How do we navigate Generative AI?

The trickster and Titan Prometheus stole fire from the gods of Olympus and gifted it, together with the skills to use it, to humans. In this tale, fire is a metaphor for technology, art, and knowledge – all the things that enabled the development of human civilization.

Angered by yet another transgression, Zeus had Prometheus captured and chained to a rock in a desolate mountainous area and, after some years, tormented daily by sending an eagle that ate his liver which would regrow overnight, only to be eaten again the next day. For good measure, Zeus also punished humans by sending them Pandora, the first woman, along with a box – Pandora’s box – filled with all the troubles and woes that ail the world today and instructions to never open it, which she famously disobeyed.

We are soon coming upon the one-year mark when, like Prometheus, OpenAI released ChatGPT 3.5 and put the capabilities of a sophisticated Large Language Model (LLM) – which heretofore were the almost exclusive domain of big-tech companies, AI researchers and hobbyists – into the hands of everyone with an Internet connection and a computer or smartphone. LLMs take a string of text as input and produce an appropriate series of words as output. They are one of a broader category of AI models called Generative AI (GenAI) models that can take text, audio, images, or videos as input and produce output in one of those forms. LLMs are GenAI models that take text (a prompt message) as input and produce text as output.

While the world can enjoy Prometheus’ gift and the ease it brings, the challenge of closing the Pandora’s box that came with it has fallen largely on academics. The realization that the price for a world with better-composed emails and blog posts would be paid by the education sector – schools, colleges, and universities – was immediate. Ubiquitous access to LLMs – first OpenAI’s ChatGPT 3.5, followed by its improved version 4.0 (available by subscription and soon to be integrated with DALL-E 3, its text-to-image generator), Google’s Bard AI, Microsoft’s new Bing and Microsoft 365 Copilot – has, for those seeking shortcuts, reduced the formerly intimidating task of essay writing that students work on for weeks to a matter of recursively crafting a good input prompt. A job that may take a student hours or days can be reduced to mere minutes.

Educational institutes of all levels are now left with the task of figuring out how to adapt their teaching and assessment of student learning in a post-ChatGPT world. Should students be permitted unfettered use of LLMs like ChatGPT or should it be banned or something in between? These and other questions have been occupying academia for all of this year.

Flare-ups of technophobia in education are not new. The invention of the typewriter brought fears that students would no longer bother learning to write. Pocket calculators were feared because they would leave students incapable of performing arithmetic calculations. Gel and ball pens would ruin students’ handwriting scripts. Scientific calculators would deprive them of the ability to use slide rules and log tables. Computers and the Internet would kill library skills. Spell checkers would make students forget correct spelling and punctuation rules. Like in eras past, the world is not about to put a blanket ban on LLMs while there are so many positive use cases for them because they pose a challenge to educators.

While some countries are trying to develop a national policy on the use of Artificial Intelligence (AI) tools like LLMs in government, education and other sectors, there are others where universities are understood to be autonomous, and the decision of how to deal with their availability is being left to them.

In the UK, universities of the Russell Group – an association of the UK’s 24 leading universities that include Oxford, Cambridge, the University of Birmingham, Edinburgh, Imperial College London, King’s College, the London School of Economics and Political Science, etc – have developed a common set of guidelines for ‘Generative AI and its role within teaching, learning, and assessment’ (GenAI Guidelines, for short). These guidelines are not the result of reflexive action but were developed after much deliberation. These are worth paying attention to because of the reputations and number of the universities that contributed to their development means they are likely to inform and influence policies and guidelines other countries and universities will adopt for their own use.

The Russell Group’s GAI guidelines are based on five principles. First, universities ‘will support students and staff to become AI-literate.’ This means being aware of the fact that information passed to an AI tool as input could possibly be retained and used to further fine-tune its operation. All AI systems are trained on data which means biases and lack of diversity in the training data will reflect in the output produced by GenAI tools. In addition to being biased, the output produced can be factually wrong, what is also called a ‘hallucination’, as many have demonstrated on the latest wave of LLMs.

AI systems are also unaware of any ethics codes that their users operate under. While LLMs usually do not perform something as crude as a copy-paste, the outputs they produce are a very complex, almost unexplainable, function of the training data they were fed in training phases. That means that there remains a risk of plagiarism. In any case, the output produced by a GenAI tool is not considered the work product of the user giving input prompts.

Second, ‘staff should be equipped to support students to use GenAI tools effectively and appropriately in their learning’. While remaining aware of ethical considerations and limitations of what an LLM can and cannot do (listed above), teaching staff should be able to guide students about good discipline-specific use cases for GenAI models that aid and improve learning.

Third, universities ‘will adapt teaching and assessment to incorporate the ethical use of generative AI and support equal access'. The role of GenAI tools in classrooms and assessments is where the rubber hits the road and is the core issue most educators are trying to wrap their heads around. Developing new or adapting existing teaching and assessment methods suitable to the discipline in the age of GenAI is a creative task in which all educators should participate. It has been less than a year since the first LLM achieved widespread use. Many more can be expected to follow in the years to come. That is why this space will require continuing innovation by educators.

Educators have to remain mindful that, like right now, there will be some GenAI tools that will be free, some that educational institutions will adopt officially but also other more capable tools available behind paywalls and that only some students will have access to.

Given that we know LLMs can hallucinate, one might argue that using LLMs to aid learning is problematic. However, consider that less-than-reliable sources of learning can also be useful. Take peer learning for example – students teaching each other – that may be used in organized in-class group activities or outside the classroom in student-led study groups and online class forums. As long as the limitations of LLMs are understood, they can serve as round-the-clock learning support to augment the efforts of professors and teaching assistants. The fact that, since its inception, ChatGPT has been able to pass several standardized tests suggests that this could emerge as a realistic use case in at least some disciplines.

Fourth, universities ‘will ensure academic rigor and integrity is upheld'. The three preceding principles champion an understanding of the capabilities and limitations of GenAI tools, an understanding of what constitutes original work, and an openness to use GenAI as another tool for learning. The fourth principle prohibits the use of GenAI by students in assessments unless explicitly permitted by assessment guidelines.

A major vendor of a tool widely used to identify text copied from Internet sources in student assignments recently announced a new product feature – the ability to detect text generated by LLMs. Testing has shown that it is currently approximately 98 per cent accurate. While this number seems high, it is not a perfect guarantee and has the potential to bias investigators on suspected cases of unsanctioned use of LLMs. That is why many universities have opted to disable this feature in their subscriptions.

Developing sufficiently reliable detection tools for LLM and GenAI-produced outputs will take time. However, whether it takes months, years, or forever, it is important to establish guidelines on academic integrity that govern the use of LLM and GenAI products.

Finally, universities ‘will work collaboratively to share best practices as the technology and its application in education evolves'. Experience sharing for the sake of gradual and continuous improvement is a common practice in education, even when there are no new disruptive technologies to adapt to.

Together, these five principles add up to a clear-eyed approach that neither declares an outright prohibition in education nor an unconditional surrender granting blanket permission to use Prometheus’ gift.

The writer holds a doctorate from Michigan State University and is the PG Program Director of Computer Science at the University of Birmingham Dubai. He can be reached at: m.ilyas@bham.ac.uk