“Artificial Intelligence” and Sydney Grammar School¹

A developing situation

In November 2022, a new generative AI tool called ChatGPT (Chat Generative Pre-trained Transformer) was released by OpenAI. This was a new chat-style interface to OpenAI’s flagship Large Language Model (LLM) series (GPT 3.5 and now GPT 4.0) and people immediately found it compelling, easy to use, and strikingly lifelike in the quality of interaction possible.

While LLMs are not in themselves new technology, ChatGPT (and the new versions of OpenAI’s LLMs that it was built upon) had tipped over into a whole new level of convincingly grammatical and well-informed output. This struck people as a genuine taste of the sort of artificial general intelligence long promised in the pages of science fiction. There was an immediate media flurry, and a broad range of overhyped responses, both positive and negative, were offered. Variously forecasted as a direct consequence of this new era of text-focused generative AI tools were the undermining of traditional education, the total transformation of workplaces, and the imminent onset of an “AI apocalpyse”.

Quite swiftly the New South Wales Department of Education banned the use of ChatGPT in schools (as did schools in other jurisdictions, including NYC). The Australian Group of Eight universities stated that as an immediate response they were increasing in-person supervision and examination style written assessments.

By contrast, other voices in educational and scholarly communities commented that ChatGPT could be used as a tool for learning, and that the advent of these newly sophisticated LLMs should not just be considered a threat to assessment in the context of schools and universities.

Since the initial flurry of responses, universities in particular have made more nuanced and varied statements on how they view the technology and its potential impact, both positive and negative, on academic work. At Grammar, we have also been reflecting on these developments.

This article sets out some groundwork for the ongoing conversation that will continue to be had within the School community about these new tools.

What is ChatGPT and how does it work?

In essence, ChatGPT and similar text-focused generative AI tools are programs that take an initial input and then serially predict the next possible word based on the initial input and the program’s own subsequent additions to it. This is a bit like taking the idea of “predictive text” that we are familiar with from our phones and ramping it up to a whole new level based on a gigantic data set scraped from all over the internet.

To create these LLMs, the statistical relationships of this truly vast amount of training data are mathematically modelled by building vectors and neural network structures in an incredibly computationally expensive operation to produce the model into which inputs can be fed to generate output. A short and clear explanation created by the Education Innovation Media Team at the University of Sydney can be found here.

Given how these models are constructed, at least four key issues have been identified with the potential status of these tools as “artificial intelligence”. Firstly, at least at this stage these models are quite static in nature, programmed on the data that they had access to during their training period: the world changes, then existing models won’t know about this until their next version is produced, and this process takes considerable time and money. Secondly, these models by their construction reproduce any existing bias already existing in the materials they have been trained on, and – given that this data includes a significant proportion of what is on the internet – caution is certainly advised. Thirdly, these outputs of these models are prone to “hallucinations”, statements that are plausible at a surface level but factually incorrect. Fourthly, and directly related to the previous point, these models have no actual ‘intelligence’ at all: despite the fluency of their outputs, there is no grounded understanding and lived experience and LLMs are not “thinking” in any humanly recognisable sense. In the final analysis, LLMs have no actual conception of what is real and what is not. The current generation of models may even represent something of an evolutionary dead end, in the sense that their well-documented problems may not be able to be worked around without complete reinvention.

_{^{1 Many Grammar staff members have made valuable contributions to the discussions leading to the publication of this article. Particular thanks are also due to Mr Jamie Twiss, a current Grammar parent. Any errors that remain are those of the authors alone.}}

How much are pupils using LLMs?

At Grammar, we have had ten specific instances across Term II and Term III this year (that is, up to nine or ten months after the first big splash about ChatGPT hit the headlines) where individuals have been penalised for academic misconduct for using ChatGPT or similar LLMs to complete homework tasks.

This may, of course, not actually reflect the true level. A recent article in The Sydney Morning Herald (Harris 2023) noted the common understanding that students are using ChatGPT more than schools or universities are picking up on. While the School does use the services of Turnitin and while they and similar companies are continually developing their software tools to identify AI generated text and paraphrase, it remains the case that not all homework at the School is run through Turnitin.

A misconception?

Nonetheless, placing too much of a focus on software tools and the arms race between generators and detectors may be to miss a fundamental point. As has been recently observed in an acute essay about AI-generated disinformation on social media (Kapoor and Narayanan 2023), generative AI models are not providing new abilities so much as they are offering a significant cost reduction in access to existing abilities. Put simply, students in schools and universities have always had the capability to engage in academic misconduct in various ways: ChatGPT and its ilk are not transformative because they can be used for academic misconduct. Instead, such tools are transformative to the extent that they make certain forms of academic misconduct much easier to access.

Here we come back to a fundamental principle of the School’s approach to education: direct interaction between teachers and pupils in the context of a joint exploration of a specific academic discipline. At Grammar, we pride ourselves as teachers and mentors for our pupils not least because we have a good understanding of the type of work that the boys in our classes are capable of. Masters become quickly aware of homework that seems out of keeping with a boy’s usual patterns of content and language, and to this extent the advent of the use of generative AI tools does not look that different from various other practices involved in academic misconduct.

Our biggest fear in this respect is that the ease of access to such tools creates an increased possibility for boys to seek shortcuts rather than spending the time thinking things through, particularly when they are short of time or otherwise stressed in some way. Such tools have the potential to reinforce a poor work ethic and to lead to failure when boys are asked – as they regularly are – to show us what they know and what they can do in situations where such tools are not readily to hand.

In the end, what we know and what we can do defines us. While this knowledge and capability might include knowing how to use LLMs and similar generative AI tools to enhance our work and productivity in ways that are consistent with academic integrity, none of us should see this as any kind of substitute for the embodied knowledge and skills that we gain in our journey through school and university and in our working lives.

What are the possibilities for teachers to use LLMs?

The Sydney Morning Herald article cited above (Harris 2023) noted that many private schools – where ChatGPT has not been “banned” in the same way as in NSW Department of Education schools – are exploring how to integrate ChatGPT and similar tools into lessons, and how ChatGPT might be used to provide feedback to students. Similarly, articles and social media posts point to teachers using ChatGPT for administrative work such as writing newsletters as well as for writing reports and feedback to students, and going as far as determining grading.

We feel very confident in saying that this is not a path that Grammar as a school would wish to consider. As one educator posted on social media, “the debate is far more nuanced than looking for cheap ways to claw back a little time for overworked educators” (Furze 2023), and the ethical use of generative AI tools is a topic that should be of concern to all.

This is not to say that teachers should not know about such things. It seems quite reasonable to suppose that generative AI tools will form some part of current and future workplaces. Consider the present incorporation of LLMs into common productivity tools such as Notion and Grammarly, and Microsoft’s current push to build generative AI tools into its modern software worksuites from Microsoft365 through to GitHub Copilot. Among other things, tools like ChatGPT are arguably only as good as the person asking the questions and there is clearly a developing art in knowing how to set the wheels in motion effectively via “prompt engineering”.

Nonetheless, generative AI tools cannot engage in “active, persistent and careful consideration of any belief or supposed form of knowledge in the light of grounds that support it” (Dewey 1933, p. 6). This is how John Dewey – an influential early twentieth century philosopher of education – defined “reflective thinking”, the development of which is a key aim of schooling and university. Similarly, a 2023 report from the US Department of Education’s Office of Education Technology notes (2023, p. 12) that while we may speak of “human-like” reasoning as a shorthand when discussing generative AI tools, this can be misleading:

The “human-like” shorthand is not always useful, however, because AI processes information differently from how people process information. When we gloss over the differences between people and computers, we may frame policies for AI in education that miss the mark.

Education remains a fundamentally human pursuit that we needn’t fear is being derailed in any simple way by these latest tools.

Teaching with ChatGPT?

In the same way that generative AI tools may have a place in the world of work, there are opportunities in schools for teachers and pupils to engage with these tools in a meaningful way. Just as Wikipedia has enhanced some aspects of education by opening up access to a certain type of consensus-based knowledge, deciding what questions to ask based on an initial overview of a topic could be a richer process with ChatGPT as one of the tools to be used. Similarly, ChatGPT could be employed to help pupils think critically about a topic they are already familiar with in ways that can then inform their own writing.

The image shown opposite is an example of an approach that a boy or master might use to engage in revision of a particular topic. While this particular example is modest, such ideas have potential, particularly in the junior years.

In the end, though, teachers will continue to give careful consideration to what exactly they are trying to teach their pupils and what developing abilities they are trying to measure at any one time, as is usual in any worthwhile lesson planning in schools. It is precisely for these human capabilities that teachers are not going to be displaced by “robot tutors in the sky” any time soon._²

_{^{2 Offering a “robot tutor in the sky” was a central pitch by the company Knewton in 2015: see e.g. Neurons.AI 2015.}}

A teacher considers how ChatGPT might be used for topic revision (Esterman 2023)

What questions should we be asking ourselves?

Two researchers from the Faculty of Education at Cambridge University have recently offered an interesting set of questions and responses for educators to consider (Kirk 2023). The questions include:

• Why does ChatGPT matter for educators?
• Should ChatGPT be banned or embraced?
• How might ChatGPT change the way people teach and learn in universities and/ or schools?
• What are the main challenges that ChatGPT raises for the education sector?
• How should educators handle the potential for ChatGPT to reproduce social biases and inequities?
• At a local level, how can universities and schools respond to ChatGPT’s emergence in a meaningful way?
• Is education set up to cope with AI?

With a similar spirit, a group of American, Australian and UK academics put forward for consideration four concrete applications of AI in teaching as examples to reflect on. Students could be asked to:

1. Reflect on articles produced by ChatGPT which have fabricated references and distorted information and then deliberate on the potential consequences of this in an era of fake news.
2. Compare ChatGPT’s answers to ones they have developed and ascertain whether they know the material and how it might be represented differently.
3. Learn legitimate techniques for using ChatGPT to support their essay writing and to help foster a sense of mastery and autonomy, and then analyse ChatGPT responses to note how the software has drawn from multiple sources and to identify flaws in the ChatGPT responses which would need their attention.
4. Consider the extent to which their use of ChatGPT has enabled or constrained their access to powerful knowledge, as a critical reflection on where and how the use of AI is taking place in society and their potential future professions.

Such applications of generative AI tools in education have the potential to reduce the mystery – and indeed the acclaim – of ChatGPT while taking advantage of what these tools offer, in a situation where skilled teachers and university lecturers work with their pupils and students to assist them in understanding how to use these tools well. In the arc of human history we arguably always “stand on the shoulders of giants”: in using tools wisely we affirm our own human capability while also taking advantage of the latest technological wonder produced out of the deep well-springs of human creativity.

What about assessment?

The majority of assessment at the School is conducted under direct supervision and written with pen and paper, as we seek to uncover what the boys know and can do on their own account. This output ranges from drawings of still-life to complex mathematical reasoning through to literary essays and creative writing. We can continue to be confident about the integrity of our assessment programmes, despite the advent of generative AI tools.

Subjects that have hand-in tasks for assessments are looking closely at the place of generative AI tools in relation to those tasks. We are not alone: schools and teachers everywhere are engaged in this process. Consider for example the traffic-light system for assessments proposed by Adrian Cotterell (2023) where different types of task allow for different levels of generative AI tool use while still preserving the validity of the “assessment construct” – the specific feature the assessment is aiming to evaluate.

As noted above, the detection of use of generative AI tools is a developing area. You can follow the progress of companies like Turnitin via their websites (Turnitin n.d.). But as with all previous tools to assist the detection of academic malpractice, the role of teacher judgement remains central. Each master at the School will continue to need to reflect on how the boys write and the level of work they usually produce, in the same way that we continue to conduct assessments in ways that assist us to reduce the opportunity for other forms of potential misconduct.

Concluding thoughts

As with most things in education, it is important to take time to reflect and not to respond in a knee-jerk manner. Grammar prides itself on thinking things through and offering considered responses to topic issues in education. That being said, the place of generative AI tools in the world of work and education are developing fast and it seems an appropriate time for the School to share some thoughts on this topic with the Grammar community.