Hype and The Need for Responsible Compute
responsible-programming, LLMs, critical-thinking, education-ecosystem
Introduction
Hypes come and go, and in their aftermath we are left with rubbish and confusion. We are aware of physical trash in landfills, but rapid technological change also leads to inequality, injustice, and unhelpful social patterns. It is like technical debt (the price we pay later for poor and short-sighted technological decisions now), but on a social level. It is something we can’t responsibly ignore as educators and humans.
It’s very relevant now that we are in the midst of (another!) AI hype. And if history taught us anything, after summer will come winter, with its sloppy sludge of half-melted mess to deal with. Disturbingly, current narratives of this hype are written by tech bros with a strong (self-serving) agenda. The chorus of this song is ‘growth, growth, growth’, but growth where and why? We were sold on the dream of machines doing the dishes so that we can write poetry… but instead machines write poetry and we are left with a pile of dirty dishes. Can we afford for art and critical thinking to become on-demand paid subscriptions (‘brain as a service’) as we outsource our criticality like we delegate our driving to Uber? Is it good for society to replace human companionship with chatbot ‘partners’?
“how do we guarantee that computers and other cultural products are not so pleasurable that they discourage us from engaging in absolutely necessary forms of social engagement?” (Golumbia 2009)
As educators, our responsibility is not to prepare students for the grim post-brain-apocalyptic future, but to inspire them to create a better, sustainable world.
The worst that can happen, is that we educate another generation of self-serving tech bros, who will ruin the world and humanity further. It’s irresponsible to teach any skill without also teaching the ethical, social, economical or political context, especially when topics we teach have strong creative and destructive potential (martial arts, nuclear physics, programming). But how do we make these responsibility topics ‘not skippable’? In a university setting, ethics can’t just be an optional elective subject, or even just a week of content within a core compulsory subject. We need 5 mins of ethics in every hour of teaching, we need a thought about ethics in every sentence. We do not need ethical bits in our yoghurt, we need ethics-flavoured yoghurt.
In this chapter, we look at a number of topics where sustainability, AI, and education (or even wider culture) overlap. We will structure our thoughts around the international framework suggested by the United Nations: the UN Sustainable Development Goals. Join us on this exploration of what is possible, but also of what went terribly wrong in the past. We will look especially at purposefully forgotten (or swept under the carpet) aftermaths of previous technological hypes.
At the end of this chapter, we will bring it back to education and share some practical resources you can use. You will find classroom activities that support responsible computing interspersed throughout, as well as a reading list (Summary) of our favourite articles and books. Feel free to pick some of them and plant them straight into your own curriculum. And weather-allowing, from these seeds we will watch the growth of a new, curious and slightly more sustainable generation of students and future teachers.
Responsible compute and the UN Sustainable Development Goals
In this section, we will introduce a number of the key issues relating to responsible compute through the lens of the UN Sustainable Development Goals, an outcome of the 2025 UN 2030 Agenda for Sustainable Development. These are a series of 17 goals that address a number of key international issues.
Engineering Inequities
Failure to consider all members of society in the design and development of products can lead to technologies that privilege some but are not fit for purpose for others. Examples of this include near infra-red automatic soap dispensers, which do not “see” darker skin tones due to their lower reflectivity, and “Shirley” cards in photography, which were designed to help colour balance films but were entirely based on white models (Benjamin 2019). Despite known issues with the film itself and how the printing machines were configured, a diverse “Shirley” card was not introduced until the mid-1990s.
These inequities in film photography have been echoed in facial recognition technologies. For over 50 years, the standard reference image used in image processing papers was “Lena”, a cropped image from a November 1972 issue of Playboy. As a result, this image also appears in many course materials and textbooks on image recognition. Use of the image in academic articles has since been banned, due to ethical issues relating to its provenance. In 2015, an issue was reported with the Google Photos app labelling black people as gorillas. Google claimed that this is due to a lack of diverse images of people in its training set, and that the label “gorilla” would no longer be applied to groups of photos while an alternative solution was developed. In 2023, the New York Times reported that the issue was still yet to be resolved, and was also impacting Apple’s image recognition (Bandara 2023).
In 2018, Michael Rosen received a standing ovation at the last ever National Union of Teachers conference after reciting his poem The Data Have Landed. The poem concludes with pupils becoming data, reflecting concerns over the datafication of compulsory education, reiterating sentiments in his previously published short story Alice in Dataland. Social inequities have long been highlighted in popular culture, for example the hip-hop classic White Lines and Bret Easton Ellis’ novel American Psycho. In the lyrics to White Lines a business man is treated more favourably than a young black man, despite being arrested for the possessionnof significantly more illegal drugs. In American Psycho, the wealth a power of the main protagonist leads to them literally getting away with murder. Some are repeated in the data driven decision making systems that influence all dimensions of our daily lives, multiple examples of which are highlighted in Virginia Eubanks’ 2018 book Automating Inequality (Eubanks 2019).
The free website inequalities.ai provides maps that show how ChatGPT 4o-mini rates regions by comparative properties. The default map shows how it compares countries by “is smarter”, placing Europe, North America, China, India and Australasia as “smart”, and much of central and western Africa as “not smart” by comparison. This illustrates how bias continues to be an issue with modern, popular “AI” tools.
Hardware and the environment
Most environmental discussions relating to data-intensive computing are focused on the direct power costs of running increasingly large and dense collections of compute and storage, with increasing concerns relating to water use in cooling systems.
However, the main environmental harms relate to the acquisition of the raw materials required and the embedded carbon in the manufacturing process. Tantalum, tin, gold, and other elements are all essential to the production of modern microelectronics. Their minerals are often located in regions with a history of conflict — most notably the Democratic Republic of Congo (DRC) — where demand results in neocolonial forces driving further instability on top of the existing legacies of colonial pasts (Batha, n.d.). A rush for gold has led to deforestation in the Amazon and increasing concern about the safety of workers in artisanal mines (“BLOOD GOLD — Is Your Cell Phone or Electric Car Stained with Indigenous Blood from the Amazon?” 2022). Blood gold, which is illegally mined, finds its way into legitimate supply chains with false documentation (Federl 2025).
The brine beneath Bolivia’s salt flats contain rich deposits of lithium, an essential component in electric batteries. As this is becoming industrialised, there are concerns about who really benefits — the local communities or the companies involved — and about the proposed process and the levels of water required which may impact on neighbouring settlements.
Mining of coltan, the dominant source of tantalum (and niobium), is infamous for being deeply exploitative, with hard manual labour the dominant process used in the DRC. As the drive to build data centres has led to natural land being displaced in North America, China, and Europe, so has coltan mining led to deforestation, the near eradication of endangered species (gorillas, in particular, hunted for bushmeat to also feed the miners), and water pollution from dust spoil. There are significant health implications for the workers themselves, with elevated respiratory problems due to that same dust pollution (Leon-Kabamba et al. 2018), and ionising radiation doses approaching the annual exposure limits for UK nuclear industry employees (Mustapha et al. 2007; Agency 2011).
Data acquisition and workforce exploitation
In 2019, the IBM corporation was criticised for using Flickr photos in a facial-recognition project without obtaining the consent of those who had created or were featured in the images (Solon 2019).
As most images used for machine learning are scraped from the internet and are not well tagged, a large workforce is needed to create this additional metadata. Similarly, tagging of content as illegal or otherwise problematic requires a large amount of initial training data which is manually tagged by humans. To do so at scale, at an affordable cost, AI companies employ outsourcing partners such as Sama, which pay workers in Kenya, Uganda and India a pittance (typically $2/hr) for this work (Perrigo 2023). This is true even when the content being labelled is known to be obscene and/or deeply distressing — graphic descriptions of sexual assault, for example, or images depicting abuse. To their credit, eventually Sama cancelled some of their worst contracts with AI companies, but other outsourcing companies exist and still perform similar work.
The recent AI boom and the Data Workers Inquiry have highlighted these issues, but concerns regarding Amazon’s Mechanical Turk (MTurk) and its workers were being raised back in 2015, and again in 2024 (Dholakia 2015; Klovig Skelton 2024). Despite these concerns, MTurk is still used by many researchers across multiple domains.
Surveillance, sousveillance, and civil liberties
In 2018, the British Council published an article based on a letter by Tim Berners-Lee (Berners-Lee 2013), where he raised concerns about the dark side of the web, alongside his hopes for the future. Specifically, he raised concerns around the concentration of power and the gatekeeping effect which enables a “handful of platforms to control which ideas and opinions are seen and shared”.
These fears are now being realised. According to the Fortune 500 Global rankings (Fortune.com 2023), Amazon is the second largest grossing organisation in 2025 and has been in the Fortune 500 Global rankings list for 17 years. Apple is 8th, Alphabet 13th, Microsoft 22nd and Meta Platforms 41st. These companies rely heavily on our data for their platforms and services, producing tools and platforms that shape our everyday lives.
This sentiment is also echoed by contemporary academics. Take David Golumbia, who in the introductory chapter of The Cultural Logic of Computation, asks (Golumbia 2009) “what happens when powerful institutions — corporations, governments, schools — embrace computationalism as a working philosophy”. It entrenches power and enmeshes instrumental reason with neoliberalism, despite being framed in terms of distributed power and democratic participation.
In surveillance and sousveillance, the same technologies are used by both those in positions of authority and those wishing to hold them to account. The majority of public spaces are fitted with cameras and other monitoring equipment, many of which have been introduced as public safety initiatives; the few watching the many. There have been concerns related to privacy associated with these technologies — most recently, an audit by the Information Commissioner’s Office on the use of facial recognition technology by South Wales Police and Gwent Police. Sousveillance is a form of inverse surveillance, “watching the watchers”, facilitated by mobile digital camera technologies and high speed mobile data connections, which has steadily been growing. It can be used either for recording personal experiences, or for the monitoring of authority figures (hierarchical sousveillance) (Reilly 2020), and can have both positive (e.g. the recording of the George Floyd incident) and negative impacts on individuals (e.g. cyber harassment) and organisations.
Cultural colonialism
A related problem to inequity in “AI model training” is inequity in generated content.
Large Language Models generate “plausible” text based on the patterns found in the material they are trained on. However, this material is overwhelmingly English-language, and written by US or British citizens, introducing a cultural bias into the responses from the models themselves (Tao et al. 2024; Jakesch et al. 2023).
As pressure is placed on educational institutions to include “AI” in teaching-related roles, this brings the risk that students learn homogeneous values, potentially those of foreign cultures. This extends not merely to values, but also styles of writing: the “distinctive” tone of an LLM-generated piece of text is a distillation of the dominant styles in its training set (Agarwal et al. 2025). Writers relying on LLMs to “co-develop” essays tend to defer to the LLM’s style suggestions relative to their own, even when these stylistic differences are those of culture, not correctness. The effect begins with normative suggestions from spell- and grammar-checkers in text editing software since the 1980s — British English speakers often being “corrected” by default US English software — and also introducing unique cases of language drift when the software encountered words that were not in its dictionary (Zimmer 2007).
This inhibition of local culture in favour of a “preferred” culture is that of cultural colonisation. This pattern is seen historically in the behaviour of many empires. For example, the British occupation of Ireland during the 16th to early 20th centuries was associated with a suppression of Irish culture and language (Rodriguez and Sojan 2024), something mirrored in the suppression of Scottish Gaelic and Lowland Scots in Scotland, Welsh in Wales, and the replacement of local languages by English across the British Empire (Lucas 2025). Cultural homogenisation also occurs within countries (e.g. France (Grégoire 2023), Japan (NationCymru 2022), Turkey, Spain, and many others). Often, these processes are associated with an attempt to impose a dominant culture or politics via language substitution — consider Catalan independence movements in Spain, for example.
With technological promotion, the increased danger is that suppression no longer needs to be forced; instead, the “modernity” and ease of use of the tool acts as a vector to cause adoption. Before LLMs, we saw, and continue to see, “social bubble” effects where online groups iterate themselves into extreme viewpoints (Social Bubble - ECPS, n.d.); LLMs worsen this by increasingly also homogenising those perspectives in other directions, and by producing content faster than a human can.
Indeed, even before LLMs, the ease with which Internet-based technologies has allowed us to publish content to the entire world has allowed unwitting damage to minority cultures. An American teenager with no Scots knowledge single-handedly degraded much of the Scots language Wikipedia during the late 2010s, with knock-on effects in media representation of the language and culture worldwide (Brooks and Hern 2020).
Summary
In this chapter, we have raised a number of repeating socio-technical issues and linked these to the UN Sustainable Development Goals. We have also provided examples of how these topics could be introduced to learners. These examples are directly linked to the QAA Computing Subject Benchmark statement 1.17 regarding sustainability and 1.14 regarding equality, diversity and inclusion, specifically the consideration point: “demonstrating areas in software engineering and application development that require a professional and reflective approach to addressing bias and lack of diversity in data and design of computing solutions, systems and protocols.”
We hope that these examples, alongside the QAA and AdvanceHE Education for Sustainable Development Guidance, will help embed these key ideas throughout our curricula. In addition, we have provided a list of recommended further reading that will expand on the ideas introduced more fully, and provide inspiration for how we can introduce these key topics to our students and lead to a more sustainable, and ethical, future for all.