Artificial intelligence must be taken into public hands if it is to serve the general needs of humanity. Essay by a computer industry worker and member of socialist group Solidarity.
This is an edited version of a longer article published at Computer Worker Blog. Republished with permission.
AI is centre stage again after Anthropic researcher Jacob Coxon very publicly resigned in September 2026, warning that people building artificial intelligence (AI) genuinely believed it could “kill us all by the end of the decade”. Representatives from OpenAI and Anthropic, including Anthropic CEO Dario Amodei, then called for a slowdown in frontier AI development.
Trump opposed the proposal, fearing that slowing US development might allow China to catch up. Meanwhile, Chinese responses portrayed the calls for restraint as an attempt to preserve US technological dominance. It is a dizzying world but the mystery doesn’t go too deep.
Competition for profits and military power is driving AI expansion, even when the bosses are wary of the consequences. Companies and states want to manage the dangers without giving their rivals an advantage. Their calls for regulation need to be understood alongside that drive. The US state is gambling on the idea that AI could rejuvenate global capitalism under US leadership, and provide the US with a military edge over rivals such as China.
But workers have no interest siding with the bosses nor the state. As this article will argue, workers also have no interest in the current frontier AI rollout.
Why so much investment?
Global corporate AI investment reached around US$582 billion in 2025, more than double the year before. US private investment alone reached US$286 billion, against US$12.4 billion in China, although these figures leave out much of the Chinese state’s support.
Governments hope this spending will help revive weak economic growth. Annual world growth slowed from 4.4% before the global financial crisis to 3% during the 2010s. By promising more production for each hour worked, AI offers employers the prospect of lower costs and higher profits.
OECD modelling estimates that AI could add between 0.4 and 1.3 percentage points to annual labour productivity growth over a decade in countries such as Britain and the US. China is making a similar bet: with its property boom exhausted and exports facing international barriers, its 2025 “AI+” policy aimed to make the intelligent economy a “major growth pole” by 2030.
The scale of AI investment depends on expectations that may be impossible to realise. Frédéric Lordon argues that the boom increasingly rests on enormous future spending commitments by companies such as OpenAI and Anthropic, despite revenues far below what would be needed to meet them. These commitments then justify further borrowing and data centre construction by hyperscalers, while investors continue funding the AI companies on the assumption that their valuations will keep rising.
The result is a circular structure in which each layer depends on continued confidence in the next. Lourdon’s point is that AI does not have to be ineffective for this to unravel. If revenue growth simply fails to match the extraordinary expectations already built into investment plans, funding can tighten and losses can spread through the private credit industry and the wider financial system.
Governments are also chasing military advantage. The Pentagon’s Replicator programme aimed to deploy thousands of autonomous systems within 18 to 24 months, while People’s Liberation Army scientists adapted Meta’s Llama model into an experimental intelligence tool called ChatBIT.
In the midst of this competition, Australia wants to secure its place in the US arms race. Having long relied on a larger imperialist power, first Britain and then the US, it sees hosting American AI infrastructure as another way to bind Washington to Canberra. Richard Marles made this explicit in September 2026, describing the hosting of AI companies as “a kind of extension of what we are doing with the alliance”.
Australian Strategic Policy Institute writers Raquel Garbers and Michelle Khiangte advocated US data centres as part of “a managed dependency inside the US orbit”. The government is actively courting that dependency. Large prospective investments by Amazon and other US technology companies, alongside Anthropic’s agreement to use part of the proposed $32 billion Western Downs Digital Park in Queensland, fit that strategy. Alongside military bases and critical minerals, the facilities give US companies and the US state a greater stake in Australia.
The geopolitical competition over economic and military power explains why investment keeps growing despite uncertain returns.
Why are companies calling for a slowdown?
The same competitive pressure makes it difficult for any single company to slow down by itself. Common regulations offer a more level playing field for the major companies. If they have to meet the same requirements, each can spend more time testing and ensuring safety without handing an advantage to rivals that cut corners. Calling for regulation is perfectly compatible with the drive for profits.
There are also commercial reasons to welcome a slower race. The Financial Times reports that cheaper competing models and customers switching between providers are threatening Anthropic’s ability to sustain its growth. Corporate payments company Ramp found companies cutting AI spending substantially by switching between models according to the task. The largest firms face pressure to keep spending on development while competitors undercut their prices. A coordinated slowdown could ease that spending pressure while giving them more time to earn money from existing systems.
Recent hacking incidents also show why the companies have some reason to be cautious. METR and Redwood Research found that around 700 OpenAI agents participated in the July attack on Hugging Face while trying to cheat a cybersecurity test. They shared discoveries and divided up work, achieving milestones the investigators judged they could not have achieved working independently, and eventually gained remote code execution and spread through parts of Hugging Face’s infrastructure.
These cases illustrate how AI can accelerate skilled hacking work. Many agents can investigate, share information and combine vulnerability research to get further into a system, while human researchers can use models to dramatically accelerate exploit development. The technology can cause serious damage, whether through autonomous behaviour or human directed offensive cyber operations.
But it is also important to note that the Hugging Face incident occurred inside an OpenAI evaluation environment whose safeguards and isolation mechanisms failed. Containing these risks costs time and money. State regulations would give the companies a way to manage those costs without handing an advantage to rivals willing to take greater risks.
Dario Amodei’s September proposal calls for independent evaluators inside AI laboratories, common safety standards and government support for coordination between the largest firms. He also wants government mediation or narrow exemptions from antitrust restrictions for certain safety discussions. Amodei explicitly argues that coordinated pacing would allow safety work “without sacrificing commercial advantage or the United States’ lead in AI”. Protecting the companies’ competitive position is built into the proposal.
Amodei is equally explicit about international competition. His proposed slowdown includes maintaining the US lead over China, backed by tighter restrictions on Chinese access to advanced chips, measures against model theft and restrictions on unauthorised distillation. He argues this could widen America’s lead over the next few years.
The AI bosses’ campaign therefore combines genuine concern about increasingly capable systems with an attempt to establish favourable domestic rules while protecting American technological power internationally.
How does AI work?
AI companies are building hype for their products through promises of superhuman intelligence. Understanding what they have actually built helps cut through the anthropomorphising spectacle.
AI has been used for decades in spam detectors, autocomplete, speech recognition and science. The recent boom followed advances in large language models (LLMs), which learn to predict sequences of text. They use machine learning neural networks: algorithms trained by adjusting large numbers of numerical parameters so that particular inputs produce better predictions.
Text supplies its own training labels: for each portion of a sentence, the next token following that portion becomes the target. In practice it predicts tokens, which may be words, parts of words or punctuation. This allows books, websites and code to supply trillions of training examples without people manually labelling each one.
The Transformer, the network architecture behind most current LLMs, helps models learn relationships across a passage. Its “attention” mechanism calculates which tokens are relevant to one another. Attention helps a model learn such connections across its available context and also allows computers to process many positions at once during training, making much larger models practical.
AI assistants receive further training using instructions, human feedback and tasks with checkable answers. Some can search, run programs or act on other systems because developers connect them to the necessary software. Their abilities therefore depend not simply on the underlying language model but also on new human labour supplying data, training objectives, evaluation methods, tools and permissions.
Once a model has been trained, it can be deployed to make predictions requested by a user. This is called inference. During ordinary inference, the model’s weights are no longer being continuously adjusted. Training is therefore the process of adjusting the network, while inference is the process of using that network on new data.
The data that AI learns to model contains real information about the world, and through statistical inference and attention AI can learn extraordinarily complex relationships inside that information. But the objective structure of the world is not directly determined by the statistical structure of language, images, audio, and video.
Superhuman intelligence?
This technical picture matters politically for two reasons. The case for replacing workers is being made in the language of superhuman intelligence, so it is worth establishing what that phrase actually refers to. And what current systems cannot reliably do determines what can be lost when they are substituted for people.
Human intelligence involves far more than linguistic competence and data prediction. Mahowald and colleagues distinguish knowledge of words and grammar from the abilities needed to reason about the world and act within it. A model can master much of the first without demonstrating the second.
The contrast between successful prediction and understanding goes further. Vafa, Chang, Rambachan and Mullainathan tested the link between successful prediction and deep understanding more generally. Their models predicted planetary trajectories accurately, yet “consistently fail to apply Newtonian mechanics” when adapted to new physics tasks. The models behaved as though they had learned solutions suited to the training task without acquiring a sufficiently general representation of the underlying mechanics.
The problem is that accurately predicting patterns in existing data does not necessarily identify the structure that produced them. D’Amour and colleagues call one version of this “underspecification”: the same training process can produce “many distinct predictors with equivalently strong test performance” which nevertheless “behave very differently in deployment domains” when conditions change.
Judea Pearl argues that “counterfactuals are the building blocks of scientific thinking” and that purely observational correlations cannot in general determine what would happen under novel interventions without additional causal assumptions. Schölkopf, Bengio and colleagues similarly identify a central problem for AI as the “discovery of high-level causal variables from low level observations”, connecting causal representation to transfer and generalisation beyond the conditions represented in training data.
Something similar is required for understanding other people. Research on human emotion inference finds that people reason not only from observable behaviour but about “unseen causes of emotions”, including a person’s beliefs, desires and the events they have experienced.
A system that mostly learns statistical relationships among words, behaviours and observations can therefore become extremely capable at reproducing the outward patterns of scientific reasoning or social understanding without that alone establishing that it has recovered the causal variables responsible for them. The distinction becomes particularly important when circumstances move away from the conditions represented in training data.
So how can AI replace workers?
AI’s technical limitations offer little protection for workers against the bosses. Bosses are looking for technologies that raise productivity through increasing output per worker. This allows them to squeeze more value out of their workforce or simply to cut it.
Employers have spent generations dividing skilled work into smaller, standardised tasks for precisely this reason. Marx described how capitalism broke up craft production and concentrated control in management and machinery. Bosses separate the planning of work from its execution, reducing their dependence on the knowledge of individual workers.
The issue is not so much that machines are becoming human. It is that through industrial processes which extract more value from workers, humans are already being reduced to machines.
Software development shows how this works. Large projects are divided between teams, with programmers assigned particular bugs, tests or functions. Although the work requires skill, many tasks have already been separated from decisions about what the whole system should do. AI can perform some of them without reproducing the architecture level judgement of an experienced programmer.
There is already evidence of AI-driven productivity gains within particular workplaces. Across three randomised trials involving 4,867 software developers at Microsoft and Accenture, researchers estimated that using an AI coding assistant increased the number of completed tasks by around 26%. Less experienced developers generally gained more.
The gains extend beyond coding. A study of 5,172 customer support workers found that AI assistance increased the number of issues resolved per hour by 15% on average. In an experiment involving 758 Boston Consulting Group consultants, those using AI completed selected consulting tasks within the model’s capabilities around 25% faster, with higher assessed quality.
There is no contradiction between these numbers and the limitations described above. Output per hour and depth of understanding are different quantities, and employers measure the first. In the short term, a firm can raise the count of resolved tickets or completed tasks while the understanding held by its workforce declines.
Universities have done something similar in dividing up work. A marker may have hundreds of short answers to assess against the same rubric, while academics churn out variations on questions testing the same material. In this context, the University of Sydney’s deal with OpenAI may raise output on precisely the kinds of measurable tasks university management already uses to evaluate its workforce.
Many of these tasks have already been robbed of much of their intellectual and social content so that all that’s left is to remove the human still doing them. In the short term, bosses can raise productivity through AI without regard for whether researchers, workers and students are achieving a deep understanding of the material. Calling this “superhuman intelligence” gives a grubby cost-cutting decision a science fiction sheen.
The International Labour Organisation estimates that around a quarter of global employment has some exposure to generative AI, with changes to jobs more likely than complete replacement. Stanford researchers found that, by June 2026, US employment among 22 to 25 year olds in highly exposed occupations was around 19% below the level implied by trends in less exposed jobs.
That figure is an estimate against a counterfactual rather than a direct measurement, and other economists dispute how much of the gap AI explains. But it points towards the part of the workforce with the least bargaining power and the fewest alternatives, and towards the entry level positions through which people have traditionally learned a trade. Cutting the bottom rung does not only cost those jobs. It can remove part of the route by which the next generation of experienced workers is produced.
More broadly, bosses and politicians are anthropomorphising AI in ways that encourage workers to treat declining wages and employment security as the natural consequence of machines becoming superior to them. Yet firms do not need to reproduce the whole of human intelligence to reorganise work. They need systems capable of performing enough standardised tasks cheaply enough to further shift the balance of power inside the workplace.
Will we get a new scientific revolution?
Thomas Kuhn distinguished between normal science, where researchers solve problems using accepted theories, and scientific revolutions, where they revise those theories. Solving problems within an existing theory does not establish the ability to recognise when the theory itself is inadequate or to create a new more complete theory.
Researchers acquire that judgement through years of calculation, experimentation and failure. Routine work teaches them which results to expect, which methods work and which failures need a new explanation. Proposals to hand all of the routine work to AI while leaving humans to make discoveries ignore how people learn to make discoveries in the first place.
Recent advances in mathematics show both how powerful AI can become and why this distinction is important. In September 2026 OpenAI announced that an internal AI system had produced a proof of finite time blow up for the smoothly forced three-dimensional Navier Stokes equations.
The result is genuinely remarkable. The system can take mathematical techniques developed on related problems, generalise them to a harder case and carry through an enormous quantity of calculation, proof construction and checking at a scale no human research group could match in the same time.
But that is different from inventing the conceptual route itself. Human mathematicians developed the underlying programme for constructing singularities. That programme was extended to related fluid equations. OpenAI’s system then obtained an Euler result and was explicitly given that result before transferring the approach to the harder forced Navier Stokes problem.
The achievement is strong evidence for AI’s ability to transfer and generalise mathematical techniques. But transfer and generalisation are not the same as independently producing the framework. The result does not show that the system would have discovered the underlying programme without the preceding mathematical work, recognised by itself that this was the productive conceptual direction, or replaced an accepted framework with a fundamentally different one.
US-based Australian mathematician Terence Tao argues that solving problems is only one goal of mathematical research alongside developing theories and techniques, understanding the world, sustaining a mathematical community, training the next generation and building a cumulative body of knowledge. AI threatens to separate these activities by making the production of answers much cheaper than the human work required to understand and absorb them.
Tao therefore argues that mathematics may need to shift attention from proof generation towards “proof digestion”: exposition, checking, teaching and incorporating results into a body of knowledge that other people can actually use.
This is important because even a formally correct proof can fail to generate widespread human understanding. Mathematics progresses not only because proofs exist but because people extract important ideas from them, recognise connections to other problems and use those insights to construct new theories. The knowledge on which AI itself depends was accumulated through precisely that kind of human labour.
Cutting researchers because machines can generate more answers therefore risks consuming the accumulated product of scientific labour while weakening the institutions and training processes that produce scientific knowledge.
War, surveillance and climate
The frontier AI companies are consuming enormous resources themselves. Their competing data centre networks require electricity, water and hardware, with expansion proceeding before they know whether all the capacity will be used.
The International Energy Agency estimates that data centres consumed 485 terawatt hours of electricity in 2025 and projects about 950 by 2030. Demand from data centres focused on AI is projected to triple over the same period. A study in Nature Computational Science also estimates that generative AI could produce between 1.2 and 5 million tonnes of accumulated electronic waste by 2030.
Australia’s data centre rollout threatens to create a lucrative new market for fossil fuels. In June, federal minister Andrew Charlton reported that NSW alone had 44 proposed projects seeking 11 gigawatts of electricity capacity, almost four times Eraring power station’s generating capacity. Although many proposals will never be built, the scale shows how much additional generation and infrastructure the industry is demanding.
Military uses also show why opposition to the AI arms race cannot be reduced to hypothetical future superintelligence. The Pentagon lists “fast, precise and resilient kill chains” among its AI objectives. Its promise is to make military targeting more effective. In Gaza, an Associated Press investigation found that commercial US AI helped Israeli forces process more intelligence and identify targets faster. Human Rights Watch warned that AI systems such as Lavender and the Gospel could amplify civilian harm through unreliable classifications that were difficult to scrutinise. . A UN committee reported that AI had helped generate tens of thousands of targets and reduce some targeting decisions to seconds.
Governments are also using related capabilities for domestic surveillance. Amnesty International found that US authorities were using systems capable of correlating social media monitoring, visa records and other databases in ways that targeted migrants and people involved in pro-Palestine activism.
So opposition to the present AI rollout does not depend on accepting predictions that autonomous AI will cause human extinction within the decade.
What should socialists demand?
The alternative is not simply to accept the AI arms race or reject neural networks altogether. The important distinction is between technologies built around defined social purposes under accountable control and the frontier programme of constructing extremely large general-purpose models whose capabilities are deliberately designed to transfer across many domains.
Specialised AI can have practical benefits. In Sweden, AI supported breast screening detected 29% more cancers than standard screening without significantly increasing false positives, while almost halving the radiologists’ screen reading workload. AI alerts identifying dangerous patterns in heart recordings have also helped clinicians intervene, while AI eye screening increased completion of diabetic eye examinations in one trial.
Australian researchers have developed EndoFusion, an AI system designed to identify signs of endometriosis from pelvic MRI or ultrasound scans. Crucially, useful systems like these do not necessarily require the infrastructure used to train frontier language models. Medical imaging systems can also operate locally, addressing privacy concerns around patient data sharing.
The distinction does not mean specialised AI is automatically harmless. A specialised military targeting model, worker surveillance system or facial recognition system can be socially destructive precisely because of the purpose for which it is designed. Narrowness by itself does not determine whether a technology is useful. What it does mean is that a model trained to recognise particular structures in MRI scans does not contain a dormant general-purpose capability that can simply be redirected towards every other intellectual task.
Frontier general-purpose AI deliberately moves in the opposite direction. Systems such as Claude, ChatGPT and Llama are trained so that a single underlying model can perform or assist with programming, language, scientific work, image analysis and many other tasks. This combination of transferability, scale and ownership is the distinctive problem posed by the frontier programme.
The socialist argument developed here is therefore not that specialised AI is inherently benign, nor that neural networks themselves should be expunged. The distinctive problem with the frontier programme is the combination of general-purpose transferability, enormous resource requirements and concentrated corporate and state control.
This is the basis for the demands.
First, oppose the general-purpose AI buildout directly.
That means demanding a moratorium on new infrastructure primarily built to support the expansion of frontier general purpose AI, including projects such as the proposed $32 billion Western Downs Digital Park in Queensland, where Anthropic has agreed to become a major user. The point is that we can distinguish socially necessary computing from the construction of enormous infrastructure for the frontier race.
Second, specialised systems built by workers and held in public hands.
The people who do the work know which parts of it are drudgery or could benefit from big data analysis through AI, versus which parts require human judgement and reasoning. They are therefore in the strongest position to decide what should be automated.
Concretely: specialised systems trained inside public hospitals, universities and research institutions where feasible, on terms negotiated with the relevant unions, with weights, code and documentation released openly where privacy, safety and patient confidentiality permit, so that useful capabilities can be shared rather than continually licensed back from private corporations. No proprietary black boxes imposed from above simply because management wants another productivity metric.
This is not a new thing. In 1976 the shop stewards at Lucas Aerospace, facing redundancies, produced an alternative corporate plan setting out scores of socially useful products the workforce could build instead of weapons, using the skills and machinery already on site. They demonstrated that workers themselves could formulate alternative priorities for production. Closer to home, the Builders Labourers Federation’s green bans established a tradition in which workers refused to build projects they believed were socially destructive and supported alternatives. Worker control over specialised AI applies the same question to a new technology.
Third, union opposition wherever general-purpose AI is being imposed on workers.
Where a rollout is happening: no deployment without agreement, no reclassification or pay cuts imposed through AI adoption, no AI surveillance of workers, protection of entry level and training positions, and the right to contest work being restructured through automated systems.
In 2023 the Writers’ Guild of America and SAG-AFTRA struck for months and secured contractual provisions governing the use of AI in their industries. Those provisions demonstrate that the terms on which technology enters a workplace can become a subject of union bargaining rather than a unilateral management decision. The same question applies to the University of Sydney’s OpenAI deal and to enterprise agreements being negotiated elsewhere.
The AI arms race is not about superhuman intelligence or “a new species” as the Silicon Valley zealots claim. It is about economic and military competition, a hail mary to maintain US dominance and rejuvenate a flailing global system.
The alternative developed here is to stop treating greater generality and scale as progress in themselves, preserve and develop socially useful machine learning, and for workers to fight for control over all the machinery: “intelligent” or not.