A Chat with ChatGPT by a veteran AI researcher
John Self entered the field of AI in the early 1970s. He’s a distinguished scholar who can claim to have introduced the idea of user model in his 1974 article (while visiting The University of Melbourne). He was writing in the context of AI in Education, a field in which he was a pioneer and long-time leader. Released in 2005, his last AI book is humane and broad, on open access and an excellent read: Whoever Said Computers Would be Intelligent
John generously hosted what was for me an important sabbatical, in Lancaster in 1988. For a couple of decades back then my research was as a psychologist in AIEd. Many in that field (not John, and not some of the other leading lights) were IT tragics with a view that human learning was little more than turning on the tap to fill the bucket. John’s landmark contribution in 1974 was to recognise that any IT (‘intelligent tutor’, the pretentious term back then) that individualised its response to a student must contain a model of that student. The crudest might be merely a note of where that student was up to in the book. Typically it would comprise a record of previous student work, correct responses and errors, and earlier comments by the IT.
Early Intelligent Tutors (ITs)
Many of those early ITs supported learners to achieve impressive gains on tests, especially in science and computing fields. Interactions resembled those in ‘direct instruction’ classrooms. Direct instruction is highly structured and is coming back into vogue in many countries, often for early reading and numeracy. However, then as now, many students find the interactions stultifying. Grit your teeth and use an IT to quickly get up to speed with LISP! But a fulfilling education, perhaps not 🙁
Learner Models in ITs
My role in AIEd was often to be the maverick critic, despairing of the poverty of typical learner models. I and colleagues studied transcripts of tutoring interactions of expert teachers with individual learners when given occasional chances to interact. Think of a teacher wandering in a class working as individuals. The teacher has a brief interaction with an individual, initiated by a student requesting help, or the teacher walking by and choosing to interrupt.
Human Tutoring and Learner Models
Not surprisingly, the interactions were often brief, just a single Q&A in either direction, or little more. But they were highly diverse. The Q might be about motivation, feelings, seemingly irrelevant things that affect the learner’s work and thinking. Brief explanations can of course be valuable, but possibly the most valuable comments were often much higher-level, about strategy, or motivation. “Inspiring” is possibly the most valuable teacher ability, as hopefully we all remember. Human teachers have learner models for their individual learners. These models are likely fragmentary in many respects, but they are, most importantly, highly diverse. Achieving this richness was the enormous challenge for IT researchers seeking to model good tutoring. It remains a core challenge for any AI intended to be used by a person.
John the Fell Runner and Hill Walker
For decades John was an immensely fit fell runner, spending hours and days in the hills of North-West England. He stopped running in 2017, then from 2018 has been posting online reflections and great photos from his Saunterings in the hills, dales, and moors near Lancaster and surrounds.
ChatGPT’s User Model
John started by asking “What do you think of Saunterings by John Self?” I’m guessing John had front of mind trying to diagnose ChatGPT’s model of its user–John. What did it know about, what did it assume about John? John gives his own commentary, commenting on the AI responses and telling us a little of his own thinking. Lots of fascinating stuff there, especially as John picks apart what seems to be underlying the conversation.
ChatGPT adopts a chatty, deferential, friendly, explanatory style, quick to apologise and explain its own errors. John could of course ask it to adopt a different perspective and style–it immediately did so, and quite convincingly.
John’s Insights
John’s first overall reaction was astonishment that ChatGPT could do so well, and so blindingly fast. Then follows pure gold as John extends his Q&A to investigate various thoughts about how the IT works, what assumptions it makes, how it formulates opinions, and where its boundaries lie. To what extent does it build a model of John? To what extent is it merely integrating the results from a huge number of searches, or is it reasoning about these? John eventually concludes that he cannot trust ChatGPT and cannot follow its reasoning–because there isn’t any. I won’t try to summarise: you need to read John’s final paragraphs to get the rich story.
Extremely well-informed gold!
Geoff