The text on this page was automatically translated and hence may differ from the original. No rights can be derived from this translation.
What has changed in the past five years?
In 2019, we saw the first applications based on deep learning – think of image recognition and text applications, like translation. “It was already clear back then that this would bring many opportunities but also have an impact on education.”, says Tommy. The topic of 'AI in education' has since become even more relevant. A follow-up paper that Nick and Tommy published shortly after the research, was initially barely noticed, but has been cited dozens of times in the past year. “There is attention for AI in education from all sides, including fortunately a lot of research interest”, says Tommy. In his current role, Nick uses language models for research and teaching at the university.
One of the conclusions back then was that 'artificial general intelligence', or AI that is as smart or even smarter than humans, was really still a thing of the future. “It still is, but with the introduction of ChatGPT something fundamental has changed: AI is no longer an abstract mathematical model, but 'feels' like a person you talk to, ask questions to.”, continues Tommy.
Did the research underestimate the development of AI at that time?
“I think, based on the knowledge at that time, we made some very good predictions”, says Nick. “We correctly predicted that AI would continue to develop significantly, but the specific direction it is taking now was impossible to foresee back then.”, adds Tommy. “I found it really surprising that something like a 'large language model', essentially an extremely large version of 'autocomplete', could work for generic tasks. In 2019, when you thought of chatbots, you envisioned 'closed' systems – more like a customer service menu where you could talk, but the options were limited. The language models break that open.”
AI tools are very disruptive for teachers. They need to find new ways to assess their students' knowledge and skills.
Another aspect that was overlooked in 2019 is the impact of generative AI – students can let the computer generate large portions of their assignments. Nick: “AI tools are very disruptive for teachers. They need to find new ways to assess their knowledge and skills of their students. Instead of essays, oral presentations are used more frequently, allowing teachers to ask follow-up questions.”
Has the debate about the risks of AI in education changed from five years ago?
In the 2019 research, it was noted that an AI – just like a human teacher, by the way – is not value-neutral. This may not be a problem, as long as you know what values exist and that they are in line with those the school stands for. The societal debate on AI in education is closely related to the broader discussion on the use of AI. “Around the language models, we also see, outside the educational context, a lot of discussion about the materials used to develop the language model, who owns it, and what influence that has. It is very good that this is being scrutinised. It starts with the language itself – a model like ChatGPT is very good in English, but less efficient in Dutch. Hence, work is being done on a fully Dutch variant.”, says Tommy. Nick would welcome more attention to the value proposition: “the manufacturer decides on the values in the final training step of the AI” – these values should be a crucial factor in selecting AI applications, but it is highly questionable whether they are transparent.
Are schools and teachers now more aware of AI? “I dare not say – during the research, we visited schools and saw significant differences. A school only needs one enthusiastic teacher to be at the forefront.” In scientific literature, attention is now given to the relationship between teachers and AI (e.g. Molenaar, 2022). Organisations like SURF are also engaged in the topic.
What does the next five years look like for AI in education?
“The previous research teaches us that it is very difficult to make statements about how specific implementations will be. But I believe that the conclusions from that time still hold true. I hope that AI can deliver on the promise of hyper-personalised education, without the negative impact that can also arise. It's good to see that this is being carefully considered.” says Tommy. Meanwhile, Nick is working on AI in education himself: “we are working on a language model that we ultimately can use to provide (future) students with information about the range of courses at our university”.
As innovation scientists, both closely follow the subject. “Adoption of new, disruptive technologies in society is always fascinating. My high school was recently in the news because they were one of the first to completely ban mobile phones in class. I found that interesting because when I was there, twenty years ago, mobile phones were already a thing. The school later allowed mobile phones and now banned them again. We must continuously evaluate innovations”.
