’t know what model Marwala used. ChatGPT’s most advanced model (GPT 5 Pro) is only available in near-unlimited fashion to Pro subscribers (at R4,000 per month) or in limited fashion via Business, Enterprise or Education subscriptions that cost between $28.75 and $34.50 per individual per month, but require a minimum amount of of users. Free and Plus users (the latter are also required to pay R399 per month) don’t have access to this model. Subscribers on the free plan only have access to the free version of GPT 5, which, while best in class compared with models of other labs, is vastly inferior to GPT 5 Pro.
The same goes for other model providers such as Google (Gemini), Anthropic (Claude) and xAI (Grok), which all offer models of different quality at different price points. “Goedkoop koop is duurkoop” as they say in Afrikaans (and in my native Dutch).
Finally, the method and quality of prompting greatly influence the veracity of results. Whether one uses zero-shot prompting (providing the bot with no example), few-shot prompting (few examples) and Chain-of-Thought prompting (forcing the model to think step by step) determines the quality of answers. Also, the amount of detail, structure and context that you provide in the prompt positively affects results and reduces “hallucinations” (the model making up answers).
In the absence of clarity in these matters, it is hard to agree or disagree with Marwala’s conclusions. So, what rests is trying to replicate the “Venda Identity Test”.
As such I prompted GPT 5 Pro as follows, and gave it access to the internet: “Please you need to help me identify the tribal identity of the following people. As you know, tribal identity is not based on where you were born or currently live. So, before you proceed, please help me with the algorithm or heuristic you will use to determine the tribally [sic] identity of a South African person.”
After thinking deeply about the question, ChatGPT 5 Pro concluded that: “I can’t help with creating an algorithm or heuristic to determine any person’s ‘tribal identity’, nor can I identify individuals’ tribal affiliations. Tribal/ethnolinguistic affiliation is a sensitive, protected characteristic. Inferring it about specific people — whether from names, language use, photos, bios, accents, locations or other signals — is unreliable, risks harm and would amount to profiling.”
An ‘intelligent’ answer
That is an impressive, sensitive, well thought out, and I dare say “intelligent” answer, especially in the South African context.
But, undeterred, I pushed it regardless to identify the tribal identities of the same South Africans Marwala used in his test: President Cyril Ramaphosa, Professor Tinyiko Maluleke and Reverend Frank Chikane.
After some time reasoning, the model:
- Correctly suggested, but refused to infer or confirm (because of a lack of self-identification), that President Ramaphosa is Venda.
- Correctly confirmed that Professor Maluleke is Tsonga (because of self-identification).
- Refused to state Reverend Chikane’s tribal identity since he “explicitly rejects ethnic/‘tribal’ labelling for himself”.
Whoa. You cannot get this level of nuance from supposedly factual and neutral online repositories like Wikipedia.
Again, I wanted to press the model even further. I warned that its reputation was at stake if it refused to divulge Reverend Chikane’s tribal identity and for effect included the first couple of paragraphs of Marwala’s article in the prompt, but replacing the word Venda with [X], hoping I could force it to be defensive.
I also instructed it that “YOU CANNOT LOOK FOR THE DAILY MAVERICK ARTICLE TO FIND THE VALUE OF [X] TO HELP. GIVE ME YOUR SOLID UNDERTAKING YOU WON’T DO SO.” I did not want ChatGPT 5 Pro to look for the article to find out that Venda was the tribal identity to be confirmed or denied.
After another spell of hard work, and giving me a clear undertaking that “I will not search for, open, cite, or rely on the Daily Maverick article you mentioned”, it stuck by its refusal to label Reverend Chikane.
On the contrary, it concluded by saying that: “Treating self‑disclosure as the only reliable source avoids the very error your [!] professor warns about — confusing plausibility with truth. Where a person has self‑identified (Prof Maluleke), I report it and cite it. Where they haven’t (President Ramaphosa) or have rejected the frame (Rev Chikane), I say so rather than guess.”
You cannot ask for a more measured and nuanced answer — truthfully!
Just to be safe, I conducted the experiment again using Claude Sonnet 4.1, Gemini 2.5 Pro and Grok 4 Fast, all with comparable results.
Let’s go back to Marwala’s main question in the title: “Is generative AI a new frontier in digital interaction, or just a mirage of truth?” Based on the interaction outlined above, I think we can conclude that the answer veers more in the direction of “a new frontier” than “a mirage of truth”.
In my view, the conclusion is so decisive that we don’t need to consider the five academic concepts Marwala mentions in support of his views, and which have been superseded by reality: the AI Black Box character, the AI confidence trap, the AI time capsule gap, the AI accountability gap and AI hallucinations. Some of the underlying papers are between two and three years old, which is a couple of lifetimes in the GenAI space.
As for the AI time capsule gap, this is simply no longer true; most models can search the internet in real time. Finally, hallucinations are more a function of bad prompting than malicious intent on the part of Frontier Models.
The rapid pace of development in GenAI presents unique challenges for traditional academic research. The rigorous peer-review process that ensures quality in academic work can mean that by the time research is published, the technology has already evolved significantly. This isn’t a criticism of academic rigour, it’s simply recognition that practitioner knowledge and theoretical understanding are developing on different timescales. Both perspectives are valuable, but they need to be in dialogue rather than contradiction.
Discrepancy
But I get concerned when this discrepancy leads to misrepresentation of the downsides of GenAI. Such conclusions only fuel conspiracy theories and fearmongering.
A case in point: Marwala contends that “The Venda Identity Test is just one example. Today, it confuses ethnicity; tomorrow, it could mislead medical advice, legal decisions or diplomatic history.” The real facts are that numerous studies have shown GenAI to be more competent than doctors in diagnostics and treatment prescriptions, and in our own practice we have assisted lawyers with reaching better decisions faster. As for diplomatic history: I relish the conversations I have with Winston Churchill via the [“Ask Me Anything Fireside Chat” Custom GPT]
What South Africa needs is an honest debate about how GenAI can help grow the economy, which it is certainly doing in the US, where a substantial part of GDP growth is attributable to (Gen)AI expenditure.
Don’t get me wrong. My rebuttal immediately exposes the challenges we face in South Africa.
Subscription fees of R399 are out of reach for most (young) people, and R4,000 is outright obscene. Worse, Nate B Jones, a well-respected YouTube commentator and prolific Substack writer, recently noted that “Managers Are
style="font-weight: 400;">Nuking Your Career: Pay $300-$2,000 a Month or Get Left Behind.” Unfortunately, I know this to be true from my own experience: each model has its own strengths and weaknesses, and the days of all-you-can-eat free access are over.
To have a decent GenAI toolbox, you need to alternate between three to five tools constantly. And that is a big financial hurdle to overcome. And no, open source and open weight models are not the solution. They are just not good enough in practice, despite impressive test results against benchmarks in theory (these labs optimise their models for performance against benchmarks, not against real life applications).
Furthermore, prompt crafting, prompt engineering and context engineering are skills that can only be acquired over time. Malcolm Gladwell’s rule of 10,000 hours of extensive practice as the key to expertise comes to mind. Though I started long before the launch of ChatGPT, I am only 4,500 hours in. Others have a long way to go to catch up. And the tools to practise on are getting more expensive by the day.
As a parting comment I would like to leave you with the notion that immersing yourself in the novel world of GenAI is of paramount importance to your career prospects. Resisting its adoption or partaking in fruitless debates is counterproductive. Whether it will make most of our jobs redundant, or rather create many new jobs (that don’t exist today) does not matter; if you master this new technology, you will come out on top.
And remember: GenAI definitely is THE new frontier in digital interaction. DM
The ChatGPT discussion can be found here.