These notes are less about what was presented and more about my own thoughts that were sparked as I listened to the session. The panel members included:
- Rakesh Patel (Chair/Moderator): Introduction
- Daniel Salcedo: Introduction to AI terminology and concepts.
- Raquel Correia: Where and How in the curriculum can AI be placed?
- Martin Pusic: How can AI be used in HPE?
- Ken Masters: Ethical Issues in the use of AI in HPE.
We need to make sure that we’re not holding up human intelligence as some kind of peak in the intelligence landscape. AI will exceed the computational ability of humans, in breadth and depth.
Machine learning algorithms aim to improve their own performance by learning new things. But what they learn isn’t always specified in advance (depending on what type of ML algorithm is being used). It’s important to recognise that, sometimes, we don’t know what we want ML algorithms to learn. This makes it very hard to know if what they’ve learned is correct. The best example I can think of is the ML model that scored well on the problem of recognising malignancy of tumours. But what it had really learned was to recognise rulers, because the images of malignant tumours in the training dataset included a ruler for scale comparison. We thought we’d built a ‘malignant tumour detector’ but what we actually built was a ‘ruler detector’.
Algorithms won’t work in isolation. When an ML system / agent learns something, it will share what it’s learned with every other agent in the system. We’ll have ecosystems (sometimes called swarms) of autonomous agents that pass information between each other, updating the collective intelligence of the system immediately. Imagine a self-driving car that learns something new. That lesson will feed back into the central model, and update every other self-driving car. Think about how long it takes for cutting edge advances in practice to percolate to all the places it needs to be. Software updates of autonomous to AI-based systems will enable immediate enhancements to collective intelligence.
There was a comment about how chess-playing algorithms can’t drive your car. But what if chess-playing (or single-purpose) algorithms are transitional, and that we’re moving towards general-purpose algorithms? If decision-making / problem-solving is a general ability (and there’s no reason to think that our cognitive systems for solving problems differs across knowledge or disciplinary domains), we may be able to develop a ‘problem-solving algorithm’ that can solve chess problems, Go problems, driving problems, and diagnosis problems. Maybe. I’m just speculating.
I was a bit disappointed that there was a distraction in the form of discussion around sentience, consciousness, and Artificial Generally intelligence. Personally, I don’t believe there’s any reason to think that consciousness isn’t computational (by which I mean, information processing). Because of this, I also believe there’s no reason to think we won’t eventually build AGI systems that are competent across all knowledge domains (and possibly, that will also be conscious). However, I don’t think we need to include that discussion in what was intended to be an introduction to some of the ideas relevant for health professions educators today. We have more immediate concerns around policy and practice in higher education, as a result of generative AI, than what might happen when / if AI systems become self-aware.
There was a nod towards the many challenges we face with AI-based systems in general, including:
- Bias
- Compute
- Legal frameworks
- Ethics concerns
- …and so on.
One of the presenters made the point that we don’t have experts in AI, which is obviously incorrect. We may not have experts in generative AI, although at this point I’d say that we definitely have a significant number of people building diverse skillsets and expertise in how to use it effectively. But in terms of AI experts, of course we have those.
Ken included references to his overview on AI in medical education, as well as his AMEE Guide on the ethical use of AI in HPE, which are both worthwhile reads.
“If we get the ethics of AI wrong, we get everything wrong”. I agree with the principle but I’m not sure what value it adds, in the sense that it applies to everything else as well. If we get the ethics of assessment wrong, we get everything wrong. If we get the ethics of admissions and selection wrong, we get everything wrong. I feel like this focuses attention on AI in a way that we don’t necessarily pay to assessment and admissions. Ethics cuts across everything we do. Maybe I’m just being pedantic?
Generative AI is not optional; it’s going to be built into society. Like AI is currently built into search, maps, and so on, language models will be built into many (if not, all) of the software we use daily. There will be no getting away from this. And, given that, we need to learn how to deal with it. For example, we have to rethink what ‘cheating’ means in learning environments that now include ‘intelligence on demand’.
It’s a mistake to focus on the weaknesses of generative AI (although we absolutely need to be aware of what they are). However, these weaknesses are going to get smaller every day. If we keep hiding in the spaces that AI can’t go (yet), we’re going to find ourselves with increasingly smaller spaces to hide. Having said this, it’s essential to recognise that the models we’re using to build generative AI may have fundamental flaws that mean we can never fully trust their outputs.
There was a suggestion from one speaker that we teach our students to get better at writing prompts. But, what happens when AI models get better at writing prompts?
Universities need the equivalent of institutional review boards (IRBs) to provide oversight on how students and staff use AI for learning, teaching, and assessment. Maybe. I’m not sure how they’re going to regulate ‘authorised’ use when this is going to be built into everything.
Q&A
- How we identify responses from ChatGPT that are confident but wrong? My – perhaps facetious – response would have been to ask how we can identify the specialist surgeon who is confident but wrong?
- How can we keep up with the pace of information that’s coming out with respect to generative AI? My response: Follow people you trust, who are curating and sense-making; find high-quality sources and pay attention to what they talk about e.g. domain-specific podcasts and newsletters; insert yourself into a community of practice; read ‘philosophy of AI’ papers.
- How are we preparing for cohorts of incoming students who are more sophisticated users of AI? We need to engage in a process of faculty development around the use of AI in almost all aspects of our work. We should also not assume that students know how to use AI to support learning.
- How are we going to address the issue of students using AI-support to prepare their applications? Again, a somewhat facetious response: How are we going to address the issue of wealthy students having access to more writing support and networks of privilege, that help them prepare their applications? My point is simply that we already have these kinds of problems, so it seems odd to focus on AI. This may be because I tend to think of AI as a super-helpful friend, and we already have those.
- Someone noted that AI companies are for-profit and that this somehow makes them more of a problem. Microsoft, Google, Apple, and Blackboard all exist to make money and we don’t seem to have any problem using their products in higher education. Weird point to make.
- What is unique about human relationships and how do we ensure that they remain front-and-centre in HPE? What if there’s nothing unique or special about human relationships? What if the empathy we feel for others is a sub-routine / algorithm that runs in our subconscious? What if the emotions we feel towards others are entirely predictable and therefore able to be manipulated (hint, they are). What if everything we do and feel is the result of computation, and therefore able to be replicated in silicone?
The main thing missing from the session (IMO), was a meaningful focus on why this matters for the audience. I didn’t think there was much in either the presentations or discussion that would help an audience member do something different when they’re back at work. I’d love to know how this symposium was received by a ‘normal’ person i.e. someone who doesn’t spend a lot of their working week focused on these questions.