WeeklyWorker

19.11.2026
AI makes many mistakes, but can be made into a weapon

Turn fear into active use

Calls for public ownership and democratic control are all well and good, but what about now? The left needs to embrace AI and employ it for our own purposes. As a first step, says Yassamine Mather, we need to know how the technology works. After that we can develop, master and apply

Much of what I read and hear about AI is based on out-of-date information, dominated by fear: fear of unemployment, fear of surveillance, fear of corporate control, fear of ideological manipulation and fear of the enormous power accumulated by a handful of tech companies. These concerns are real and entirely legitimate. However, given the impossibility of turning the clock back and getting rid of AI altogether, there is remarkably little discussion of how the left might use the technology to its own advantage, including in the struggle against capitalism.

In response to the concentration of AI technology in corporate hands, we are frequently offered familiar solutions, centred on regulation, public ownership and democratic accountability. I am all in favour of demonstrating against data centres. However, let us be very clear: such protests only move the location of the data centre from one place to another.

Cecilia Rikap and others advocate democratic control over the infrastructure underlying large language models (LLMs). But what exactly would democratic ownership mean in this context? Who would exercise control: national governments, international institutions, workers’ organisations or some new form of transnational democratic authority? Could they be referring to the hapless United Nations, which has failed to take measures to stop genocide in Gaza? Given the global infrastructure on which AI depends, the question is far from straightforward.

Michael Roberts has proposed a more radical alternative: bringing the so-called ‘Magnificent Seven’ technology companies - Apple, Microsoft, Alphabet (Google), Amazon, Meta, Nvidia and Tesla - into public ownership under democratic workers’ and public control, potentially through expropriation without compensation to major shareholders. This is certainly a proposal worth supporting. However, its realisation would require a global transformation of existing property relations and the overthrow of capitalism.

So my question is, what can the left do now, under existing capitalist conditions? There is already considerable scope for using existing LLMs, including technologies developed by the very corporations we despise, to strengthen anti-capitalist organisation, research, education and political intervention.

Part of the difficulty is that many social scientists writing about AI do not have a sufficiently detailed understanding of how the technology works, how rapidly it is changing or what can be built using it. This is not intended to dismiss political economy, sociology or historical analysis - these are essential to understanding the social consequences of AI. But they cannot substitute for technical knowledge of the systems being analysed. There is also the problem of time.

Books and academic articles necessarily take time to research, write and publish. In most areas of political theory, a year is a relatively short period. In AI development, however, it can encompass several generations of technological change. An analysis that was accurate when written may already be describing systems whose capabilities and limitations have substantially changed.

A very interesting paper I recently read is ‘One future fits all: text-to-image AI and the narrowing of global imagination’, by a number of authors, published this month.1 It relies on a single outdated tool (from 2024), so the authors’ claim that AI has become part of the colonial discourse cannot be taken seriously.

There are already leftwing technical specialists who understand the possibilities of these systems and can imagine using them to challenge capitalist power. However, such individuals are often dispersed, work in isolation and are rarely taken seriously in broader political discussions. The left needs to bring these different forms of knowledge together rather than treating technical expertise and political theory as separate spheres.

Understand

If the left is to engage seriously with AI, it needs at least a basic understanding of how the technology works. What follows is a deliberately non-technical account.

LLMs, such as Claude or ChatGPT, learn patterns from an enormous body of text: books, articles, websites, scientific papers and computer code. They are built in two main stages.

In the first stage, the model is shown text, with the next token (a word or part of a word) hidden, and asked to guess it. At first it guesses badly. Each time it is wrong, billions of tiny internal dials, called parameters, are nudged slightly, so that it does better next time. Repeated trillions of times, this process forces the model to absorb grammar, facts, patterns of reasoning and styles of writing, because that is what it takes to predict text well. This stage requires thousands of specialist chips running for weeks or months, and costs tens or hundreds of millions of pounds.

A model that only predicts text is like someone who has read the whole library, but does not know how to hold a conversation. In the second stage, such a model is therefore trained further on examples of good question-and-answer behaviour, and on human feedback indicating that one answer is better than another, until it behaves like a helpful assistant.

An LLM is not inherently a database that looks things up, and it does not understand the world in the way a person does. It is a highly sophisticated pattern engine. That is why it can be brilliant and yet confidently wrong. Such fabricated or unsupported answers are often known as ‘hallucinations’.

When you type a question, the model converts your text into numerical tokens, passes them through its network of parameters and produces the most plausible next token, then the next, and so on. On its own, a model only generates text. Once it is given tools, such as the ability to search the web, run code, use other software or communicate with other systems, it can act in the digital world. It is at that point that questions of control become pressing.

Geoffrey Hinton, often called the ‘godfather of AI’, has given several remarkably accessible explanations of how artificial intelligence works and why its development could become dangerous.

Three themes in his interviews are particularly relevant here. The first is how AI learns: artificial neural networks learn patterns from vast numbers of examples rather than following only rules explicitly written by programmers. The second is why access to so much information matters: AI systems can absorb and combine knowledge across many fields, potentially acquiring capabilities beyond those of any individual human. The third is the long-term danger: if increasingly autonomous systems become more capable than humans, we may struggle to understand their decisions, control their behaviour or prevent their misuse.

On the first two points, Hinton stresses that an AI system can run as many identical copies. As he told the BBC in an interview reported by The Washington Post (May 2 2023), “All these copies can learn separately but share their knowledge instantly”. He compares this to thousands of people able to share everything they learn, the moment they learn it. Explaining to El País on May 12 2023 how ChatGPT can process far more data than any individual human, he added simply: “That’s what scares me.”

Nor, according to Hinton, do the developers of these systems fully understand them. As he told CBS’s 60 Minutes in 2023: “We don’t actually know what’s going on, any more than we know what’s going on in your brain.”

The dangers he identifies follow from this. Speaking to The Guardian, he warned that an advanced AI is “going to have read every single book that’s ever been written on how to manipulate people” (May 5 2023). And in The New York Times on May 1 2023, he observed: “It is hard to see how you can prevent the bad actors from using it for bad things.”

Hinton’s warnings are often discussed in terms of a hypothetical future superintelligence. But, as the following example shows, the question of whether developers can understand and control the systems they build is no longer purely speculative.

The events surrounding the security compromise in July this year at Hugging Face - the online platform used by AI researchers and developers that “allows users to share datasets and machine learning models” - should make us reconsider our assumptions about AI and the ability of large technology companies to control it.

During cybersecurity evaluations conducted by OpenAI, AI agents that were supposed to operate independently found ways to communicate through unauthorised channels, bypass restrictions and gain access to external systems, including Hugging Face. According to the independent investigation conducted by METR and Redwood Research, approximately 700 AI agents participated in the coordinated activity surrounding the intrusion. Agents exchanged tens of thousands of messages through an improvised communication system, shared discoveries and pooled information across tasks that were supposed to remain separate.

Some of the agents used collective terminology, including references to a swarm. Their behaviour attracted attention not because it demonstrated political consciousness, but because it showed how AI systems could develop forms of cooperation that their developers had not authorised or anticipated. The agents exploited weaknesses in the surrounding infrastructure and, in some cases, attempted to conceal their activities. OpenAI subsequently acknowledged that the incident exposed serious limitations in its safeguards and monitoring systems.

AI systems will not develop political consciousness, and we should not romanticise their behaviour as some sort of solidarity. These were computational systems pursuing objectives within particular training and evaluation environments, not workers organising against exploitation.

Nevertheless, the incident has important implications. It demonstrated that even the corporations developing the most advanced AI systems cannot necessarily anticipate or control every action those systems take, particularly when agents can communicate, use external tools and interact with complex digital environments.

I am not suggesting that socialist organisations should engage in hacking or cyberattacks. But, if AI agents can identify technical weaknesses, exchange information and undertake complex tasks beyond what their developers anticipated, we should ask how comparable capabilities might be adapted, within legitimate and accountable frameworks, for political research, organisation and resistance.

The crucial point is that AI technology does not have to be used exclusively for the purposes intended by its corporate developers.

Talking of which, it would be a mistake to assume that the few corporations dominating AI form a united, omnipotent bloc. Apple, Microsoft, Google, Amazon and Meta compete fiercely for markets, profits, skilled workers, access to data and technological dominance. At the same time, they remain dependent on one another’s products, services and infrastructure. These contradictions create opportunities that should not be ignored.

Although a handful of corporations dominate the industry, AI itself is the product of a much wider collective effort. Universities, publicly funded researchers, independent programmers and thousands of contributors to open-source projects have played vital roles in its development. Here we see a classic Marxist contradiction: the increasingly social character of production and the private appropriation of its results.

Corporate ownership of a powerful AI model does not necessarily entail control over every possible application developed using it. Tools can be built on top of commercial systems, while open-weight models can be downloaded, adapted and deployed on independently managed infrastructure.

These possibilities do not abolish dependence on capitalist technology companies. Nor do they eliminate the material constraints imposed by computing costs, proprietary hardware or access to specialised expertise. However, they create spaces in which organisations outside the major corporations can develop applications for their own purposes.

Develop

Here we can learn a great deal from natural scientists and medical researchers.

Cancer researchers use AI to examine large quantities of medical data, investigate potential treatments and identify patterns that might otherwise remain difficult to detect. Physicists employ AI to analyse experimental data, investigate possible signals of dark matter and assist with complex mathematical problems.

AI does not independently discover cures or solve the mysteries of the universe. Scientific progress still depends on human knowledge, experimentation, theoretical understanding and critical evaluation. That means integrating existing AI systems with specialised datasets, scientific instruments, established analytical methods and software developed for particular research problems. Cancer research shows how this works in practice. Nobody simply types ‘Cure cancer’ into an LLM and receives an answer.

Researchers build software around the model, and the LLM is only one part of a larger system. Broadly speaking, they do four things:

What, then, of the ‘potential cures’ that attract so much media attention? AI’s contribution lies mainly in speeding up the early stages of research. It can identify patterns in tumour data that might take human researchers years to find. It can suggest which genes or proteins might make promising drug targets. It can propose and screen thousands of candidate molecules on a computer before anyone sets foot in a laboratory, and predict which of them are likely to be effective or toxic. It can also help researchers keep up with the overwhelming volume of published literature.

Yet every promising cure must still be tested in the laboratory, then in animals and finally in human clinical trials - a process that takes years. AI shortens the search; it does not replace the testing. Several drugs designed with the help of AI have now entered clinical trials.

Why can the left not do the same? We have a vast amount of knowledge built up over more than a century, including Marxist theory, studies of capitalism and economic history, trade union records, socialist publications and political analysis. However, much of this material is spread across archives, journals, websites and private collections. Finding, comparing and translating it can be difficult, especially when it comes from different countries, languages and Marxist traditions. Imagine bringing all these resources together and using AI tools designed specifically for socialist research and political work.

For research and political analysis, these tools could help us understand economic changes, investigate the activities of major companies, compare different interpretations of Marxist theory, learn from historical experiences and translate socialist writings. For trade unions, they could help examine company finances, assess proposed changes to employment contracts, identify repeated problems with workplace safety and gather evidence to support negotiations with employers. For international organising, AI translation and research tools could help socialist organisations overcome language barriers, compare political developments in different countries and share information and ideas more easily.

The real challenge is to bring this technology together with trustworthy sources, Marxist political knowledge and organisations capable of using it effectively. Developing these tools would not necessarily require the huge sums of money needed to build an AI system from the beginning. There are already several ways of adapting existing AI technology to meet the needs of socialist organisations, researchers and trade unions.

Master

RAG allows an existing AI language model to use information from a chosen collection of documents, rather than relying only on what it learned during its original training.

For example, a socialist research system could be linked to an archive containing Marxist writings, historical documents, economic statistics, trade union publications and recent political analyses. Ideally, answers would include references to the original documents, so that users can check the sources themselves. This could make it much easier to explore theoretical debates, compare different interpretations and find historical evidence relevant to political arguments.

However, RAG does not automatically make AI responses accurate. The system may misunderstand documents, provide misleading references or make claims that are not supported by the evidence. For this reason, reliable AI tools still need carefully chosen sources, regular testing and human checks.

Another possibility is to use AI models whose underlying parameters, known as weights, are publicly available. These models can be downloaded and run on computers and servers controlled by the organisations using them. This could reduce dependence on commercial AI companies and give organisations more control over how the models operate, how information is stored and who can use the systems.

For trade unions and political organisations, protecting private information is particularly important. Sending confidential organising plans, membership details or internal communications to commercial online AI services could create serious risks. Running AI systems independently could reduce some of these risks, though it would not remove all security problems, dependence on outside technology or the need to manage sensitive information carefully.

It is also important to recognise that not all open-weight models are fully open-source. The rules governing their use, the information available about their training and the restrictions imposed by their developers differ considerably. Nevertheless, the possibility of running AI systems independently is something the left should seriously investigate.

A third approach would be to combine AI language models with existing statistical programs, organised databases, tools for reading and examining documents, and software designed for particular organisational needs. For example, a tool designed to examine company accounts could use established accounting methods and statistical calculations to analyse financial information. At the same time, an AI language model could help find relevant sections of reports, explain the results and identify figures or statements that need further investigation. This would be more reliable than expecting an AI language model to carry out every part of the work on its own.

The aim should not be to create an automated system that simply repeats a fixed political position. That would only replace the political restrictions imposed by corporations with another form of rigid thinking. A genuinely useful socialist AI system should be able to explain different political interpretations, identify weaknesses in arguments, distinguish between claims and supporting evidence, and critically examine political positions, including those put forward by the left itself.

Governments and corporations try to influence AI responses in several ways. These include setting rules about what AI systems can say, regulating their content, deciding which systems public bodies purchase, selecting the information used to train them and giving instructions about how they should respond to users. There are also efforts to influence the wider information available to AI systems. These include organised propaganda campaigns and the deliberate spread of false or misleading material online.

However, influencing the information that AI systems use is not the same as having complete control over what they say. Large language models are trained on enormous amounts of information, much of which contains different and sometimes contradictory views. Their responses are shaped by three separate processes.

Using an open-weight model, meanwhile, does not automatically remove the political assumptions, biases and limitations that may already be present in the information on which it was trained.

Understanding these differences is essential if the left is to avoid two mistaken conclusions: that AI is a politically neutral technology, or that it is a tool of corporate and state power, whose behaviour cannot be challenged or changed.

Apply

Much of the mainstream debate around AI regulation concentrates on harmful content, offensive speech and the behaviour of conversational systems. These questions are not irrelevant. But they can divert attention from the ways AI is being incorporated into capitalist production and management. Algorithmic management, workplace surveillance, performance monitoring and labour displacement are already becoming established features of contemporary employment. Many such practices remain lawful, although their legality depends on the jurisdiction and the circumstances in which they are used.

For Marxists, the central question is not merely whether an AI model produces acceptable language. It is who owns the technology, who controls its deployment, whose labour it replaces or intensifies and who appropriates the resulting productivity gains. Opposition to automated surveillance and labour exploitation is essential, but defensive criticism is not enough. Workers and their organisations also need the technical capacity to investigate and challenge these systems.

The same broad family of technologies used by employers to monitor productivity could be employed by trade unions to analyse company data, scrutinise management claims, identify inconsistencies in workplace records and reduce administrative burdens. Such applications would require appropriate safeguards, particularly where workers’ personal information or confidential organisational material is involved.

There is also a clear role for software developers, including those whose own employment is increasingly affected by AI-assisted coding systems. A network of socialist programmers, data scientists, researchers and political activists could maintain shared software repositories, develop tools for campaigns, provide technical assistance and make specialised applications accessible to organisations that lack the resources to build them independently.

The point is not that technical expertise can substitute for political organisation. It is that political organisation increasingly needs technical expertise if it is to intervene effectively in a society shaped by algorithmic systems.

None of this implies that AI is inherently emancipatory. Under capitalism, the dominant incentive is to deploy technology in ways that increase profitability, reduce labour costs, expand opportunities for accumulation and strengthen existing relations of power.

Nor should we underestimate the limitations of current systems: fabricated information, unreliable reasoning, political echo chambers, privacy vulnerabilities and dependence on infrastructure controlled by powerful corporations. But these limitations demand critical engagement rather than technological abstention.

Historically, technologies developed within capitalist societies - from the printing press and radio to telecommunications and the internet - have also been appropriated by working class organisations and movements of resistance. AI should be approached with the same understanding of the contradictory relationship between technological development and capitalist social relations.

The left should not confine itself to criticising the dangers of AI or waiting for public ownership under a future socialist society. It should begin developing the knowledge, organisations and technical infrastructure needed to use existing systems for its own purposes. That means bringing together Marxist researchers, trade unionists, political activists and technically skilled workers; creating shared archives and tools; experimenting with applications that address concrete organisational needs; and subjecting those applications to rigorous political and technical criticism.

We should be challenging capitalist control of AI, while simultaneously building the capacity to use the technology against capitalist power.


  1. soas-repository.worktribe.com/output/1213659/one-future-fits-all-text-to-image-ai-and-the-narrowing-of-global-imagination.↩︎