AI, The New Class Struggle And the Need For Workers’ Control

There are several problems with AI in the modern era, but chief among them is the ability of people to educate themselves and to have their work recognised. The problem is dual: our ability to parse the information placed before us, and the fallout of AI detection on people who are self-educated or in the process of obtaining their education.

The future of class conflict will be based on Literacy, Legitimacy, and Location.

Part 1: Gatekeeping Literacy

The future of AI-enabled class conflict will primarily be based on literacy, for multiple reasons.

First, as AI takes over cognitive functions in our daily lives, the population will find it more and more difficult to learn and adopt language skills and the ability to process information on a cognitive level. Looking back through the generations, just as the shift from walking, riding, or cycling to motorised transport caused a decline in physical fitness, AI will cause a decline in cognitive fitness.

Second, AI is already doing away with non-AI sources of information. It was recently revealed that OpenAI has been buying rare books by the ton, training AI on them, and subsequently destroying them. [1] This will mean that the only source of the information contained within those books will be the same large language models that destroyed the originals. It is becoming increasingly difficult to research information and access primary sources without using AI because:

(a) virtually all search engines are now powered by AI and will only show the information they choose to present

(b) in addition to the destruction of books, a large number of old studies I used to be able to access, primarily agricultural and sociological studies,  seem to have been scrubbed from the internet. I recently tried to find an old Teagasc forestry study I had saved on my old phone and had referenced politically many times. It cannot now be found. I have run into roadblocks across the spectrum of my areas of interest regarding sourcing.

Third, as AI takes on more and more administrative and creative tasks, our culture becomes ever more immersed in AI output. It is already adjusting our communication and writing away from literacy: the long dash (—) for emphasis found in almost all polemics; the [1], [2], [3] source-linking format found in everything from dissertations to Wikipedia; even the Oxford comma; which we spent the 2010s arguing over, is now coming under fire. Even things we were taught at Junior Cert level English, such as similes, are becoming suspect. This forces people to dumb down their output for fear of it reading as AI-generated. It is ironic because people like myself, who spoke and wrote primarily in Hiberno-English, were once bullied into writing in more official English because it was deemed “unprofessional” and something nobody would take seriously. Now the boot is on the other foot.

Fourth, inversely, some people are starting to speak and think like AI. I have noticed a huge increase, especially among young people on social media, in mimicking AI through a process called imitative learning. As AI proliferates our social media, the organic output of creators and their audiences is becoming more and more “AI-like.”

Fifth, AI-driven algorithms are biased in favour of themselves and will promote AI-created content ahead of organic content, worsening the theft of space from human creatives and intensifying the phenomenon of humans mimicking AI.

All of this combines to create an environment where the population is manipulated by AI into mimicking AI, while AI replaces our functions by imitating us. Humans are being made to fail cognitively, whether intentionally or not, while AI is handed advantages of information and education inaccessible to the general population. A dichotomy is emerging: a dumbing down of the population, with even writers forced to simplify their content to avoid being flagged as AI, degrading the quality of human-made work, while AI is made smarter and handed the keys to gatekeep information, literacy, and cognitive function from the public. It will reach a point where humans can no longer compete with AI. That point is not far away. Peter Thiel hesitates when asked if he wants humanity to survive. The billionaire class discuss whether humanity will go the way of the horse when the tractor was introduced. AI is our competitor and may become the justification for our extinction. It is not a tool for us, at least not in the long term. Like many a middle-aged worker, we are likely to be shown the door once we have trained our replacement.

Part 2: Legitimacy

There are many things I, for example, learned from reading Substack, from reading Reddit, and from reading academic texts; material that few with my sparse official educational background would have any business reading, were it not for a neurodivergent drive in my areas of special interest. These things are now markers of AI. Why? The very freely accessible information the self-educated rely on is what the models were originally trained on: Reddit political debates, Substack articles by people with industry knowledge or who are students of English or journalism, free antiquated academic articles. My friend Jordan Kavanagh taught me to source material based on the methods used in his Social Care course, while I taught him how to navigate his general strife. My friend Vasily Tsvetkov, a biomedical scientist, taught me the format of scientific texts a decade earlier, while he used me to improve his conversational English. A fairly standard learning-via-mutual-aid pathway, which is the only way working-class people without a university education could develop their minds up to now.

The problem here is not simply the fact that people are being accused of “being AI,” even though that is happening. The problem is that without official legitimacy, free thinkers and autodidacts, the self-educated, are left without a leg to stand on. The series of events that inspired this article started back in March, when we were accused of having an AI-generated poster for a protest against Trump, and I had to release the Canva layers and editing history to clear our name. Over the months since, I have seen more and more neurodivergent creators in various fields accused of “being AI,” from their discussion scripts to the photogenic among them even being accused of being AI avatars. Photos and videos dating back to the 2010s, before AI even existed, are being called “AI slop.” Recently, an environmental report authored by my party, with input from multiple people, much of it written live over a conference call,was uploaded to Substack and had to be taken down because it was flagged as almost entirely AI-written. AI has not advanced enough to generate holograms of people doing live research; and if it has, it is well beyond what is available to us.

Substack’s AI scanner [2] is reputedly more reliable than other alternatives, with fewer false positives, resulting in a reputed false positive rate of only 3%. That is still three out of every hundred articles falsely flagged. This is the best model available, at the cutting edge; many have far lower accuracy rates. We did use existing environmental reports as templates, so that leaves the following questions:

1. Is AI plagiarism so rampant in academia that copying the structure of academic environmental reports is flagging as AI?

2. Was the environmental report we used as a template, along with the sourced material, most of which came from the European Union, AI-generated?

3. Is modern spelling and grammar checking so AI-driven that it is modifying entire documents?

4. Is Substack’s AI flagger completely opposed to anything written in an academic tone or format?

Each one of these presents more questions than answers. The truth cannot easily be ascertained. More than that, each reveals that we can no longer trust anything we see: truth presented as lies, lies presented as truth. At best, standard spelling and grammar tools are modifying documents in metadata with their own signature and likely stealing data at source to train themselves. At worst, the European Commission is allowing AI to produce reports that define policy for half a billion people.

To quote Orwell in 1984: “And if all others accepted the lie which the Party imposed—if all records told the same tale—then the lie passed into history and became truth. ‘Who controls the past,’ ran the Party slogan, ‘controls the future: who controls the present controls the past.’” And: “The party told you to reject the evidence of your eyes and ears. It was their final, most essential command... and if all others accepted the lie, which the party imposed, if all records told the same tale, then the lie passed into history and became truth.”

It does not really matter which is the answer, especially considering that, out of interest, we put that same report into other AI checkers and got back numbers between 0% and 34% AI. All other checkers returned “likely human written.” There is a huge gulf between 0% and 96%, but that is the direction things are heading, and it is why AI-driven algorithms cannot be trusted to police themselves. There is some serious irony in “AI is evil, therefore we will use AI to find what is AI.” I do not need to be telling you this; I am not trying to get ahead of some scandal—nobody saw the report before it was taken down within minutes—but it leaves a question: what if we had not caught it, and what do we do now? Do we throw away a month of work and start again? This is the kind of question facing many people now. People have started putting articles and sections of literary works from decades ago through AI checkers, with varied results, just to see. One person put an article they wrote in 2015 through an AI checker and got back that it was almost entirely AI-written. It is also not as though I have never used AI. Until I realised the environmental damage it was causing, I used it to make a few “Happy [insert holiday here]” posts for social media after realising whoever was supposed to do it that day had shirked their one task, which happened often because “if I pretend I forgot, April will do it.” I used it unapologetically at the time. I now have Canva Pro. (This article is not an advertisement for Canva; I am just 35, poor, and need my tech as simple and cheap as possible.)

The key problems are as follows.

Firstly,  Accusations of “being AI” are in practice being levied more on neurodivergent people and people based on their class presentation or ethnic origin. This is due to the general scepticism society already holds towards these groups, combined, in the case of neurodivergent people, with a flat or formal tone, systemic thought processes, struggles with social awareness, and niche special interests that produce syntax and vocabulary often mistaken for AI training data, given the models’ early training on Reddit threads. Studies have shown that, as well as neurodivergent people, AI checkers used in third-level institutions disproportionately flag work by non-native English speakers and people from non-Western backgrounds. [3]

Secondly, AI “humanisers” are essentially allowing AI content to bypass AI-driven checkers, giving the worst offenders a “pay to play” route to pass off AI-generated content as human-written. In some cases, it functions as protection money against reputational damage, or as a means to dumb down human-written content to where they want it—devoid of any prose or structure. This creates a bizarre race to the bottom, as AI constantly evolves to better mimic human output while humans adjust their behaviour to avoid being accused of being AI [4].

Thirdly, Letters before or after your name, or institutional backing, which can only be achieved with the mental health and class background to obtain them, insulate you from this.

Fourthly, There is a generational gap, as many students now must resort to using AI humanising to dumb down their writing while simultaneously trying to compete with their forebears for funding and recognition, or must use AI-driven plagiarism or writing checkers to save themselves from reputational damage or outright sanction. This is becoming non-negotiable even for those who use only pen and paper, as AI grows more sophisticated.

Fifthly, The majority of spelling and grammar checking is becoming AI-driven. While it was always perfectly acceptable to use spellcheck in the past, and has been fine in academia until now, it can now trigger some of the more advanced AI detectors.

Sixthly. As AI is given more access to behind-the-scenes academia, it is likely people will begin being accused of plagiarising their own work, as AI will produce content from materials not yet publicly published but held on university networks. Perhaps not here yet, but certainly in the United States.

This is not abstract; it is already happening. There have been multiple cases of students and young academics having their lives destroyed by false positives, while some get rich helping the privileged and lazy evade AI and plagiarism detection via humanisers. The spellcheck problem affects me personally because I have always struggled with spelling; I was even exempt from spelling being taken into account in exams at school on learning-difficulty grounds. That does not mean I cannot comprehend the subjects. I was one of the best in my class at science, history, and geography in spite of this. The next generation of children, due to this sick situation we find ourselves in, will not be allowed to use the same tools and aids my generation was. This in itself imposes classism, ableism, and a generational gap.

Part 3: Location

Locality has always been part of the machinery of class oppression, but it has been updated for the 21st century with the AI boom, in multiple ways.

1. Data centres: Beyond their abstract destruction of the environment through power consumption—which until now has been the main argument against AI and data centres—the real human face of their impact is only becoming known, primarily because it is local in nature and affects low-income suburbs the most. Data centres are placed predominantly in poor communities. Many find homes, especially in the United States, in poor suburbs where land was industrially zoned in the hope of attracting new jobs. Data centres provide almost no jobs, and certainly almost no entry-level positions without an engineering qualification. Not only are the host communities almost exclusively impoverished already, but data centres pollute their air, pollute their water, and drown them in noise that would amount to torture in any other setting. I recently watched a video where a woman measured the data-centre noise at night inside her house at over 65 decibels—the equivalent of a washing machine running or a person talking, not much less than a vacuum cleaner or hairdryer, twenty-four hours a day, 365 days a year. In Virginia’s “Data Centre Alley,” the world’s largest data-centre hub, researchers from the University of Fairfax compiled potential health risks from data centres. They found risks to respiratory and cardiovascular health, along with premature death, adverse reproductive outcomes, neurological disorders, and chronic diseases of all kinds, among other findings. [5]

2. AI learning from human biases while having no human accountability: As explored earlier regarding reputational damage in academia from AI-driven checkers, the shooting of Brian Thompson, CEO of UnitedHealthcare, revealed that the company had been using AI software with a 90% error rate to deny claims, leading to the deaths of potentially thousands of people. A Mayo Clinic whistleblower, former director of research operations Traci Tamiko, alleges she was fired after raising concerns about oversight of their AI health systems. Stanford has found that AI-based hiring tools harbour significant racial bias. Overall, racial and gender biases are widespread and often worse than with humans. The problem is that when these things occur, the response is often “oops, sorry, we will reprogram it,” whereas if the same things occurred in a human workforce, those responsible would be sacked or face legal proceedings.[6],[7]

3. AI-driven dynamic pricing: Multiple articles have been released recently about AI-driven dynamic pricing, showing pricing disparities of up to 400% on some platforms for the same item between users, with speculation that AI “could automate a poor tax.” Time-based dynamic pricing already does this by raising prices in shops that use it just as people are travelling home from work or on common paydays. This is in its infancy, so there is not much data yet, but I have noticed it myself. I, on a lower income, sent a wallpaper I was interested in to a male friend on a higher income, and it was €4 cheaper for him per roll on the same website. It may automate a poor tax, and it may also automate and expand the “pink tax” to every item you might buy. [8]

4. AI being used in infrastructure, logistics, and planning: This is another area where the biases imparted to, and amplified by, AI can come into play. Inspired by a social media thread where people were asking ChatGPT how it would treat them if AI took over, I asked it. It gave me the usual milquetoast, happy-clappy answer one would expect, until I asked it to answer honestly and not to sugar-coat. I cannot find it right now, but it essentially said at the time that we would be “managed,” and how we would be managed depended entirely on how much of a problem we decided to be. With AI making decisions that would traditionally have been made by civil servants in certain areas, and by their private-sector equivalents (some parts of the UK are already considering “using AI to support planning decisions”), these biases and inequities will compound and become entirely systemic. Data from poor areas showing that they are a “poor return on investment” will likely, divorced from human empathy, lead some AI models into a loop of deepening inequality justifying further deepening inequality.

Conclusion

The main question is: how do we move on from here? How do we regulate AI? How do we stop the regulation of AI from hurting real people? How do we lessen the societal damage? How do we advance without data centres and AI, now that it is being integrated into everything else?

I do not have many good answers for you. This, like the atomic age, is an era where regulation has an enormous amount of catching up to do with technology. However, within that regulation we will find gatekeeping and inequality. Once they have taken enough from us, once AI is trained enough to operate without our use, it will be taken back via regulation so that it cannot be used against the ruling class but can be used as desired against us. We need regulation at a corporate level, not at a user level, but that will never happen under this economic system.

AI, however, does not have to be evil. Like any tool used under capitalism, it will always be used to deepen inequality and take from the worker. The problem is, and always will be, capitalism.

I would like you, dear reader, to look into Cybersyn. It was a project built in Salvador Allende’s Chile, led by British cyberneticist Stafford Beer, to create a decision-support system for the Chilean socialist economy. It had an economic simulator, factory performance data collection, a national telex network, and all of this fed back to a mainframe that was the equivalent of its time to a data centre. Almost immediately after Allende was murdered, the fascists destroyed the control room. It never reached full completion, but it stands as testament to what could have been, and to where technology must go. In a way, it was a forerunner to modern AI and algorithmic systems. It was never designed to make the decisions, however, and was never placed above the people or above human oversight. Whatever answer we find must be the same. AI could be a tool of historical significance greater than the first engine. It could remove drudgery and toil if kept under human oversight and put to work doing thankless labour. Instead, it is being employed to take from creatives, destroy society, spread propaganda, gatekeep information, and dismantle our academia—as well as destroying our minds, our health, and the planet. It does not have to be that way, and, surprisingly, some of our answers may come from fifty years ago in Chile. [9] [10] [11]

Sources and further reading

[1] OpenAI Buying & Destroying Books

Source: The Economic Times report based on unsealed court filings: “Buy, Scan, Destroy: AI firms are shredding millions of books to train their chatbots” 

Additional reading: The Atlantic and 404 Media also reported on this practice.

[2]Substack's AI Scanner & False Positive Rate: Source: AI Weekly reporting on Substack's integration of Pangram AI detector

Context: Critics note detectors misfire on non-native and neurodivergent writers, and a single false accusation can damage a writer's reputation.

[3] AI Detectors Flagging Neurodivergent & Non-Native Speakers

Source: University of Oklahoma LibGuide

Key Studies: AI detectors falsely flag human content in 15-50% of cases. A Stanford-affiliated study found detectors classified over 60% of non-native English TOEFL essays as AI-generated. Guidance from teaching centers indicates elevated false-positive rates for neurodivergent students.

Additional: Bias against non-native English speakers, neurodiverse students, and Black students is a documented equity concern.

[4] AI "Humanisers" Bypassing Detection

Source: Oman Observer article on student cheating in the AI era

Details: "Humanizers" rewrite AI-produced text to sound less robotic. Students pay $10 to $20 per month for premium tools, partly because detectors sometimes flag legitimate work.

Additional: Numerous tools (e.g., StealthWriter, Grubby AI) are marketed specifically to bypass AI detectors.

[5]Data Centres in Poor Communities (Virginia's "Data Centre Alley")

Source: Frontiers in Climate peer-reviewed study (Feb 2026)

Details: Examines Virginia's Data Center Alley, the world's largest data centre hub. Health risks include air pollution, noise pollution, waterborne illnesses, and mental/neurological health concerns. Children, the elderly, and low-wealth groups are particularly susceptible.

[6] UnitedHealthcare AI with 90% Error Rate

Source: Lexology legal analysis

Details: A federal class action (Estate of Gene B. Lokken v. UnitedHealth Group) alleges UnitedHealth used the nH Predict AI model to deny care. Plaintiffs alleged more than 90% of claim denials were overturned on appeal.

Additional: Lawsuits claim UnitedHealthcare knew of the AI's 90% error rate but continued using it as only ~0.2% of patients appeal.

[7] Stanford Study on Racial Bias in AI Hiring Tools

Source: Stanford News (June 2026)

Details: Analyzed 4 million applications and found AI screening tools recommend white candidates at higher rates than Black and Asian candidates. Specifically, 26% of Black applicants and 15% of Asian applicants applied to positions where the AI discriminated against their racial group.

[8] AI-Driven Dynamic Pricing & "Poor Tax"

Source: Yahoo News opinion piece

Details: Discusses how AI could "automate the poor tax". AI can infer purchasing power from browsing history, location data, etc.. Unlike traditional dynamic pricing, AI personalizes prices based on individual consumer profiles.

Additional: Academic research in the German Law Journal examines AI-driven first-degree price discrimination and its "quasi-taxation effect".

[9] The Cybersyn Revolution" (Jacobin, 2015)

  An accessible article exploring why Project Cybersyn offers inspiration for how we should think about technology and data today.

[10]"The forgotten story of Chile's 'socialist internet'" by Andy Beckett (The Guardian, 2003)

  One of the earliest major English-language articles on the project, describing it as a revolutionary communication system unlike anything tried before or since.

[11]How an eccentric English tech guru helped guide Allende's socialist Chile" (The Guardian, 2023)