What is generative AI?

While AI has frequently appeared in the news since late 2022, it is not a new technology. AI has existed since the 1960s, and humans have used it for many different applications: automated telephone services, chatbots, computer-based games, and satellite navigation tools. United Nations International Children’s Emergency Fund (UNICEF) defines AI as ‘machine-based systems that can, given a set of human-defined objectives, make predictions, recommendations, or decisions that influence real or virtual environments’ (UNICEF, 2021). This definition covers a broad spectrum of capabilities. Until recently, publicly available AI systems would decide on the most appropriate response or action following human interaction by using a range of predetermined responses or actions. This response could be complex, for example a navigation system deciding the best route drawing on information from maps, road closures, and traffic levels. However, these systems did not generate new responses, independent of the responses that were pre-programmed. In the case of a navigation system, this system is only capable of deciding the best possible route out of the pre-existing possibilities according to the available data and pre-coded rules on how to interpret these data. The navigation system is not capable of generating an entirely new route.

OpenAI’s release of ChatGPT 3 in November 2022 dramatically changed public awareness of AI and its potential possibilities. ChatGPT 3 was the first publicly available open-access generative AI. Unlike previous AI systems, ChatGPT 3 could simulate human learning to generate new content in response to human prompts. Following the release of ChatGPT 3, the number of AI systems and their capabilities have risen exponentially. These AI systems learn from data, but rather than directly copying information from them, they use the information to create new content. These data may come from databases known as large language models (LLMs) – an artificially intelligent system that humans have trained to comprehend and analyse human-produced text using human-generated content such as books and articles (Maslej et al., Reference Maslej, Fattorini, Perrault, Parli, Reuel, Brynjolfsson, Etchemendy, Ligett, Lyons, Manyika, Niebles, Shoham, Wald and Clark2024). Alternatively, more recently created or updated AI systems may use the internet as their source of data. This transition from databases to using the internet occurred in a matter of months.

Why is it important for teachers to know about AI?

The development of AI systems has been exponential over the last few years. OpenAI released the first public, free and open-access generative AI, ChatGPT 3, in November 2022. Within 2 months, ChatGPT 3 had exceeded a hundred million users (Hu, Reference Hu2023). By February 2023, while some universities sought to explore the potential opportunities of AI, 40% of UK universities had banned or were preparing to ban ChatGPT (Wood, Reference Wood2023). Economists at Goldman Sachs cautioned the impact of AI on the workforce – potentially replacing over 300 million jobs (Hatzius et al, Reference Hatzius, Briggs, Kodnani and Pierdomenico2023). At the time of writing, AI has become the fastest technology in history to reach widespread societal usage across the globe (Harrison, Reference Harrison, Araya and Marber2023). Unlike previous technologies which have spread gradually allowing knowledge of their use, capabilities, and limitations to develop over time, the rapid spread of AI has not allowed this. The ethics of using AI, efficacies, and pedagogies for learning about and learning with AI remain nascent.

Despite this, AI has already and continues to have considerable societal and educational impact. The UK Department for Education (DfE) advised schools that ‘[t]he education sector needs to: prepare pupils for changing workplaces; teach pupils how to use emerging technologies, such as generative AI, safely and appropriately’ (DfE, 2023). As more workplaces make use of AI to automate tasks or augment human capabilities, it becomes increasingly important that pupils become aware of how to use AI and the ethics of using AI not only to act as responsible citizens and be able to make critical judgements about AI generations but also to provide them with skills in AI literacy that they may need in the workplace as well as an awareness of issues surrounding AI-generated content such as deepfakes in the media (images, video, or audio depicting real or non-existent people with the intention of deceiving others).

In addition, a substantial number of pupils in the UK have begun to use AI for educational purposes. Recent estimates of the number of UK secondary school children using AI for education vary from 14% (DFE, 2024) to 25% (Bisoondath, Reference Bisoondath2024) and even 42% (RM, 2023). While there is considerable variation in these results, this may be explicable through the wording of the survey questions as some of the surveys asked about specific AI systems or specific contexts of use. However, the surveys do suggest that a significant number of school children are using AI. From my own observations, in many of my classes, I am aware that I usually have at least two pupils using or trying to use AI on tasks. As more devices are built with integrated AI software such as Microsoft’s Co-Pilot and Apple’s Apple Intelligence, it will become easier for pupils to access AI, especially as many schools transition to greater use of technology and pupil devices. Many new devices with Microsoft Windows come with Co-Pilot installed, and using AI is becoming as accessible as an internet search.

What are the potential disadvantages of using AI?

Many UK pupils have received little or no guidance from schools on how to use AI safely and responsibly. Consequently, many pupils are unaware of how to use AI to enhance their learning and may use AI in ways that are detrimental. If pupils use AI to complete a task for them, they are not accessing the learning intended by that task. While written communication may appear at a high standard, if pupils frequently use AI, their performance on tasks carried out without access to AI could potentially decline if they are not otherwise developing required skills (Brynjolfsson, Reference Brynjolfsson, Araya and Marber2023; Bartoletti, Reference Bartoletti, Holmes and Porayska-Pomsta2023). This risk of technology hindering learning is not new to AI and has been researched with other digital technologies (Mueller & Oppenheimer, Reference Mueller and Oppenheimer2014; Zierer, Reference Zierer2019). Using AI to provide answer scaffolding or support could potentially lead to reliance on the AI if pupils do not use the AI to develop their skills and confidence to use these skills independently of the AI. This is akin to giving pupils a scaffold every time but never reducing it and challenging pupils to progress to a point where they no longer need it. A recent survey found that many pupils were concerned about their performance in exams when they did not have access to AI (RM, 2023). This does not necessarily mean that pupils are copying and pasting AI-generated content to use as their own work. Doubtless, some are; but other pupils may be using it because they recognise it can help them learn yet are unaware how to use AI effectively to reduce reliance on it. There is a risk that if AI use is merely to replace or automate existing tasks that it will decrease human agency and hinder learning (Bartoletti, Reference Bartoletti, Holmes and Porayska-Pomsta2023; Mueller & Oppenheimer, Reference Mueller and Oppenheimer2014; Zierer, Reference Zierer2019). Puentedura’s Substitution, Augmentation, Modification, and Redefinition (SAMR) model set out four ways that people can use technology (Puentedura, Reference Puentedera2014). Puentedura argues that, if technology is merely used to replace (substitute) or augment (substitution with some functional improvement), it does not necessarily lead to any learning gains and may hinder learning, as pupils do not sufficiently develop other skills (Puentedura, Reference Puentedera2014, Zierer, Reference Zierer2019). Alternatively, new technologies to transform tasks and what is possible (modification and redefinition) will promote higher-order thinking and provide tasks that are high in challenge but low in threat to pupils (Zierer, Reference Zierer2019). Using new technologies to substitute or redefine existing capabilities may mean that users develop different skills. For example, the rise of navigational systems has led to a decrease in map-reading skills. This could potentially exacerbate existing social inequalities as pupils’ attainment suffers through overreliance on the assistance of AI (Brynjolfsson, Reference Brynjolfsson, Araya and Marber2023). This is an area that is not well researched at present, but teachers need to be mindful of what skills their pupils should develop and whether any technology use potentially hinders their development of them. AI is a technology, and like any other technology, pupils need to learn the skills for using AI and understanding of how to use it responsibly (Shiohira & Holmes, Reference Shiohira, Holmes, Araya and Marber2023). While schools devote considerable time to teaching pupils to use technologies, at present, some educators (unintentionally) expect pupils to understand the ethics of using AI or how to use AI without providing the same extent of clear guidance as with other technologies.

AI also poses risks for academic integrity. It can be incredibly difficult to detect AI-generated work as it can complete most tasks with varying degrees of success. While anti-cheating software, such as Turintin, claim to have the ability to detect AI, this software relies on AI systems to detect AI and cannot reliably detect it, often flagging AI-generated content as human generated and vice versa (Ahmad, Reference Ahmad2024; Oxbridge Editing, 2023). At the time of writing, it is not possible to detect AI usage conclusively by technology alone (at least in primary and secondary education, teachers are likely to be familiar with their pupils’ style of work and prior attainment; however, there are other ways to identify AI usage, as I explain in the next section). Moreover, it is impossible to prevent AI usage outside of school and increasingly difficult to prevent pupils from using it within schools. Strategies to prevent AI usage on tasks are also becoming less effective as AI systems become more powerful.

As AI is a learning technology which generates new content rather than providing a series of pre-existing results, such as a search engine, it also produces flawed results. Similar to a learner, AI can and does also make mistakes. It can invent new (factually incorrect information), a phenomenon known as ‘hallucination’. In addition, AI may not necessarily be able to appropriately evaluate between conflicting sources of data (Shiohira & Holmes, Reference Shiohira, Holmes, Araya and Marber2023). Consequently, AI can replicate and exaggerate biases existing in its data set or make content on the basis of factually incorrect or completely irrelevant data. On one occasion, my attempt to ask ChatGPT for a model answer for a GCSE Classical Civilisation question for the unit on Roman City Life, resulted in detailed (but incorrect) instructions on how to carry out a completely irrelevant Chemistry experiment. On another occasion, an AI pretending to be an Ancient Egyptian provided a history of global carrot consumption when asked about their favourite food. While comical, this can also be confusing and frustrating for teachers and pupils.

Few AI tools are transparent about their data origins or cite their sources without prompts. Users need to be aware to fact check the information provided by AI and be mindful of the potential for biases. Many LLMs (used to train AI systems) are based on data created before 2021 (Maslej et al., Reference Maslej, Fattorini, Perrault, Parli, Reuel, Brynjolfsson, Etchemendy, Ligett, Lyons, Manyika, Niebles, Shoham, Wald and Clark2024). Therefore, these systems do not have access to more recent information. While this may not seem like a significant problem in Classics, it can affect how users can interact with AI. If users ask the AI to cite its data or ask for recommendations of websites, these links may be outdated and no longer work or even non-existent.

AI systems also raise data protection concerns. Once a user provides an AI with information, the AI will learn from this information and draw upon this learning to answer other users (Moares & Previtali, Reference Moares and Previtali2024). Many AI systems require users to input their personal information to a website to create an account or ask for access to additional information on a device or digital accounts such as contacts of Google Drive files. The issue is that these AI do not always make their third-party affiliations and data storage locations clear or explain how they will use these data (i.e. whether they will share it with third parties (O’Rorke & Talbot Rice Reference O’Rorke and Talbot Rice2024). The complexity of finding this information means that users may potentially risk their personal information. While deleting an AI system chat or account may give the impression of deleting the data, this is not necessarily the case. In addition, users need to take care not to input personal data or intellectual copyright (both of published works and of pupils) to avoid breaching relevant copyright or data protection legislation such as GDPR in the UK. A further issue with copyright is that AI can unintentionally infringe copyright through creating content that closely resembles existing work.

What are the benefits of using AI?

Despite the limitations, AI provides considerable opportunities for education, such as aiding with teacher workload. AI is good at completing tasks that are useful but time consuming, such as the creation of resources, templates for letters, or creating adaptive resources. AI can also assist teachers with planning and lesson preparation as well as marking. AI tools take a bullet point list of feedback for a report or UCASFootnote ² statement and turn it into prose. The capabilities are useful for teachers with limited time or those who are less experienced, such as early career teachers. It also allows teachers to be more efficient as they face increasing demands on their time. AI also facilitates adapting existing resources such as set texts to the needs of pupils and enables teachers to provide greater support than they may otherwise be able to.

At the same time AI can support pupils as a learning aide. It can guide learners on how to plan tasks and provide non-judgemental answers to questions that pupils may be nervous asking their teacher or their peers. Surprisingly, some AI tools can also be supportive and encouraging when pupils make mistakes or lack confidence. I have observed interactions where the AI provides guidance to pupils to help them through a task and build their confidence on the task when pupils told the AI they did not understand what to do. Another method is to use AI as an alternative to a search engine. While pupils risk finding incorrect information, I have seen pupils use the AI to considerable effect when they ask it to recommend websites. AI can act as a filter, providing pupils with a smaller selection of relevant websites to start their research, reducing the cognitive load. Pupils can also ask AI to provide them with model answers, sentence starters, or examples of persuasive language that they can use to improve their communication. I have observed pupils take agency in their learning by using AI to find methods of improving their work, for example by asking for suggestions on structure, asking for examples of persuasive language, or asking how they could show one of the success criteria. These pupils showed awareness of the qualities of a good answer but were unsure of how to apply these qualities, and the AI supported them by providing a range of examples. For pupils who have English as an additional language, AI can help them readily access content by acting as a translator or providing glosses or key vocabulary to learn.