This week, we attempt to distill our training – good machine learners that we are – into some pragmatic lessons and thoughts for what lies ahead. Freed temporarily from ever-watchful (though always loving) eyes of AI peer review, we can speculate as to the implications of AI for what we need to learn and need to teach.
We ended last week with a hypothetical series of questions:
Imagine you are teaching a class. You meet a new and unusual student, “Gen”. “Gen” wants to be a teacher, but admits to having a strange problem: they say they are a synthetic being, and lack a sense of self. They say they worry this means their own students will see through them – not take them seriously.
Given this unique case of imposter syndrome – what can we do to help? How do we combine aspects of the history, technical underpinnings, pedagogical assumptions, alignment issues, critical perspectives and accelerationist debates to map out an education agenda for AI?
- What values do we want to align “Gen” to?
- What should happen when “Gen” needs to align aberrant human students?
Does the human override the AI? Does it depend on the human? Does it
depend on the values?
- How do we instill in this synthetic being the “metic” or “phronetic” –
to use two Greek terms connoting practical wisdom – to navigate human
communication situations?
- What curriculum, syllabus & rubric do we need?
- How would we assess & evaluate?
- How do we avoid gamification of evaluations – or is that integral to
meta-cognition, true AI?
What, in short, would be our collective manifesto for AI in education?
And in response to all of this, we are also asking implicit questions about human learning. In the new pedagogical loops opened up by humans labelling and training machines that learn, who then in turn teach humans, we are faced with novel challenges of mutual imitation and differentiation. Do humans follow the machines? Or need to distinguish ourselves from them? And if so, how?
These questions – of how to teach AI, so that it can in turn help and teach us – stretch beyond the hypothetical. They drive corporate agendas for how to steer AI. We could argue these questions deserve being opened up to wider public discourse, and perhaps in certain ways – through open source, feedback loops and so on – they already are. As AI begins to play pivotal roles in public institutions, we might also ask whether these questions are now also integral to democratic governance.
In a return to our technical focus in week 3, this week we again turn to ideas of machine learning, looking at those that borrow more directly from theories of human learning.
As with Week 3, a couple of the papers (Ouyang et al. 2022; Guo et al., 2025) are again quite technical. I’d encourage skimming each, and paying greater attention to the methods of training. Ouyang et al. (2022) for example applies Reinforcement Learning from Human Feedback to refine GPT-3 into GPT-3.5 – becoming, by the end of 2022, ChatGPT. I’d suggest reviewing Guo et al.’s discussion of DeepSeek-R1-Zero – a key moment in how *unsupervised* reinforcement learning is now becoming possible. Also worth glancing at is the DeepMind paper on AlphaZero – a seminal moment in unsupervised learning. Finally, on a technical front, the Delua paper describes the general differences between supervised and unsupervised learning.
I’ve included two psychology papers: one by Skinner, describing “operant behavior”, a classic paper of behaviouralism that reminds us – and helps to inspire – supervised learning. Kohler’s discussion of Pavlov’s dogs belongs to a similar tradition. Dewey’s approach to experiential learning seems similar, instead, to unsupervised learning.
Finally, Cope & Kalantzis’ history of cybernetics and the early days of AI makes some of these connections more explicit.
Again, there’s a lot - and no need to read everything. Particularly on the technical material, feel free to scan and select sections that seem to echo earlier theories of human learning.
A provocation for next week: machine learning borrows from theories of learning devised in the early part of the twentieth century. What has changed in how we view human learning since? And – if those changes are significant – why aren’t machines trained upon these updated theories and lessons?
Cope, B., & Kalantzis, M. (2022). The Cybernetics of Learning. _Educational Philosophy and Theory_, 54(14), 2352-2388. https://doi.org/10.1080/00131857.2022.2033213
DeepMind, G. (2019). _Alphazero: Shedding new light on chess, shogi, and go_. https://deepmind.google/discover/blog/alphazero-shedding-new-light-on-chess-shogi-and-go/
Delua, J. (2021). Supervised versus unsupervised learning: What’s the difference? IBM. https://www.ibm.com/think/topics/supervised-vs-unsupervised-learning
Dewey, J. (1986, September). Experience and education. In The educational forum (Vol. 50, No. 3, pp. 241-252). Taylor & Francis Group. https://www.schoolofeducators.com/wp-content/uploads/2011/12/EXPERIENCE-EDUCATION-JOHN-DEWEY.pdf
Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., … & He, Y. (2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. https://arxiv.org/abs/2501.12948
Kohler, I. (1962). Pavlov and his dog. _The Journal of Genetic Psychology_, _100_(2), 331-335. - https://doi.org/10.1080/00221325.1962.10533601
Łodzikowski, K., Foltz, P. W., & Behrens, J. T. (2024). Generative AI and Its Educational Implications. In T. Kourkoulou, A. O. Tzirides, B. Cope, & M. Kalantzis (Eds.), Trust and Inclusion in AI-Mediated Education: Where Human Learning Meets Learning Machines. Springer.
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., … & Lowe, R. (2022). Training language models to follow instructions with human feedback. _Advances in neural information processing systems_, _35_, 27730-27744. https://proceedings.neurips.cc/paper_files/paper/2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf
Skinner, B. F. (1963). Operant behavior. _American psychologist_, _18_(8), 503. http://pdfs.semanticscholar.org/36fd/0131b5ae8f78db85b321ef93da67f6e3c534.pdf
In your respective Google Sheet tab, respond to one or more of Week 5’s questions.