In it they warned of various risks and consequences they could see on the horizon as Large Language Models (LLMs) of AI (such as ChatGPT) gained traction outside the computer laboratories. These included the environmental costs of the processing involved in such technologies as well as the fact that, by its very nature, such a system ‘heightens the potential for automation bias, deliberate misuse, and amplification of a hegemonic worldview’.
They also highlight that such models give the appearance of genuine human communication but are nothing more than clever machines working out what word should probably come next and giving it a go. In other words, LLMs are, in their memorable and controversial (Gebru lost her job shortly after) term, nothing more than ‘stochastic parrots’.
What the parrot knows
If you are like me, you will have to look up ‘stochastic’ and, even then, the left-click ‘Look up’ function may still leave you cold (‘having a random probability distribution’). So, like me, perhaps you will turn to the parrot itself and ask for the sort of definition that would satisfy a ten-year-old.
‘Sure!’, it chirps, as confident and chipper as ever. ‘It means “random-ish”,’ it suggests helpfully, before adding, ‘It’s a nickname for AI language models (like me).’
So, the parrot knows it a parrot.
Except the parrot doesn’t actually know anything. It just regurgitates one word at a time all the knowledge, flotsam and jetsam it has scraped from the web in a facsimile of human communication to produce, with unswerving confidence, sentences and paragraphs that, while not being random, are, well, random-ish.
AI and ‘reasonable proficiency’
That infamous 2020 paper was mentioned in a recent essay called ‘The Future of Search’ in the London Review of Books by another professor, this time a Professor Donald Mackenzie from Edinburgh University. In it he suggests:
For anyone who, like me, teaches a subject in which students are no longer assessed primarily by means of traditional exams, the most pressing concern is these (AI) models’ ability to generate essays that read like the work of a reasonably proficient if intellectually unambitious student.
A more disturbing thought is that the models’ capacity to do this may have revealed that something is wrong with our pedagogy.
Have we been teaching students to be stochastic parrots?
I suspect his concerns are valid.
My initial foray into the world of education was as a French teacher. I identified early on in my training that we were in danger of teaching students what I called ‘phonic blobs’.
One lump of sound meant, ‘I am 14-years-old’. Another meant ‘Where is the ironmongers?’. You could even append your own name to a lump of sound to tell your imaginary pen pal what to call you. Memorise as many lumps of sounds as possible and parrot them in the right order at the right time and, voilà, Polly’s a French speaker.
On paper at least.
Assessment and ‘lived experience’
ChatGPT ends its simplified definition for simple folk like me by explaining that the term ‘stochastic parrots’ highlights the fact that AI, ‘doesn’t understand the world the way humans do — it’s using patterns, not lived experience’.
How much of our current assessment system – essays or otherwise – is simply about pattern recognition and repetition rather than lived experience? Or, to put it another way, how can we ensure that young people have a lived experience worthy of academic assessment in some form or another?
The IB’s Systems Transformation: a way forward?
The day after I read the Professor MacKenzies’ article, a friend sent me a video about a new IB DP subject being trialled in a small number of schools, including Mulgrave School in West Vancouver, Canada.
Called Systems Transformation, it is an exciting transdisciplinary real-world course that takes the learning out of the classroom and applies it to challenges that are both genuine and pressing. And there are plenty of them around, wherever your school is situated.
It is a programme that clearly has the potential to transform the nature of how young people learn as well as how they are assessed. After all, when you are working in real-life situations with multiple agencies and alongside indigenous peoples addressing real-world challenges, there is neither the room nor the need for the random-ish parroting of facts.