Imagine a world where a machine could offer students personalised feedback, generate new content tailored to their needs, or even predict their learning outcomes. With the rapid emergence of generative AI – notably the likes of ChatGPT and other large language models (LLMs) – such a world seems to be on our doorstep (Kasneci et al., 2023). However, as the horizon of education broadens with these advancements, we must also consider the maze of ethical challenges that lie ahead (Schramowski et al., 2022).
Educational research has seen accelerated growth in its relationship with LLMs, as evidenced by our scoping review of 118 peer-reviewed empirical studies (Yan et al., 2023). These studies unveiled that LLMs have found their way into a staggering 53 types of application scenarios in the automation of educational tasks, ranging from predicting learning outcomes and generating personalised feedback to creating assessment content and recommending learning resources.
While this paints a vivid picture of the vast potential LLMs offer in reshaping educational methodologies, the picture is not devoid of challenges. Many of the current innovations utilising LLMs have yet to be rigorously tested in real-world educational settings. Furthermore, the transparency surrounding these models often remains confined to a niche group of AI researchers and practitioners. This insularity raises valid concerns about the broader accessibility and utility of these tools in the educational sphere.
Issues of privacy, data usage, and the looming costs associated with commercial LLMs like GPT-4 add layers of complexity to this discussion. Beyond the financial concerns, the ethical ramifications of how student data is handled, the potential for algorithmic biases in educational recommendations, and the erosion of personal agency in learning decisions also present significant challenges to widespread adoption. One can’t help but ponder: Are these technologies primed for widespread educational adoption, or are they reserved for those who can navigate the intricacies of AI and afford the associated costs?
‘Are these technologies primed for widespread educational adoption, or are they reserved for those who can navigate the intricacies of AI and afford the associated costs?’
From our scoping review, three central implications emerge:
First, while there exists a golden opportunity to harness state-of-the-art LLMs for pioneering advancements in educational technologies, it is imperative to use them judiciously. Innovations in areas such as teaching support, assessment, feedback provision and content generation could revolutionise the educational landscape, potentially reducing the burden on educators and enabling more personalised student experiences. However, the economic implications of commercially driven models like GPT-4 might make this vision more of a dream than a reality.
Second, there is a pressing need to elevate reporting standards within the community. In an era dominated by proprietary AI technologies like ChatGPT, transparency isn’t just a lofty ideal; it’s a necessity. To foster trust and facilitate wider adoption, it is paramount that we advocate for open-source models (for example, Llama 2), detailed datasets and rigorous methodologies. This isn’t merely about boosting replicability; it is about engendering trust and ensuring the tools we advocate for align with the educational community’s broader needs.
Lastly, but by no means least, is the urgent call to adopt a human-centric approach in developing and deploying these technologies. Ethical AI is not merely about sticking to a checklist of principles; it is about weaving human values into the very fabric of these systems. Engaging stakeholders, from teachers and policymakers to students and parents, in the process of developing, testing and refining AI technologies ensures that the technology serves the community rather than the other way around. When these systems make decisions that impact real lives, those affected should not only be aware but should have a deep understanding of the rationale, potential biases and associated risks.
In conclusion, generative AI and LLMs, with their tantalising capabilities, are indeed a double-edged sword. They promise to revolutionise education but come with a fresh set of challenges concerning ethics, transparency and inclusivity. As these models steadily weave themselves into our educational fabric, an active, continuous dialogue among all stakeholders is crucial. In navigating this brave new world, we must ensure that technological advancements are both ethically sound and genuinely beneficial, leading us not just into the future of education but a brighter future for all.
This blog post is based on the article ‘Practical and ethical challenges of large language models in education: A systematic scoping review’ by Lixiang Yan, Lele Sha, Linxuan Zhao, Yuheng Li, Roberto Martinez-Maldonado, Guanliang Chen, Xinyu Li, Yueqiao Jin and Dragan Gašević, published in the British Journal of Educational Technology.
Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., … & Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274
Schramowski, P., Turan, C., Andersen, N., Rothkopf, C. A., & Kersting, K. (2022). Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence, 4, 258–268. https://doi.org/10.1038/s42256-022-00458-8
Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R., Chen, G., Li, X., Jin, Y., & Gašević, D. (2023). Practical and ethical challenges of large language models in education: A systematic scoping review. British Journal of Educational Technology. Advance online publication. https://doi.org/10.1111/bjet.13370