Artificial Intelligence, Artificial Literacies: Is the English classroom the place to develop AI literacies?

Published in Teaching English (For NATE members only)

The English classroom is in danger of finding itself the dumping ground where any in vogue notion of ‘literacies’ is to be addressed. However, talk of AI literacies raises particular challenges for subject specialists, indeed also for those with media studies backgrounds, for whom AI extends far beyond the corpus covered in traditional training in (digital) media. What does it mean to be AI literate?

As Steve Connolly (Chair of the MEA) and Mark Readman argue ‘literacies’ is a term used as a ‘shibboleth’, suggesting some kind of deficit in society which education is called upon to address. For example, ‘financial literature’, ‘computer literacy’, etc. It has become much detached from being literate in reading and writing. Although, in media studies, ideas about language and grammar have made their way into the study of non-print media such as film. Nevertheless, AI poses a particular problem for us: ‘the transparent opaqueness’ by which it operates. Simone Natale uses this term to emphasise that at the interface AI systems seem transparent because they are often based on a familiar chat-based structure: I ask a question, it answers. Nevertheless, the complexities of the programming behind that system and the rules applied by its creators to training are totally opaque to us. Essentially, we are always deficit when we, as users, are confronted with an AI system if the developers do not explain how it works, but we are under the illusion that we are not deficit which gives us a false sense of assurance regarding the authority of the output. We face a ‘deficit paradox’, which means that whilst we need to address the deficit, we do not recognise its existence. Thus, there is a real urgency for us all to have at least some working knowledge of how AI systems work. Otherwise, AI produced outputs are likely to shape common sense knowledge with an air of authority usually reserved for the nuance of expert writing. Why then would anyone waste their time reading lengthy, carefully researched books and articles in the future? What are the consequences for knowledge and truth if we let this rather dystopic imaginary become a reality?

AI literacy is potentially closer to its English rather than media counterpart. It requires us to understand how to read and write to it and with it, as well as comprehend what it writes back to us. This comprehension also requires context but might also benefit from the more creativity potentials of media literacies too. It does not, however, require us to start applying structural ideas of ‘AI language’ and ‘AI grammar’ as we have previously done with film analysis. If we want to use generative AI to support the creation of writing, as some teachers do, then we have a responsibility to teach pupils how AI works. We would not after all expect them to pick up a pen and write fluently without any support in how to hold a pen (although I appreciate supporting them to write in cursive might now be out of fashion!).

Before moving onto a more detailed discussion of what AI literacies might look like and whether the English classroom is the best place for it, let us briefly review how generative AI systems work:

Firstly, they are coded by human beings. In this code are algorithms that set up the restrictive parameters that give the AI program its limits, e.g., what it will and will not do. Sophisticated programs can then develop their own algorithms to further enhance their productivity and so-called ‘learning’. Even the most unsupervised models are rooted in some human-created code somewhere.

Secondly, they are train. They are fed data sets, the boundaries of which are selected by humans. There are controversies over how such data sets are tagged (or given linguistic meaning/ value). This is often performed by incredibly low paid individuals forced to work in terrible conditions and for long hours in central African nations (this is the Amazon Turk approach for example). Whilst the values imposed on the data sets are rooted in the particular Western ideologies where the corporations are based (most often the US). We have seen the result of this in our research, where ChatGPT would call out conversations it thought were slipping towards racisms towards African-Americans or Holocaust denial, but was happy to claim that the Bosnian Genocide was a disputed fact! Kate Crawford and colleagues have raised concerns about the historic ImageNet dataset which played an essential role in training many early AI systems. The human-images were removed in 2020 due to the tagging being full of racist, xenophobic, and misogynistic labels; not before it had already been used in potentially thousands of training scenarios, however. In most cases, generative AI is trained on open access data, the easiest and largest compendium of this is the World Wide Web. All WWW data is valued equally in this context – weighting is not usually given to peer-review content over other material. The status quo is that a populist Reddit post is given the same value as a peer-reviewed article by a subject expert. (This is not to say that such weighting is impossible!)

Finally, in retrieving information, Large Language Models like ChatGPT take a mathematical approach to words. They do not understand the language which they process linguistically. Rather, they scan instances of the arrangement of words in your request across the corpus of their training data, then apply probability measures to identify a line of best fit. This mathematical logic then informs the response you see on your screen. The way in which you simply get an answer after inputting a question gives the illusion of authority. Just as we rarely question our calculator when we add 2+2 (although for more complicated equations, it is just as important that we check out inputs here too! Just like AI, calculators have a very specific way of reading information). Google has retracted its AI search responses after outputs included examples such: eating rocks every day will help you keep up your essential mineral intake.

Several scholars, including myself, have challenged the idea of artificial intelligence, arguing that we should think more about AI models as adaptive systems. They respond to their environment, those making requests, and those supervising their development, continually altering the outputs produced striving towards a targeted goal of perceived ‘accuracy’. The concept of accuracy in AI tends to refer to whether the produced outputs match appropriate inputs (based on the line of best fit approach). The extent to which this might echo notions of ‘truth’ more significant in the Humanities cannot easily be determined – it will vary depending on who is training each model and what their intentions are for its goals.

I would suggest that to be AI literate, then, means to be confident in the following:

1 Understand the Workings of AI Models

This understanding does not need to be the level of a Physics or Computer Sciences PhD student, but rather a simple understanding such as the outline sketched above is enough. The basic need from this understanding is an appreciation of the ways models can be trained, recognising the potential ideological biases in training sets and tweaking processes, acknowledging the mathematical logics AI systems apply to language and images, and questioning the notion of their intelligence and automation. This also includes developing a critical awareness of the exploitative capitalism behind the development of many major AI models. Holding debate sessions, in which students are encouraged to research specific case studies and prepare a response to a question, such as ‘is AI good for society or not?’ would be one approach to this. Part of the literacy here is also thinking carefully about what to share about ourselves online given this is fair game for compilation into AI data sets for training.

2 Know the Right Things to Ask

AI models work best in response to ‘prompts’, many of which have been trained into the system. For example, ‘take the following details of my job history and format it into a CV’, or ‘make the writing of this paragraph more concise’, or ‘rewrite this song in the style of The Beatles’. Each of these prompts tells the AI system to focus on a very narrow data set (provided by the use) and asks it to transform the data in a very specific style. Such prompts yield far more useful results than an attempt to answer a media studies essay question. It essential, AI in this vein is reminiscent of Microsoft’s ‘Clippy’, the friendly but mildly irritating paperclip, who would disrupt your thinking to let you know that it ‘looks like you’re writing a CV, have you seen the template we have for that?’ (It was rarely a CV.)

We recently ran all of our media studies essay questions through CHATGPT and whilst it would mention relevant theorists, it did not offer quotations, and it paraphrased inaccurate and non-nuanced overviews of their work (all the work we had tried to undo in our seminars resurfaced in its responses). Furthermore, the answers were short, without critical engagement or detailed analysis of actual media case studies (even if we were quite detailed in our request). Sometimes the system would ‘hallucinate’ theorists and bibliographic references.

Allowing students to try out different types of questions and examine the quality of responses would be a productive group challenge. They could be set the task of considering what ‘values’ they apply to the outputs, and why they think these values are important when analysing information.

3 Be Critical of the Outputs

Just as we would encourage students to cross-examine sources, from historical documents to newspapers, we should also support them to check what other platforms and resources say in comparison to what has been generated for them by an AI system. This is a great task to help them discover for themselves the specific limitations of AI. This is not to assume that we should critique AI more than any other source, but rather than we should apply the same critical attention we already give to other sources to AI.

4 Design AI

Using creative work to get students to think through some of the challenges programmers of AI systems face is a great way to make transparent some of the processes involved behind the interface, without requiring a great deal of technical knowledge. For example, students could be given a particular scenario for which AI might be helpful and asked to consider what data would be appropriate for training sets, access issues you might have in getting this data (and thus the ethics of using it), some of the challenges in labelling it – are there easy ‘ground truths’ or is there some debate about how some datum should be labelled?

5 Learn to Code

Even a basic coding education introduces computational logics to pupils which highlight the different ways that computer systems interpret, and process information compared to how we do. Coding should be an essential part of our learning from the earliest days of primary through secondary and Higher Education. Unfortunately, it currently drops off after the most rudimentary of training due to the optionality that shapes later study.

From this very brief sketch of the workings of AI and what it might mean to be AI literate, we might question whether the English classroom is the place where this should be tackled. What it means to teach ‘English’ has of course changed over time and there are certainly activities suggested above that resonate with existing practice in English classrooms. Nevertheless, it seems like the best approach to teaching AI literacies would be interdisciplinary. I emphasise interdisciplinary here, however. It is not enough for it to be transdisciplinary, e.g., taught in different subject classes separately. Rather, it needs to be given space outside of disciplinary silos where the range of skills from English, media, history, computer sciences, art and design, mathematics and the sciences can all play a role. Given the rapid pace at which publicly available AI changes, a rigid curriculum would probably be limiting as it would likely be out of date very quickly. Rather project-based learning would be more suited to this.

The challenge of course is the rigidity of our current curriculum. We cannot ignore AI literacies and we have a duty of care to the younger generations to ensure they can navigate a world in which their entire lives are and will increasingly be shaped by such systems. We need to find the balance between short-term solutions: introducing extra-curricula challenges and long-term solutions: lobbying for interdisciplinary spaces in the curriculum (especially at Secondary and beyond where this is most lacking) that are fit for tackling today’s and the world’s future big issues. We also need to apply a critical perspective to our own uses of AI, to model the exemplary behaviour we might hope to see in our pupils as they encounter these systems more and more in their daily lives. Perhaps English and media teachers can take the lead with these solutions, if there is no one currently doing so in our institutions.

A recent post on the Media Education Association’s website invited leading academics to consider the role of AI in media education: https://www.themea.org.uk/post/ai-literacies-and-media-education

The original article can be accessed here. (For NATE members only)