Puncturing the AI Bubble

An alternative economic strategy  which prioritises infrastructure based on human need and wellbeing is needed to tackle AI, argues Carl Rowlands, in the third of his series of articles.

How do we handle AI? The problem for socialists is that AI represents concentration: a concentration of money, processing power – hence the term “hyperscale data centres” – and energy.

Its use of training materials also represents a huge concentration. AI models have absorbed large swathes of the internet and huge amounts of printed materials. With its speculative circular financing and its dependence upon venture capital, this is not something that could form the basis of a nationalised industry.

This is an extreme version of what Stuart Holland described in his book The Socialist Challenge as “mesoeconomics.” It is oligarchical or even monopolistic, a distorted capitalism, operating outside of basic market structures. It only exists because of a concentration of money and power in tech, something which has developed from the 1990s onwards.

Pretty much all of the generative AI services currently run at huge losses, even for the versions that charge for premium services. These are services which, deprived of access to speculative finance and the resources provided by the biggest tech companies, would probably crumble to dust, as described here by Matthew Cole. 

It is helpful to break AI into its component parts. Firstly, the ‘infrastructure’. AI is usually situated in the ‘cloud’ – its emergence is very much linked with the stacked computing represented by hyperscale data centres. The scale was already building up, thanks to the move to SaaS (Software as a Service – for example, Microsoft 365, or Google Mail). This means that the actual computing is undertaken centrally. Most of the time, if one chats with Chat-GPT or Claude, this takes the conversation across the Atlantic to the US. The environmental impacts of all this are intense, but diffused globally, as I previously attempted to surmise.

Additionally, we should differentiate between “machine learning” and the use of “Large Language Models” (LLMs). Machine learning can be described as the identification of patterns. It might include surveillance, or activities such as robotic automation. LLMs can be described as “generative AI”. This is itself a form of predictive computing, whereby a chatbot might use millions of tokens, attempting to decode what it thinks should happen next, or what it should say next. 

Firstly, given the potential impact on employment of machine learning-driven automation (for example, the Starmer government’s pet project, driverless taxis), I would argue for an Automation Charter to protect the interests of workers and consumers. This would establish a clear set of guidelines, including the need for continual human monitoring, failsafe switches and the provision of a human-driven or analogue alternative. It would enforce compliance with the highest security standards. 

So far, this is an industry which has thrived in the dark. Actually detailing the impact of hyperscale (cloud) data centres is very difficult, not least because the large US corporations who have built most of them tend not to be especially forthcoming regarding precise details of water, power and pollutants. The general approach in the US has been to forge secret agreements with cash-strapped local municipalities and then form different sets of agreements with utility companies. In this sense, perhaps the UK as a whole can be said to be a cash-strapped municipality, or as the Starmer government preferred to describe the country, as a huge “AI Action Zone.”

Not only is the physical infrastructure shrouded in commercial secrecy, but the actual outputs are, too. There is widespread human intervention in real-time which influence the answers which AI chatbots provide – violating any security or privacy protocols, especially when so many users are increasingly turning to chatbots with extremely sensitive content.  This is just one aspect of the huge level of low-waged manual reviewing that occurs to attempt to ensure AI chatbots make some sense, some of the time. 

An AI Licensing Authority could regulate, evaluate and publicise all plans for large-scale data training and implementation. This would seek views across society, but particularly those in roles directly affected by the roll-out of AI. Its Advisory Board could consist of experts from around academia, the private sector and civil society. But its first task would be to establish true transparency: the corporate ownership and funding, the sources of the training data to as great a degree as possible, the different energy and resource inputs, known examples of reliability, known examples of bias, the role of (low-cost) human labour in manually reviewing outputs, and other such aspects.

These are the immediate steps which would begin, at least, to collect reliable information on what these huge systems actually do, and the secondary impacts. This is not an industry which is able to regulate or verify itself.

Demystification is critical. AI is built upon tech. It is all about IT and computer science, essentially. Regulate tech effectively, legislate against the huge corporations responsible for vendor capture and monopoly, and the questions around AI begin to become manageable.

There are already legal lines of responsibility. Ultimately, hyperscale AI has contracts which exist in the real world and these govern its implementation. These are lines of accountability. For public services, consideration should be given to hosting an official UK branch of open source AI models which provide specific functionality. These could be lightweight, low-intensity LLM models allowing for a greater degree of control and use for local applications. It may be that using LLMs to generate code – to allow codeless forms of software engineering – would only require a fairly light regulatory approach. These are perhaps questions worth further consideration.

Finally, it’s worth considering whether there should be local supervisory capacities for AI with jurisdiction to investigate local AI implementations – in a similar way to the regulation of consumer protection and food hygiene. This supervision and regulation should incorporate feedback and detailed proposals from trades unions and civil society organisations. Here we would have the beginning of a regime of inspection, at least – and to accompany that, we’d need. a move towards dispersing the concentration of infrastructure and control which hyperscale AI represents. 

But if such a regulatory regime emerges, it will be in a context where some professions have already been impacted by the use of LLMs. Copywriters, translators and generally, people whose work involves organising different sets of information, have already been directly affected. We just don’t have the numbers yet.

There are swathes of white-collar jobs which are thought to be vulnerable. We should assume there is a growing crisis of unemployment and underemployment, at least partly as a result of the implementation of AI, but that this crisis is being carelessly monitored. The government has largely abdicated its responsibilities. Not only does AI threaten jobs, it also ‘absorbs’ the previous output of everyone, including artists and musicians, while not acknowledging or crediting anyone other than itself, and the corporation providing the service.

Employers should be compelled to enter work agreements with trades unions to ensure retraining and gradual transition in cases of automation affecting roles. A Structural Adjustment and Green Transition Agency should oversee and guide these and other processes (such as de-carbonisation) which might increase structural unemployment.

But for socialists, taking responsibility is not just about regulation. It would mean planning an alternative economic strategy and the development of infrastructure based on human need and wellbeing. If we are unhappy with the social, labour and environmental outcomes from relying upon Big Tech, then we will need a comprehensive robust alternative, some of which would, I suggest, attempt to move on from having tech as a core growth dynamic. Modernity cannot be just about private equity, or data centres, or increasingly isolated individuals tapping into chatbots. A positive version of the future relies upon us depending upon each other, being in the same spaces together.

Just as there are social issues that correspond with each other, there are therapeutic approaches which can address multiple questions simultaneously. For example, there are significant capacity issues across the NHS, especially in the field of mental health, This has a wider economic impact as it results in labour force inactivity. It also contributes to the wider malaise affecting society. But there is a connection to environment – most town centres are seen as depressed. Various structural changes, perhaps most significant being the retail parks built in the 1990s and 2000s, would indicate there is little chance of reviving retail as it was.

The second part of this equation is that, partly as a result of austerity, there is a shortage of sports facilities where they are needed the most. There are no strong campaigns to encourage adult sports participation. There should be a programme of building sports facilities… and for increased investment in NHS mental health services. There is little doubt that this would help society!

But I would argue that the NHS can also be expanded horizontally, to fire up the private sector from the outside. For example, a pilot programme could offer everyone support for two optional health-related activities per month. This could include counselling and different forms of physical activity (including potentially a short physiotherapy session). The options could potentially include sessions with massage, individual or group exercise in gyms, swimming or on sports courts. This would act as a significant stimulus for town centres and the private sector, mostly affecting SMEs initially. The growth in personal services would be quite significant; we would perhaps expect a big demand for counselling and physiotherapy in particular.

This is an example of how a ‘therapeutic agenda’ could result in a people-centred form of economic growth. We can maybe glimpse a future of the ‘high street’ as a place of recreation. Society needs to decide the kind of economic growth we would want. Some of this may be in high-tech manufacturing, assisted by machine learning. Some of it, equally valid, may not.

Carl Rowlands works as a learning resource creator, teacher and writer. He is based in central Europe and the UK. Earlier articles in the series can be found here and here.

Image:https://www.picpedia.org/chalkboard/a/artificial-intelligence.html License: Creative Commons 3 – CC BY-SA 3.0 Attribution: Alpha Stock Images – http://alphastockimages.com/ Original Author: Nick Youngson – link to – http://www.nyphotographic.com/ Original Image: https://www.picpedia.org/chalkboard/a/artificial-intelligence.html