There are diverse intelligences and dispositions that we need to cultivate in citizens and students, to equip them for the challenges now confronting society [1][2][3]. One that persists in all lists of '21 st century skills' is critical thinking/analysis. One general form that this takes in formal education is the capacity to understand, critique and formulate arguments, which transfers into knowledge work in the workplace. In this paper, I describe how GenAI apps offer new capability in this regard, reporting on tests I have conducted as a continuation of several decades' research into argument visualisation [4][5][6]. I then describe a second form of critical analysis, namely, distilling a body of ideas into a more succinct summary, the example being the clear articulation of university course learning outcomes. These examples help to demonstrate what AI can now do, which until recently we considered the preserve of humans. This invites a discussion of how we balance the benefits and risks of human cognitive offloading to AI, as a perspective on the broader question of human agency in future human/AI systems.
Arguments are being made constantly in everyday public discourse, as well as within academia. We aspire for citizens to be able to make robust arguments, as well as critique them appropriately. In the philosophy of argumentation, the recurring types of argument have been taxonomized into a robust set of "Argumentation Schemes" [7]. OpenAI's GPT-4 is a sophisticated aid to analysing everyday arguments, as illustrated in Figure 1.
Social media platforms such as Twitter have established themselves as influential channels for public discourse and opinion, although the quality of conversation is of course highly variable with platform and community. In a tweet, a well-known AI researcher argued that "AI doomers", who are proponents of strong AI regulation, would also have called for the banning of pens and pencils. This is an argument by analogy [7]. Bing Chat (a version of GPT-4 integrated into the Microsoft Bing search engine) was able to critique this claimed analogy effectively (Figure 2).
The examples so far have been very short: the arguments have made a single 'move', which GPT could recognise and comment on. Let us now consider a more complex case. In March 2023, a large number of eminent thinkers wrote an open letter calling for a pause in building large language models. 3Achieving widespread media coverage, this provoked extensive debate, including a letter of rebuttal from another set of academics and industry researchers. 4 This seemed an authentically rich argument to test GenAI.
I asked Bing Chat (now Copilot) to access the letter online and identify the key claim and arguments. It provided a reasonable textual summary, output as a set of bullet points summarising key arguments. However, it is well established that students struggle to critique arguments, and that rendering them visually as an argument map can help them understand the key elements of the argument (this is a form of concept map tuned specifically to show multiple perspectives, and the key features of arguments such as supporting/challenging claims/evidence). I asked it to generate a map, but it could not. However, when asked, it confirmed that it understood Argdown, which is a markdown notation for argument maps. It generated this in a code window, which I pasted into the Argdown web app, 5 resulting in a map (Figure 3). Examination of the argument map reveals to what extent this was a rigorous analysis, but also illustrates 'hallucination in argument mapping' (Figure 4). Hallucinations of two types were found. Firstly, the red underline signals incorrect classification of a premise using incorrect, or indeed madeup argument schemes. There is to my knowledge no such argument type as Argument from responsibility, or Argument from precaution. Argument from omission seems to be a jumbling of Fallacy of omission and Argument from ignorance.
Secondly, there were hallucinated summaries. This node apparently reads well as a summary, but the authors do not talk about researchers at all.
Asking students to perform critical evaluations of AIgenerated argument maps should serve as assurance of learning about the subject matter, but can also provide important insights for them into the limitations of AI, if students are equipped and empowered to see through hallucinations.
Conversational agents are exciting for education since they are, by definition, premised on learning through dialoguehardly a novel concept. But consider this illustration of GPT's capabilities (Figure 5). The capacity to add relevant information that was neither requested nor in the source article, and explain this when queried, is unprecedented. In the next example, Bing Chat is asked if can add new nodes to the Argument Map (Figure 6). Bing Chat's Argdown code can also be rendered as textual outlines. Figure 7 shows the addition of the critical questions, and substitution with placeholders for students to complete. It should be noted that the above capabilities are all from the generic ChatGPT-4 model, but as discussed later, customizable intranet GPTs open many new possibilities for tuning chatbots educationally, to institutionally-specific requirements (e.g., within a particular degree program).
An important educational question arises, as we see this kind of performance, namely, will the students engage in excessive cognitive offloading, and fail to learn how to do this themselves? We return to this in the discussion about user agency.
Thus far, we have focused on critical thinking and reflection around arguments, primarily with students in mind, but equally, these are tools for any professional to test their thinking. In the next example, we focus on a specifically instructional task, which harnesses the generative capability of LLMs more fully to distill complex text into key themes. The text in this case is a specific 'genre' of writing, the Course Intended Learning Outcome (CILO). CILOs define what students know and can do on successful completion of the course. As part of a well-designed curriculum, each part of a coursesubjects, modules and assessmentsshould all respond to its CILOs. Effective implementation of CILOs requires both the subject matter expertise of academics and the pedagogical knowledge of learning designers (LDs). Indeed, recent evidence points to the benefits that academics gain from working with LDs on their online teaching, and how this transfers to their in-person teaching [8].
One specific element in this task that academics can struggle with is to articulate good LOs. Furthermore, these typically vary widely in quality and quantity between academics. At UTS, we are working towards summarising all courses consistently using approximately six CILOs, to achieve a better user experience as students make enrolment decisions, and to assist teaching teams in their course design and reviews. However, it is an intellectually and linguistically demanding task to distill a list of 20-30 CILOs (which is not uncommon), down to six well designed CILOs, and the university needs to implement this summarization for its entire program.
It is here that we anticipated that LLMs could assist. GenAI intranets now provide universities with authenticated, secure, private services, integrated with other internal services, and tuned to support business processes. 6 In a 2-day hackathon, iterative prompt engineering informed by feedback from academics and learning designers led to the refinement of a system prompt that configured 'CILObot', a ChatGPT to aid in drafting these new CILOs. The system prompt incorporates widely recognised design principles (e.g. open each CILO with a verb from Bloom's Taxonomy), with the addition of internal requirements (e.g., UTS Indigenous-CILOs), and the chatbot is grounded in a corpus of documents about CILO design.
The prototype is showing promise, and after a day's intensive work using the Azure 'Chat Playground' (the ChatGPT design environment), the results for several programs in our Health faculty were validated by disciplinary experts (e.g., Figure 8). CILObot generates a coherent first draft in about 30 seconds, which can of course then be refined through further conversation with it, and edited by the teaching team. We estimate that agreeing on how to distill 20-30 CILOs into 6 would normally be a minimum of 3 hours' meeting between the Course Director and the program's lead academics, which represents an impressive return on investment. Next steps will test CILObot with other degree courses.
These capabilities are, in my view, impressive. If students were to produce argumentative reasoning as presented above, we would surely conclude that they were thinking critically, and had mastered some argumentation principles. Similarly, if an academic proposed a set of six distinctive, well expressed CILOs with complete coverage of the original CILOs, we would regard that as exactly the kind of task senior academics should be capable of. The difference, of course, is that these tasks are performed in under a minute, producing coherent drafts.
We do not need to believe that agents have the same kind of understanding as people to appreciate the value of AI being able to communicate with this fluency precision in order to provoke critical human reflection. GenAI performs these tasks in seconds, and can iterate its analysis as often as requested. In principle 7 , therefore, GenAI can be used to:
• offer students, academics or any other kind of analysts instant, formative feedback on draft arguments, for instance, identifying points that could be potentially attacked; • analyse a written corpus to give insights into the quantity and quality of argumentation, which could inform both LA researchers, practitioners, and educators; • analyse a written corpus in order to derive a representative set of summary themes (noting that AI cannot 'read between the lines' as a human qualitative analyst does).
The pivotal questionwhether we are envisioning the future of learning among students or professionals in the workplace -is the "allocation of function" between human and machine, to use the original term from ergonomics. Questions of cognitive offloading and human agency now arise, as we consider different scenarios.
If AI improves short-term productivity (e.g., faster syntheses of complex information; more creative ideas; more incisive reasoning), we might anticipate (and indeed we are already seeing in certain professions) that AI apps will embed into professional work practices. Professionals are qualified to 'drive' such intellectual power-tools (in contrast to students their qualifications should enable them to recognise poor AI output); they will welcome cognitive offloading in their busy lives; and if they do not use AI may find they are unable to compete with those who do. We might see this as empowering professionalsand yet we might also see a loss of agency as they are essentially forced to use AI in order to compete. Time will tell if the long-term use of AI leads to the degrading of important human capabilities, just as GPS satellite route navigation has for many young people obviated the need, and hence ability, to navigate via printed maps.
In sharp contrast, for education the story is very different. "Productivity gains" need to be judged by a different yardstick, since while an essay written solely by GenAI in 2 minutes is a "productivity gain" in terms of artifacts/minute, the absence of the student's cognitive engagement fails other "KPIs" for meaningful education. Students must build their foundational knowledge, skills and dispositions, in order to function as citizens and professionals in the myriad contexts in which they cannot call on AI, but must think on their feet and demonstrate diverse intelligences [1,2].
Consequently, as emphasised in a recent national report for the higher education sector, assessment must be reformed for the age of AI [9]. Cognitive offloading takes on special importance in assessment design [10], since it forces us to ask what exactly we deem important to assess in the age of AI. It is beyond the scope of this short paper to expand on this issue further, but a fitting conclusion is to return to AIED research 30 years ago, and remember a distinction made by Roy Pea (emphasis added):
"Pedagogic systems focus on cognitive self-sufficiency, much like existing educational programs, in contrast to pragmatic systems, which allow for precocious intellectual performances of which the child may be incapable without the system's support. We thus need to distinguish between systems in which the child uses tools provided by the computer system to solve problems that he or she cannot solve alone and systems in which the system establishes that the child understands the problem-solving processes thereby achieved. We can call the first kind of system pragmatic and the second pedagogic. Pragmatic systems may have the peripheral consequence of pedagogical effects, that is, they may contribute to understanding but not necessarily. The aim of pedagogic systems is to facilitate, through interaction, the development of the human intelligent system. While there is a grey area in between and some systems may serve both functions, clear cases of each can be defined." [11] GenAI forces us to ask when we areor should be, as the boundary shiftsassessing joint human+AI system performance, versus capability without AI. A consequence of this distinction is that we must cultivate "mindful engagement", not "mindless engagement" [12]. In the intense debates about whether the human (student, academic or professional) remains sufficiently in the loop, these concepts from the era of symbolic AI, when they could barely glimpse what is now possible, remain as important as ever.
CILObot would not have been possible without the joint expertise of my colleagues Sharon Coutts, Ann Wilson, Michaela Zappia (Institute for Interactive Media and Learning), Susan Gibson & Carl Young (Data Analytics and Insights Unit), and Miguel Ramal & Olaf Reger (IT Unit).