A month or two ago I shared a "warts-and-all" preprint of this paper on the risks of educational uses of generative AIs. The revised, open-access published version, The Human Nature of Generative AIs and the Technological Nature of Humanity: Implications for Education is now available in the Journal Digital.

The process has been a little fraught. Two reviewers really liked the paper and suggested minimal but worthwhile changes. One quite liked it but had a few reasonable suggestions for improvements that mostly helped to make the paper better. The fourth, though, was bothersome in many ways, and clearly wanted me to write a completely different paper altogether. Despite this, I did most of what they asked, even though some of the changes, in my opinion, made the paper a bit worse. However, I drew the line at the point that they demanded (without giving any reason) that I should refer to 8 very mediocre, forgettable, cookie cutter computer science papers which, on closer inspection, had all clearly been written by the reviewer or their team. The big problem I had with this was not so much the poor quality of the papers, nor even the blatant nepotism/self-promotion of the demand, but the fact that none were in any conceivable way relevant to mine, apart from being about AI: they were about algorithm-tweaking, mostly in the context of traffic movements in cities.  It was as ridiculous as a reviewer of a work on Elizabethan literature requiring the author to refer to papers on slightly more efficient manufacturing processes for staples. Though it is normal and acceptable for reviewers to suggest reference to their own papers when it would clearly lead to improvements, this was an utterly shameless abuse of power of a scale and kind that I have never seen before. I politely refused, making it clear that I was on to their game but not directly calling them out on it.

In retrospect, I slightly regret not calling them out. For a grizzly old researcher like me who could probably find another publisher without too much hassle, it doesn't matter much if I upset a reviewer enough to make them reject my paper. However, for early-career researchers stuck in the publish-or-perish cycle, it would be very much harder to say no. This kind of behaviour is harmful for the author, the publisher, the reader, and the collective intelligence of the human race. The fact that the reviewer was so desperate to get a few more citations for their own team with so little regard for quality or relevance seems to me to be a poor reflection on them and their institution but, more so, a damning indictment of a broken system of academic publishing, and of the reward systems driving academic promotion and recognition. I do blame the reviewer, but I understand the pressures they might have been under to do such a blatantly immoral thing.

As it happens, my paper has more than a thing or two to say about this kind of McNamara phenomenon, whereby the means used to measure success in a system become its purpose, because it is among the main reasons that generative AIs pose such a threat. It is easy to forget that the ways we establish goals and measure success in educational systems are no more than signals of a much more complex phenomenon with far more expansive goals that are concerned with helping humans to be, individually and in their cultures and societies, as much as with helping them to do particular things. Generative AIs are great at both generating and displaying those signals - better than most humans in many cases - but that's all they do: the signals signify nothing. For well-defined tasks with well-defined goals they provide a lot of opportunities for cost-saving, quality improvement, and efficiency and, in many occupations, that can be really useful. If you want to quickly generate some high quality advertising copy, the intent of which is to sell a product, then it makes good sense to use a generative AI. Not so much in education, though, where it is too easy to forget that learning objectives, learning outcomes, grades, credentials, and so on are not the purposes of learning but just means for and signals of achieving them.

Though there are other big reasons to be very concerned about using generative AIs in education, some of which I explore in the paper, this particular problem is not so much with the AIs themselves as with the technological systems into which they are, piecemeal, inserted. It's a problem with thinking locally, not globally; of focusing on one part of the technology assembly without acknowledging its role in the whole. Generative AIs could, right now and with little assistance,  perform almost every measurable task in an educational system from (for students) producing essays and exam answers, to (for teachers) writing activities and assignments, or acting as personal tutors. They could do so better than most people. If that is all that matters to us then we might as well therefore remove the teachers and the students from the system because, quite frankly, they only get in the way. This absurd outcome is more or less exactly the end game that will occur if we don't rethink (or double down on existing rethinking of) how education should work and what it is for, beyond the signals that we usually use to evaluate success or intent. Just thinking of ways to use generative AIs to improve our teaching is well-meaning, but it risks destroying the woods by focusing on the trees. We really need to step back a bit and think of why we bother in the first place.

For more on this, and for my tentative partial solutions to these and other related problems, do read the paper!

Abstract and citation

This paper analyzes the ways that the widespread use of generative AIs (GAIs) in education and, more broadly, in contributing to and reflecting the collective intelligence of our species, can and will change us. Methodologically, the paper applies a theoretical model and grounded argument to present a case that GAIs are different in kind from all previous technologies. The model extends Brian Arthur’s insights into the nature of technologies as the orchestration of phenomena to our use by explaining the nature of humans’ participation in their enactment, whether as part of the orchestration (hard technique, where our roles must be performed correctly) or as orchestrators of phenomena (soft technique, performed creatively or idiosyncratically). Education may be seen as a technological process for developing these soft and hard techniques in humans to participate in the technologies, and thus the collective intelligence, of our cultures. Unlike all earlier technologies, by embodying that collective intelligence themselves, GAIs can closely emulate and implement not only the hard technique but also the soft that, until now, was humanity’s sole domain; the very things that technologies enabled us to do can now be done by the technologies themselves. Because they replace things that learners have to do in order to learn and that teachers must do in order to teach, the consequences for what, how, and even whether learning occurs are profound. The paper explores some of these consequences and concludes with theoretically informed approaches that may help us to avert some dangers while benefiting from the strengths of generative AIs. Its distinctive contributions include a novel means of understanding the distinctive differences between GAIs and all other technologies, a characterization of the nature of generative AIs as collectives (forms of collective intelligence), reasons to avoid the use of GAIs to replace teachers, and a theoretically grounded framework to guide adoption of generative AIs in education.