In March 2025 Google officially rolled out its AI Overview feature. Every user of the site will have observed it making glaring factual errors (those who haven’t simply did not notice them). This move has proven unpopular for the company yet they have no plan in reversing the decision.
Below are two of the most glaring factual errors which reflect the biases these ChatGPT-style AI features contain and the issues with them. They provide an overview on the social benefits of nuclear war, and name Barack Obama as America’s first Muslim president. They are found in a listicle of 17 examples, any of which could be used to illustrate this farce.
(Image Credit tomshardware.com)
While AI has only recently become a media talking point following the launch of Chat GPT, programmes such as these have been around for a long time and have already shaped the world in a major way. Social media companies use AI to determine what our feeds look like, recruiting firms use AI to review CVs and insurance companies use AI to determine whether or not to pay a claim. These are all longstanding practices which predate the most recent AI boom and are worthy of examination. This article will look at two distinct eras for AI, pre and post the launch of Chat GPT which catapulted AI to the forefront of popular discourse. Firstly though, it is worth examining what exactly AI is before discussing how it has been used and its consequences.
What is AI?
Artificial intelligence (AI) is a very misleading name: these programmes are not intelligent, as this article will go on to show. What is specifically being referred to when the term AI is used is what is referred to as Machine Learning or Neural Networks. So what is a neural network and how do they work?
One way of thinking about how traditional computer programmes work is that they process a series of logical statements of the form if this, then that. For example, a traditional computer code that was designed to tell the difference between images of squares and circles could count the number of corners (assuming there is some programmed definition of a corner):
“If the number of corners = 4 then it is square
Else, if the number of corners = 0 then it is a circle”
The thing to focus on here is the clarity on how the computer is arriving at this decision, the logic is transparent.
Neural Networks on the other hand are a lot less clear. To programme a neural network to perform the same task – identifying squares vs circles – think of an analogy with physical wires and connectors:
- Start with a black box filled with thousands of wires and connectors that are connected randomly at first. They scan a photograph in and spit out an answer: square or circle.
- Next, this box has to be configured so that it will give the right output. This is referred to as training. To do this, thousands of photos, correctly labeled as squares or circles, are required (which necessitates human labour to label them).
- Test how the box performs: feed each photo in as an input and observe the answer the box produces. In this initial test it will be answering randomly, getting roughly 50% right.
- After that first test, small changes are made to the box, to the connections between the wires and connectors. (Complex mathematical techniques are used here to determine what changes to make but for the sake of this simplified analogy of physical wires rather than nodes with weights, it is sufficient to think of it as a random small change to the current set up of the wires. Over many of these small changes the final product will be arrived at).
- Run the test again, observe if the results improve. If they do, great. If not, reverse the changes.
- This improvement will only be small, so the process has to be repeated many many times: edit, test, repeat, edit, test, repeat…
- Over vast numbers of iterations, eventually some configuration of the wires and connectors is found that works. It will correctly identify the circles and squares in the training set. The AI has been built!
In terms of Machine Learning, then, the boxes and wires of our analogy are in fact numerical values and sophisticated mathematical functions.
The are three important takeaways from this simple analogy:
- What results is not intelligent. It is limited to the scope of what it was trained on. If this machine were to be shown a triangle or an octagon it would simply label them as either square or circle as it was never trained to identify triangles or octagons.
- It was only possible to train the box because we had a vast amount of data to run the tests with. If the machine was less than perfectly accurate, the way to improve its accuracy is to find more labeled photos of squares and circles to train it on. More often than not the labeling of these data sets are done by workers in the global south on slave wages.
- Nobody understands why the model — this particular arrangement of boxes and wires — churns out the answer that it does. It’s clear from testing that it does, and we do know how to set up the frameworks which train models and select the ones which give consistently satisfactory results, but we don’t really know what’s going on ‘inside the box’ so to speak.
All of the AI models that will be later discussed, no matter how complex, fundamentally work in the same manner as this analogous black box of wires. And the same three takeaways apply to all of them.
Pre Chat GPT: Google and advertising
AI has already shaped both the internet and modern capitalism. In particular, the advertising industry has led this change with Google as the largest and most dominant player in the field. (While not thought of principally as an advertising company, ad sales account for more than three quarters of revenue at Google, and all rely on AI.)
In the mid-1990s Google was a small venture capital backed start-up with the goal of organising the internet, and from the beginning it was one of the pioneers of AI. Very early on, Google started to gather data on its users; what they searched for, what they then clicked, etc. That data would provide the training material for its AI which was tasked with sorting search results. As its user base grew, so did the amount of data it gathered, further improving the AI, which improved the product, attracting more users and therefore more data – a ‘virtuous’ cycle for the company.
Its founders were initially very hesitant to sell ads on the site – in 1999 the entire ad department consisted of just seven people. Ads were simple: buyers would pick which key words they wanted their ads to appear next to. A table manufacturer might have placed an ad wherever a user searched for “table” or “kitchen”. This was sufficient in the era of seemingly endless venture capital money to burn through; the company had little pressure to make revenue while its stock price continued to go up.
All this would change following the dot-com crash at the turn of the millennium. Suddenly the patience of the venture capitalist backers had run out and Google was under intense pressure to increase revenue. The seven-person ad team would find a solution that would shape the internet for the next 20 years. As co-founder Larry Page later put it, “Advertisers shouldn’t be involved with key word selection, Google should decide”. From then on Google’s AI would pick where to place ads such that they were most likely to be clicked.
In 2003, Google filed a patent for “the generation of user data for use in targeted advertising”, changing radically how the company operated. They were from this point onwards fundamentally an ad agency with tech bells and whistles. Google used the user data that they had been gathering for years to build AIs that would predict what to show to which person and at what time. This was remarkably successful with Google ads achieving a much higher click-through percentage than their rivals who were still relying on key words. Google shifted to selling ads per click rather than per view, meaning that any improvement to their AI ad selection translated directly and immediately to increased profit for the company. The race was on to improve it.
The next 15 years would prove to be a gold rush for user data, with personal data overtaking oil as the world’s most valuable commodity. No piece of data was too small to be usable and as terabyte after terabyte was added to the training set the prediction AIs got better and better, and revenue went up and up. Google’s harvesting of data was done by any means necessary, largely in secret, with little to no attention paid to the laws of whatever region they were operating in. For example, when Google Streetview was being launched, as the cars went around photographing the streets they carried onboard devices which would automatically try to connect to nearby wifi networks to scrape user data. Or consider how the Android operating system is given away for free; whatever money the company could earn from selling it pales in comparison to the value of the data it can harvest from the increased number of users using it for free.
This new data goldrush would spread beyond Google into the entire tech world. No matter what kind of website, app or electronic device you were selling, its usage would generate some user data, which could then be sold to feed the AIs and make better ad predictions. This is not just limited to online space, all “smart devices” with an internet connection feed data back to train AIs.
All of this required immense computing power in the form of enormous data centres. These centres come with similarly immense energy and water demands. Between 2000 and 2010 the world saw a 5-fold increase in the number of data centres from 100,000 to roughly 500,000. It was during this era that Ireland became the data centre hub of Europe, and the situation would only get more extreme with the advent of generative AI.
Chat GPT and generative AI
AI in the public consciousness changed in 2022 with the launch of Chat GPT, made by OpenAI. ChatGPT is an example of what is called generative AI –bots that can produce seemingly original text, images and videos. I write ‘seemingly original’ because, for all the hype, generative AI is still the same technology as our circles or squares example. It is a black box trained to produce something that looks like it was written by a human, and judged on the extent to which it looks like a human wrote it. It is ‘taught ’to do this by training it on every piece of text that could possibly be scraped off the internet, without ever compensating the original creators. This type of AI is known as a large language model; they were originally designed to be text generators for writing and even its own developers were surprised when the text generated appeared to be “intelligent”.
Like the ‘squares and circle’ AI, generative AI is not intelligent. The illusion of “intelligence” comes from the training data. If the prompt asks about Roy Keane walking out of the Ireland squad in 2002 before the World Cup, Chat GPT can competently plagiarise some combination of the many thousand column inches that have been churned out over Saipan. Crucially, there is no intelligence at play here. There is a somewhat sophisticated mathematical process occurring, sure, but it is not capable of thinking, it is not capable of creativity and it is not capable of producing anything authentically new.
This is also the source of an unsolvable problem for AI: hallucinations (a euphemism for lying). If there is no answer the AI model can readily find, it will spit out a lie. Why? Because it is a text writing device, it has no conception of the truth, reality or what it is even being asked to do. It just wants to write something coherent and passable for what a human might write. “You will never get them to not hallucinate,” one former Google executive and founder of a generative AI agent start-up recently admitted.
Why, then, are large language models spoken of as if they are about to launch skynet? Every day articles are written about the oncoming AI wave. Ireland now has a government minister for AI (Niamh Smyth, Fianna Fáil). According to AI evangelicals like OpenAI’s CEO, teachers are going to be one of the first professions replaced by AI, quickly followed by doctors! (Consider what it might mean to purposefully roll out technology with in-built hallucinations to ‘aid’ the education and healthcare of human beings.)
The explanation for the wall-to-wall hysteria is that the hype boosts share prices. Silicon Valley is in a perpetual boom-bust cycle, from the dot-com bubble, the automation hype of the early 2010’s, Crypto bubbles, and more. It is an industry dominated by PR smoke and mirrors and generative AI is no different. The coverage raises share prices, making money for the venture capital funds behind the tech industry, who make their exits before the pop.
This latest bubble has engulfed the tech world with all the major tech players pivoting massively to AI. Last year approximately one third of US start-up investment went to AI-related projects. At the beginning of the year the Trump administration pledged that $500bn of investment would be going towards AI. The utopian vision of the future these tech giants promise will never come to pass, however the cost they are paying in its pursuit is very real. Generative AI has been a climate disaster.
Generative AI is incredibly resource-intensive and requires enormous data centres to function. These data centres were already built for AI projects from before the emergence of generative AI, but the speed of their construction has accelerated to keep up with the new demand. Training the ‘boxes of wires’ depends on this gargantuan computing power, and the boxes never stop needing to be trained. Microsoft alone now has 5GigaWatts of servers, using more energy than the entirety of Hong Kong. They are joined by Google and Amazon who have even larger server farms. Billions of Euros are now being funnelled into the construction of new data centres each year, accelerating the climate catastrophe at a moment where the emergency breaks need to be pulled.
Data centres are a plague on the environment, consuming vast amounts of both carbon and water at a time when we need to be conserving both. The already-insufficient climate targets of the tech giants have been completely dropped, and with that has gone the pretense of the softer, greener capitalists in the tech industry. Google’s emissions are up 50% in the last 5 years, Microsoft’s up by 30%. Microsoft CEO justified the u-turn on his company’s climate pledges, stating they were “set before the AI moment”. The narrative has now shifted to ‘how will we find enough energy to fuel the dawning of the AI age?’ Here in Ireland, the new LNG terminal the government is seeking to build will end up being used to power this AI misadventure. All this at a time where the climate catastrophe is fast reaching its crescendo. Generative AI is the latest in a long long list of examples of capitalism’s incompatibility with our continued existence.
AI: A Climate Disaster
A cooler and more stable climate, paired with EU access and a tax-haven regime for multinationals has helped catapult the south of Ireland into the data centre hub of Europe. Approximately 10% of European data centres are in Ireland, a country which makes up 0.7% of the European population. Dublin in particular has the most hyperscale data centres of any city outside of the United States or China.
In 2015, data centres accounted for less than 5% of Ireland’s electricity usage – that figure is now close to 30%. A 1 MegaWatt (MW) data centre — that uses enough electricity to power 1,000 houses — can gobble up to 25.5 million litres of water annually. Ireland has 1400 MWs of data centres as of 2024. And unlike the water used in households, water used in data centres never returns to circulation and is lost to the atmosphere.
These figures will only increase in the coming years if there is no popular opposition to data centres. The need to keep these sprawling computing centres cool means that the data centre industry is now sniffing around the north of Ireland now, despite the lack of an EU connection.
What Can Be Done?
When People Before Profit passed a motion on South Dublin County Council to put a moratorium on new data centres it was immediately overturned by the Irish Government. There is the possibility to mobilise against data centres, as a recent report from The Sunday Independent shows that household energy bills have been subsidising data centres consumption, at a time when many families are left with the choice “heat or eat” during the winter. These same data centres receive billions in tax breaks from the same government that refuses to adequately fund services. The strain they place on our grid has seen a rapid increase in the number of power cuts households face each year. Between 2023 and 2024 there was a 22% increase in the number of power outages in Ireland with a disproportionate amount in south Dublin.
As ecosocialists, there is a need to link these immediate harms they cause with the broader climate disaster they are a part of. Demanding not only that we place a moratorium on the construction of new data centres, but that the existing ones be brought into public ownership and used for the common good. There are real useful reasons to operate some data centres; they are a utility of the modern age like water or electricity, but they cannot be left to Big Tech to operate for profit. Instead they should be democratically controlled and operated. This will inevitably involve switching the majority of them off altogether. In what rationally-organised society, in the middle of the climate crisis, would nearly 30% of energy go to plagiarism bots and invasive, personalised ad-machines?
