A year ago, most people would have associated Artificial Intelligence with turn-of-the-century sci-fi movies. AI was the fodder for much entertainment in the late 20th and early 21st centuries, but seemed like the prospect of a far-off future. In the space of a few months, AI has gone from a seemingly hypothetical scenario entertained by literature and cinema, obscure bloggers, and the fringes of academia to a very real presence in our world.
In AI research, the ‘alignment problem’ refers to the notion that the actions of artificially intelligent machines are not inherently aligned to the objectives of their designers, and of human beings more broadly. Some researchers believe this problem will not be solved, and hence the extinction of humanity is imminent. Others believe this is a low probability, and that AI can be used to advance human objectives. Irrespective of one’s level of apocalypticism, it is clear that AI is not going away. All institutions and workforces will be forced to navigate its increasing presence in our world, and its rapidly accelerating capabilities.
Young people are always the greatest proponents of new technologies. The most accessible tool powered by AI is the now-infamous language model ChatGPT, created by OpenAI. ChatGPT is, as the name suggests, a chatbot that swiftly provides answers to questions you ask it, and simulates a human conversation. With varying degrees of accuracy and efficacy, it can help students and teachers alike compose emails, write code, complete discussions on scientific reports, and conduct historical research. As such, students from primary schools to universities are starting to use language models like ChatGPT to help complete their schoolwork.
Naturally, schools are facing an array of critical questions about how to reconcile the imperatives of education with the inescapability of AI. To what extent is the imperative to teach students critical thinking jeopardised by AI language models? Should schools attempt to limit the usage of these language models, or should they embrace these as an educational tool? Are there new and innovative forms of assessment that schools can use to harness these new technologies, or is a return to traditional modes of assessment inevitable?
First, it’s important that we acknowledge a few home truths about the status quo when it comes to AI language models and education:
As long as academic performance is a source of pressure for students, students will search for ways to manage their workload through time and cost-efficient strategies. ChatGPT is like having an army of dim-witted, error-prone tutors and researchers at your disposal. These models have minimal utility if a student inputs basic, sparse or uninformed prompts. Setting aside the tendency to ‘hallucinate’, AI language models have no inherent knowledge of what a teacher or marker expects of the style, structure or substance of a response, so simply prompting one to “write an essay on the exploration of justice in To Kill a Mockingbird” will yield a mediocre output. But despite issues in the quality of output, for students who struggle academically or display little academic interest, AI presents an opportunity to complete study notes or take-home assignments with seemingly minimal effort. ChatGPT can seem especially useful for content-heavy subjects like history and commerce that may not require the same personal touch of English or drama. And, for now, many of these language models are free. (OpenAI has introduced a subscription service, ChatGPT Plus, that allows some access to GPT-4, their latest and most sophisticated chatbot, which raises a whole other issue about accessibility.)
Schools and state governments across Australia have already started responding to the advent of AI as a virtual assistant for students completing take-home assignments. There is little consistency between schools’ policies, and state education departments have flip-flopped on the issue. The Victorian Department of Education placed an interim ban on ChatGPT, and blocked access to ChatGPT on servers and devices in Victorian state schools, while Catholic schools in Melbourne have encouraged the use of AI as a complement to traditional education. The Queensland Department of Education similarly banned ChatGPT, but recently changed its stance and expressed a willingness to “harness[…] AI tools to support teaching and learning.” Like banning mobile phones in classrooms, simply ‘criminalising’ the use of AI seems naïve at best and Luddism at worst; students are generally more tech-savvy than teachers, and will undoubtedly find ways to circumvent such hard-line policies.
However, simply ignoring the issue and failing to design any policy is equally naïve, as there are myriad ethical and educational concerns that arise from the ascent of AI.
Firstly, there is a danger that take-home assignments will become redundant. Early in high school, students are often required to practise composing body paragraphs at home in preparation for more advanced academic writing in later years; middle school science programs rely on Student Research Projects, complete with written discussions and conclusions, as forms of assessment; HSC subjects like Extension History and Extension 2 English involve major research projects. Such formative and summative tasks may no longer function as reliable indicators of a student’s learning. As such, a serious reappraisal of the reliance on such tasks may be imminent.
Secondly, there are inevitable questions about academic honesty that arise from overreliance on AI. Of course, a student’s understanding of any subject is a composite of their own ideas and the guidance of teachers, tutors, parents, peers, text books and the internet. However, the use of AI language models to complete an entire assignment seems like an unambiguous act of plagiarism, as another form of ‘intelligence,’ albeit artificial, has generated work and the student has submitted it as their own. But is it? To whom can ownership of that work really be attributed? Much of the output of AI language models is scraped from existing content; GPT-3, for example, was fed 300 billion words from myriad books, articles and other on- and offline sources. While an essay produced by ChatGPT can evade Turnitin, as the combination of words it spits out is not identical to previously published work, academic honesty is not simply breached when a student steals the exact words of another piece of writing. It can be breached through the failure to acknowledge the origin of an idea, replication of the structure of another author’s argument, or even mimicry of another author’s syntax without acknowledgment.
Thirdly, AI may threaten the development of critical thinking and literacy skills. Students who have relied on AI tools to complete their assessments may not have exercised any intellectual effort themselves, especially if they prompt the language model to generate whole arguments. And critical thinking skills – the kind notionally accrued through the study of humanities at school – will be pivotal in a world in which AI and deep fake technology will increasingly blur the boundary between the real and the false, truth and deception.
Finally, as mentioned above, the output of AI language models is fallible. It can write a good email and solve a logic puzzle, but can it write a B-range or higher essay on Macbeth, complete with analysis of textual form, integration of quotations (and not just ones it makes up) and discussion of social context in a personal-meets-academic register, without laborious instructions from a knowledgeable human on the other end, who could have just written such an essay themselves? Students who don’t know how to use the language models may submit work that is inferior to work they could have produced themselves, and work that doesn’t align with a marker’s expectations regarding language, tone, structure and purpose.
While AI has provoked much consternation about the future of education – and, indeed, the planet – there are benefits that may arise from its use within educational contexts.
AI language models may maximise the efficiency of completing an assessment. They have the potential to reduce the more tedious aspects of research and thus open up greater opportunities for human critical thinking about the researched material; however, GPT-3 is only trained in data up to 2021, so a lot of statistical information is outdated, and ChatGPT cannot accurately cite its sources (many of which it entirely fabricates). Despite this, ‘collaborating’ with AI has the potential to produce very good work, as the machine may offer ideas or perspectives that a student my not have considered. Indeed, ChatGPT may help students edit their writing for concision or rectify grammatical and syntactical mistakes.
Additionally, as AI improves (and hopefully aligns with human objectives), jobs of the future will almost certainly use AI technologies. There has been a recent spate of layoffs in tech-related workplaces, and this will plausibly accelerate as AI advances its capabilities with coding and software engineering. Various jobs in law, media, education and finance have likewise been predicted to be affected heavily by future advancements in AI. People need to learn how to work productively with AI, and employees who can view AI as a collaborator, rather than a competitor, may have a competitive advantage in the job market of the future. It is surely unwise to insulate students from this reality by introducing blanket bans on AI language models in schools, is it not?
OpenAI has “trained a classifier to distinguish between text written by a human and text written by AIs,” but such detection models are, thus far, incredibly weak and unreliable. Turnitin has launched AI detection, but their model prioritises precision to reduce false positives, and therefore may miss huge swathes of AI-generated content. Hence, schools cannot simply keep the status quo and hope that detection models will be able to weed out potential academic dishonesty. The inevitable question, then, that arises is: Should schools return to traditional in-person exams for all assessments, or find new and innovative modes of assessment?
Here are a handful of suggestions for schools as they navigate the challenges posed by AI, and adapt their modes of assessment accordingly:
Schools will need to develop policies that are appropriate to their student cohort, mitigate the use of ChatGPT in assessments, and acknowledge its inescapability. The functionality and capability of AI will likely never be worse than it is today. Be prepared for the inevitable.
To help us better understand how AI is impacting schools and what resources may assist, please take a couple of moments to complete this survey.