Fact-Check – Was a recent speech by Cyril Ramaphosa genuinely AI-generated?
You may have seen a little storm break out online recently over the question of whether the South African government is using artificial intelligence (AI) to write its speeches.
The claim was made by a right-wing podcaster in response to a speech given by President Cyril Ramaphosa in late November about education. In particular, this podcast bro spotted some paragraphs relating to the need to decolonise African education and took to Twitter to opine that the whole speech was AI-generated.
Another chap then jumped on the wagon, ran the speech through a few free AI-detection services, and came back with the finding that in at least one case, there was an 89% chance that the speech had been written by a service like ChatGPT.
Cue outrage online, and tweets like the following presenting this statement as 100% fact: “Ramaphosa’s speech in praise of decolonised education was written by AI”.
Let’s take a step back and look at what was actually going on here.
The speech in question was given by President Ramaphosa on 21 November at the Education International conference in Sandton in Johannesburg. It was the welcoming address to delegates.
The president’s speech on this occasion was, to the say the least, nothing special. When you strip out the greetings to various dignitaries it amounted to 830 words, which is shorter than the average article on Daily Maverick, and all it says, in essence, is that education is important and teachers are important. There’s also a bit where Cyril recommends some tourist destinations to visit in Joburg.
The bit about decolonising African education which seems to have so triggered the podcast bro in question amounts to just 78 words, less than 10% of the total speech.
So in summary, this is an absolute nothing-burger of a speech. It is the kind of virtually meaningless address given by Cabinet members and government officials at these kinds of conferences and events every day, and 99% of the time they don’t make the news because there is frankly nothing of substance within them.
So, some would argue that to pay this much attention to this kind of speech at all is, well, dumb.
But the bigger issue is the idea that you can use free AI detection services to unmask text as being written by AI. This is simply not the case with any reliability whatsoever, as the makers of ChatGPT acknowledged earlier this year when they ditched their own AI detection tool on the grounds that they had realised it simply didn’t work.
In the case of the online sleuths bent on exposing the South African presidency for using ChatGPT, they were running samples of fewer than 100 words through the already unreliable AI detectors, rendering the findings even less trustworthy.
Read more in Daily Maverick: AI – the beginning of the end of the cellphone?
What are the hallmarks of AI-generated text? Well, it tends to be an extremely generic, boring, even-handed word porridge. What are the hallmarks of government speeches? Guess what? Samesies.
The truth is, the genres of government speech-writing and AI-generated text are similar. Neither of them are going to be marked by unusual word choices or controversial thought. Bottom line: We can’t tell with any certainty whether AI wrote Cyril’s speech.
And a question for another day: Would it matter, in any real way, if it did? DM