Introduction
With the advancement of large language learning models (LLMs – AI tools such as ChatGPT), there are senses of optimism, potential and scepticism within the education world in equal amounts: optimism regarding its potential impact on reducing workload, a key factor in the recruitment and retention crisis (NFER, 2024); potential in its ability to accurately assess and improve the quality of teaching (and thus also improve recruitment and retention); and scepticism about its use entirely. It seems inevitable that AI will play a part in the future of our profession, with nearly 60 per cent of us having used it in some form (Hallahan, 2024). My own position has matured over time: while, on balance, I am optimistic about its usage, I would argue that it needs to be more gradually introduced. In the past, technological innovation has moved past our ability to fully understand and regulate its use (Herane, 2024), and the same is happening with AI. Organisations such as the Turing Institute have advocated for the implementation of AI following the principles of fairness, accountability, transparency, privacy and security (FATPS) (Memarian and Dolek, 2023), which, as a framework, is one that I feel is our duty as educators to consider as we integrate AI into our practice further.
Fairness
Fairness can be best applied through two different lenses, one looking at ourselves as practitioners, in terms of our work–life balance, and one looking at our students. One of the persuasive positive arguments in favour of AI is its potential impact on teacher productivity and workload. Help in this regard can be best realised through the use of programs such as TeachMate AI, MagicSchool AI and Aila – closed systems created by educators. However, the research base for their effectiveness is limited, so the benefits of their usage are anecdotal. Early research includes the recent Education Endowment Foundation study, which found that teachers using ChatGPT to produce resources spent, on average, 31 per cent less time per week without compromising the quality of output (EEF, 2024). However, the study doesn’t specify the level of quality previous to AI usage. Outside of this trial, the quality of AI-generated outputs for resources and lesson planning can vary wildly – mitigated slightly by the closed AI systems. However, if tools such as ChatGPT, Gemini or Bard are used, it very much depends on the prompt that is given to the tool. Even with a quality prompt within the tool, the accuracy of the information provided to you can be suspect. AI tools are programmed to be certain that what they are telling you is correct – but all they are doing is predicting what the next words in the sequence that they produce will be (Williamson, 2023). If AI is being used to bolster subject knowledge, we may not automatically realise that the information generated is incorrect. This is particularly applicable within primary settings, where we need to have knowledge across a variety of disciplines, and there will always be subjects in which we have more knowledge than others. However, if we are relying on AI to do the heavy lifting for us in terms of planning and it’s wrong, we are in danger of defaulting on Teacher Standard 3 (DfEDepartment for Education - a ministerial department responsible for children’s services and education in England, 2011).
The other potential impact on workload relates to assessment. You can upload pupil work to AI and it will offer recommendations based on what it is programmed to analyse. As a teacher within Year 6, I can see the benefits of this when the spectre of Key Stage 2 moderation looms – a tool that can spot grammatical devices and offer edits and next steps is a ‘promised land’ that is hard to resist. The caveat is when we consider the algorithmic bias present in AI. AI can only generate content based on what exists – and, of course, it has no morality (Deshpande et al., 2023). If you upload the work of a pupil with SEND (special educational needs and disability) or who speaks English as an additional language (EAL), with no qualifying information as to these external factors that can affect a child’s work, then the algorithm will assess it in the same way as any other piece of work, thus causing the recommendations to be unsuitable for that child. This problem is exacerbated further if we have anthropomorphised the AI tool that we are using.
Accountability
Anthropomorphisation is treating an object as if it were human, and this is something of which we are all guilty when dealing with technology (and the tech industry itself capitalises on this – how many of us have an Alexa in our home or use Siri?). Many AI products have been designed with this in mind, as Deshpande et al. (2023) attest to. The risk of this in an educational context is considerable when you bear in mind that its use can be perceived as a dialogue between you and a helper. However, AI cannot think; it is not human and should not be treated as such (Deshpande et al., 2023). If we accept its findings in our students’ work as fact, we are reinforcing the algorithmic bias or being influenced by racial, gender or societal bias (of which AI has no concept).
We also do not know the criteria that have gone into the decision that the AI has generated with respect to a student’s work. We can be professionally challenged to back up our assessment decisions; however, we cannot do the same for AI. If we are not challenging the AI’s recommendations, and the assessment recommendations that we are using are inaccurate, then we are defaulting on Teacher Standard 6 (DfE, 2011). To my knowledge, only Oak Academy’s AI tool Aila is open-source (Searle, 2024), where the code behind the tool is public-facing. Even then, a substantial amount of technical literacy is needed in order to understand it.
In addition, if we are treating AI as human, then we are holding it accountable for decisions and planning for which we need to be taking accountability ourselves – morally, our students deserve us to be more than a passive participant in these processes.
Transparency
This moral element to the argument deepens when we take into account the thoughts, feelings and emotions of our students with respect to AI. Barnardos conducted a study where they found that young people’s feelings surrounding the issue ran from excitement to dread, amid fears of AI taking over, privacy and their future job prospects (McLoughlin and McMullen, 2024). If we are feeding AI details regarding our pupils or we are using AI to produce resources or assess their work, should we be making it clear that we are doing so? Again, the risk of this is doubled with a tendency towards anthropomorphisation – if it’s clear that we’re treating AI tools as a key part of our decision process or we openly refer to it as ‘he’ or ‘she’, are we risking ruining our credibility as a practitioner in front of our students?
In addition, one of the big selling points of products such as TeachMate AI is the existence of tools such as report writers. The research base for the efficacy and quality of using AI for this purpose is limited, but the potential time saving is massive for educators. However, with the concerns surrounding bias and the usage of AI to make decisions regarding our students, the moral argument is clear. Furthermore, the school report process is something that happens just once a year in most primary schools. It’s the one time at which children and their parents will receive a report about them as a learner, and while we all have stock phrases that will creep in, reports should be written as if we know the student. If we do choose to use AI, I would argue that transparency is paramount and should be reflected in the AI policy set out by the school.
Privacy and security
Within this aforementioned AI policy, held by either the school or the trust, will be how the generative AI tools used by the institution comply with the General Data Protection Regulations (GDPR) (DfE, 2023). Said policy should clearly outline how the tools listed within will meet the requirements for the seven principles set out by the GDPR (DfE, 2023). This works well for those tools that have gone through this due diligence; however, if we are using tools that aren’t approved by the school or trust (such as ChatGPT, arguably the most prevalent example of an AI tool) and we are inputting personal or even identifying information regarding our pupils, then there is substantial risk of a GDPR issue. Schools and their staff are prime targets for cyber security attacks, with 52 per cent of primary and 71 per cent of secondary schools experiencing a breach or attack in the past year (DSIT, 2024). It is only a matter of time before an AI tool experiences this as well. While most practitioners are aware of the risks surrounding data protection, thanks to robust training processes across LEAs (local education authorities) and MATs (multi-academy trusts), it may not be as readily apparent to a lay person that there is a risk in sharing information with these tools. If we are doing this, we need to be taking all measures possible to anonymise anything personal or identifying – especially because our students have not necessarily given us explicit consent to use these tools. These concerns with privacy and security should be at the forefront of our minds when using AI tools – technological transformation doesn’t always progress with the needs of children as a vulnerable and protected group in mind (Livingstone et al., 2017).
Conclusion
While the balance of this article highlights the risks surrounding AI, for the most part I am a proponent of its use. I do believe that it will form a part of educational practice, theory and discourse in the coming years. However, despite having been around for a while, AI is still a nascent technology in many ways, and we should be cautious about its usage. As a profession, we should absolutely engage with it but should do so with the needs and safety of the students within our care at the forefront of our minds, engaging with politicians and the creators of these tools to ensure that they are safe. Above all, we need to remember that we are dealing with tools and not people. If we fall into the trap of anthropomorphisation, then we run the risk of taking the recommendations of these tools at face value, reinforcing algorithmic bias. If we are outsourcing all our planning and resourcing to these tools without quality-checking the outputs, then we run the risk of undermining our own subject knowledge and credibility as practitioners. Without considering these factors, and the principles of FATPS, the sceptics of AI within education may be proved correct.