Introduction
This essay argues that what is commonly framed as an "AI crisis" is more accurately understood as a plain old crisis of capitalism, made visible through new technological means. The AI industry does not represent a break from capitalist logic but rather its projection into the informational age. The structures of extraction, accumulation, and alienation that characterized early capitalism and colonialism reappear here in digital form where information and computation are the new resources and valorization principles..
The discussion of AI regulation has focused primarily on dystopian scenarios: existential risk, autonomous weapons, mass surveillance. It is not my place to comment these scenarios (maybe another article coming), but they obscure a rather more immediate problem. The AI industry operates according to hyperkapitalist principles of production and accumulation that warrant scrutiny independent of any possible vision of the future. The patterns are familiar: primitive accumulation through the enclosure of commons, colonial extraction of labor from the Global South, and the systematic alienation of workers at every level of the valorization chain.
This analysis traces these continuities. It examines how the current organization of AI development reproduces historical forms of capitalist expansion, and how the experience of software developers confronting AI tools reveals the underlying structure of alienated labor that was always present but previously masked in knowledgework. The crisis is not technological. It is structural.
I. The Observation
Software developers report, with increasing frequency, a loss of satisfaction in their work. The cause is not displacement, since most retain their positions, but a transformation in the character of the work itself. Problem-solving, the cognitive core of software development, is increasingly delegated to AI systems. What remains for the human worker is supervision: monitoring outputs, reviewing code written by another entity, correcting errors in logic one did not produce.
The empirical data support this observation. According to Stack Overflow's 2025 Developer Survey, 84% of developers now use AI tools or plan to do so, an increase from 76% the previous year. According to Sonar's State of Code Report 2025, 42% of all code is AI-generated or AI-assisted. Trust, however, is declining: only 29% of developers report trusting AI outputs, down from 40% the year before. 46% actively distrust the results. The most frequently cited frustration, reported by 66% of developers, concerns "AI solutions that are almost right, but not quite." 45% report that debugging AI-generated code takes longer than debugging code they wrote themselves.
Positive sentiment toward AI tools has declined from over 70% in 2023 and 2024 to 60% in 2025. Of developers currently employed, 75% describe themselves as "complacent" or "not happy at work." These findings are consistent with qualitative accounts. In an interview with MIT Technology Review, software engineer Luciano Nooijen of Companion Group describes beginning a side project without AI tools and struggling with tasks that previously came naturally: "I was feeling so stupid because things that used to be instinct became manual, sometimes even cumbersome." He compares the phenomenon to athletic skill atrophy, suggesting that maintaining coding ability requires regular practice of foundational tasks.
But the issue extends beyond skill maintenance. "I got into software engineering because I like working with computers. I like making machines do things that I want," Nooijen states. "It's just not fun sitting there with my work being done for me." The loss is not merely technical but motivational: the work no longer provides what it once did.
II. The Parallel
Remember the precedence case: Before large language models transformed software development, a similar disruption occurred in the visual arts. The release of DALL-E, Midjourney, and Stable Diffusion in 2022 enabled image generation through text prompts, making certain forms of visual production accessible without traditional artistic training.
Many artists and designers reported diminished engagement with digital art. The concern was not primarily economic, though economic displacement was real, but related to the nature of creative work itself. When an algorithm produces comparable results in seconds, the question arises: what is the significance of years spent developing technique? What does skill mean when skill can be circumvented? Research on "AI Anxiety" among creative professionals documents this phenomenon: the experience of lost artistic identity, diminished sense of authenticity, and uncertainty about originality (Dasari, 2025). The cognitive and emotional reward associated with solving creative problems, what might be called the "lightbulb moment," depends on engagement with resistant material. When prompts replace that engagement, the reward structure changes. The sense of self-efficacy, autonomy and pride a seasoned artist gains when mastering a skill seemingly diminished. Production continues, but something else has ceased and as such art is endangered to lose its soul born from humane creative struggle with the material.
The parallel to software developers is structural. In both cases, the concern is not primarily economic displacement but the loss of what made work meaningful, art and good engineering are based on craftsmanship that should be carried with pride. So the question is not only what if humans will be displaced, but what happens to work when its intrinsically rewarding elements are removed.
## III. What Is Revealed
These disruptions reveal a structure that was always present but not in plain sight.
Software development has been characterized as paradigmatic knowledge work: intellectually demanding, creative, well-compensated, marked by high autonomy. This characterization, while not false, obscures continuities with other forms of labor. Software developers solve problems defined by others, build products owned by others, and work according to specifications set by others. The distinction from manual labor is one of degree rather than kind. What differentiated knowledge work was that the process itself provided intrinsic satisfaction beyond the wage.
Marx analyzed this structure in the Economic and Philosophic Manuscripts of 1844 under the concept of "alienated labor" (entfremdete Arbeit). He identified four dimensions:
- 1.
Alienation from the product: The result of labor belongs not to the worker but to the capitalist. It confronts the worker "as something alien, as a power independent of the producer."
- 2.
Alienation from the activity: Labor itself is experienced not as fulfillment but as compulsion. "The worker therefore only feels himself outside his work, and in his work feels outside himself."
- 3.
Alienation from species-being: Humans are distinguished from animals by free, conscious activity. Under capitalist conditions, this capacity is reduced to a means of physical survival.
- 4.
Alienation from other humans: Social relations become mediated by market relations; competition displaces cooperation.
In manual labor, these forms of alienation are manifest. In knowledge work, they have been masked. Problem-solving provided intrinsic satisfaction, a sense of competence and mastery, that made the underlying alienation tolerable. The product still belonged to the employer, priorities were still externally determined, but the process itself was engaging.
AI tools remove this mask. When problem-solving is automated, what remains is the bare structure of wage labor: time exchanged for money, supervision rather than creation, external control rather than autonomy. Software developers confront what factory workers have always known: their position in a system that treats them as replaceable inputs.
The transformation is instructive. Workers who once engaged in cognitive problem-solving now operate as monitors of probabilistic systems, steering outputs they did not produce and may not fully understand. The shift reveals something about the nature of employment as such: the provision of meaningful work was never the objective. Where creativity and skill represented costly necessities, their automation is pursued. Where they can be replaced by cheaper approximations, they will be. There is no dignity inherent in alienated labor, and no structural actor within capitalist production whose function is to protect those who do not own the means of production. This becomes visible in current patterns of workforce reduction. Layoffs in the technology sector are frequently executed not because firms face losses, but to redirect capital toward AI infrastructure. Workers are displaced not by necessity but by strategic choice, their labor costs converted into compute budgets. The logic is transparent once observed: human labor is an expense to be minimized; its reduction frees resources for further accumulation.
IV. The Layers of Appropriation
The AI industry exhibits not one but several structural features of capitalism simultaneously. It reproduces, at the level of information, the historical trajectory of capitalist expansion.
IV.I. Primitive Accumulation: The Enclosure of the Internet
Marx described "primitive accumulation" (*ursprüngliche Akkumulation*) as the historical process through which common resources were converted into private property: the enclosure of commons, the dispossession of peasants, the creation of a propertyless class compelled to sell its labor power. This process was not a singular historical event but recurs in each phase of capitalist expansion.
The internet functioned, in its early form, as a digital commons: an open network in which knowledge circulated without proprietary ownership. Large AI companies have appropriated this commons. They have scraped the web, extracting billions of texts, images, and conversations posted without compensation, and constructed proprietary models from this material. The products of collective intellectual labor have been enclosed, patented, and monetized.
This constitutes enclosure at the informational level. The mechanism differs from historical land enclosure, proceeding through infrastructure, terms of service, and first-mover advantage rather than violence, but the structural result is analogous: common resources become private assets.
IV.II The Colonial Dimension: Data Annotation as Extraction
This appropriated data is raw. To make it usable, it must be annotated: sorted, labeled, rated, filtered. This work is predominantly outsourced to Kenya, the Philippines, Venezuela, and India.
Workers there earn 90 cents to 2 dollars per hour. They work without contracts, without social security, without grievance mechanisms. Some are exposed to disturbing content that must be filtered from training data, including violence, child abuse, and hate speech. They often do not even know which company they work for or which product they are training.
The structure mirrors colonial resource extraction. Just as raw materials were extracted from colonized territories, processed in metropolitan centers, and sold back as finished goods, cognitive labor is extracted from the Global South, processed through statistical algorithms, and sold in the Global North as "artificial intelligence." The term itself obscures the human foundation. What is marketed as artificial is, in significant part, African, Filipino, Indian, Venezuelan. The intelligence is human; only the aggregation is mechanical.
This observation has technical implications. Large language models do not generalize well beyond their training distribution. Each new use case, each new domain, each new capability requires additional annotation: millions of hours of human labor to label, rate, and correct outputs until the model can approximate the desired behavior. The system does not learn in any robust sense; it memorizes patterns from annotated examples and retrieves statistically similar responses. When inputs deviate from the training distribution, the system fails, often silently, through hallucination. This means that every apparent capability rests on prior human labor. The model does not reason; it pattern-matches against the accumulated judgments of thousands of annotators. The term "artificial intelligence" is therefore misleading. A more accurate description would be: a probabilistic retrieval system trained on extracted human cognition. The "intelligence" is not generated; it is transferred, compressed, and concealed.
The concealment is ideologically significant. By presenting the system as autonomous, as "thinking," the labor that constitutes it becomes invisible. The Kenyan worker who spent hours labeling images, the Filipino contractor who rated chatbot responses, the Venezuelan annotator who flagged toxic content: none of them appear in the final product. What appears is a seamless interface that seems to generate knowledge from nothing. Increasingly, this precarization reaches the center as well. Gig workers in Europe and the United States annotate data for minimum wage or less. The colonial logic of extraction without recognition and labor without rights is spreading inward.
IV.III Alienation at the Center: The Programmer as Overseer
At the other end of this chain stand software developers in the Global North. They are not economically precarious; their salaries are high and their working conditions comfortable. Yet they experience a distinct form of alienation.
Their work increasingly consists of monitoring AI-generated outputs: reviewing code they did not write and may not fully understand, identifying errors in logic they did not produce, approving or rejecting results. The cognitive substance of the work, the problem-solving and reasoning, has been delegated to a system that itself depends on the extracted labor described above.
What remains is monotonous and mechanical. The work requires sustained attention but does not engage creativity. It demands presence but not intellectual participation. The developer becomes, in effect, an overseer of a machine whose operation rests on the invisible labor of thousands of annotators. Simultaneously, these workers face pressure to increase output velocity. The logic of AI-assisted development privileges throughput over craft: more pull requests, more features, more deployments. Quality becomes secondary to quantity. What emerges might be called a regime of token maximization, in which the metric of success is volume of production rather than soundness of result. Developers are incentivized to generate rather than to think, to approve rather than to scrutinize, to ship rather than to refine. The predictable consequence is an accumulation of low-quality output, technical debt that compounds silently until systems become unmaintainable.
This produces a characteristic form of distress. The cause is not physical exhaustion but a dual absence: the absence of meaningful engagement, and the absence of pride in what is produced. Satisfaction in work depends on connection to what one creates. When that connection is severed, when one cannot identify one's contribution in the output, when one knows that what one produces is mediocre but is measured by its volume regardless, the psychological structure that makes work tolerable breaks down.
IV.IV. The Chain of Value Extraction
At each stage, human labor is extracted, transformed, and rendered invisible. The final product appears as autonomous technology, as though it had generated itself.
The significant point is that none of these mechanisms are specific to AI. Primitive accumulation, colonial extraction, ideological concealment, alienation: these are constitutive processes of capitalism as such. The AI industry does not introduce novel forms of exploitation; it reproduces established forms within a new technological domain. What appears as an "AI crisis" is more accurately understood as a crisis of capitalist production relations, rendered visible through their application to cognitive labor. This reframing carries implications for regulatory discourse. The dominant conversation focuses on hypothetical future risks: superintelligent systems, autonomous weapons, existential threats. These scenarios may warrant consideration, but they should not displace attention from present conditions. The exploitation of annotators in Nairobi, the alienation of developers in Berlin, the enclosure of collective knowledge by corporations in San Francisco: these are not speculative scenarios. They are occurring now, and they follow patterns that precede any particular technology.
The infrastructure demands of AI development extend these patterns further. The computational requirements for training large models necessitate data centers of unprecedented scale, consuming electricity and water in quantities that reshape local environments. These facilities are disproportionately sited in structurally weak regions: the American South, rural areas of South America, parts of Southeast Asia and Africa. The selection logic is familiar from earlier phases of industrial expansion: locations are chosen for low land costs, weak regulatory frameworks, and limited capacity for organized resistance.
AI companies lobby actively for deregulation of labor and environmental protections in these regions, presenting infrastructure investment as economic development. The result is a transfer of environmental burden from wealthy consumption centers to peripheral areas. Communities are displaced to make room for server farms; aquifers are drawn down for cooling systems; electrical grids are strained or monopolized. The pattern echoes extractive industries of the past. When RWE excavates lignite in Germany, entire villages are razed and landscapes permanently altered to fuel power plants. The data center represents a similar logic: the physical transformation of territory in service of computational demand, with costs borne locally and benefits accruing elsewhere.
The hunger for compute thus produces its own geography of displacement. What appears as weightless, virtual, immaterial, the cloud, rests on material infrastructure that occupies space, consumes resources, and displaces populations. The abstraction conceals the concreteness of its foundation.
IV.V. The Logic of Automation
A further structural element warrants attention: investment in machinery to reduce labor costs. Human labor represents the largest variable cost in most production processes, particularly in contexts where workers possess legal protections, collective organization, and bargaining power. Investment in automation follows a calculable logic: when machines become cheaper than human labor, substitution occurs. This dynamic applies to manufacturing, to retail checkout systems, and now to software development. The recent increase in automation investment in China, where labor costs were historically low but are now rising, illustrates the point: capital moves toward lower costs wherever they can be found. This logic is internally rational. The problem lies not in irrational decisions by individual actors, but in the rationality of the system itself, which treats cost minimization as the primary optimization target regardless of broader consequences.
IV.VI. The Role of Belief
A further complexity: the actual productivity of AI may matter less than perceived productivity. What drives investment and adoption is not demonstrated efficiency but the belief that efficiency will follow.
But this observation raises a prior question: whether productivity is the appropriate metric at all. Work, profit, and productivity have become ends in themselves, detached from any account of what they are for. To ask "for whom are we being productive?" has come to sound almost subversive. Yet, contemporary societies face conditions that make the pursuit of productivity as such appear increasingly disconnected from collective welfare: democratic institutions under strain from polarization and erosion of trust, environmental systems approaching or exceeding critical thresholds, information environments saturated with low-quality content. In this context, the productivity gains promised by AI do not self-evidently serve the common good. The question of who benefits admits of empirical investigation, and the evidence suggests that the gains accrue primarily to those who already possess capital and institutional power.
The AI industry exhibits pronounced tendencies toward monopolization. The computational infrastructure required for large-scale model training is concentrated in a small number of corporations. The resulting systems are deployed in ways that tend to weaken rather than strengthen collective capacity: automating jobs rather than reducing working hours, generating content that displaces human expression rather than augmenting it, concentrating surveillance capabilities rather than distributing oversight. The technology is being domesticated by the wealthy and powerful and, in significant respects, deployed against social cohesion and individual autonomy.
I want to hint at Alan Turing's analysis of machine intelligence, though in a different way than usually noted. When considering whether machines can think, Turing proposed a reformulation: not "Does the machine possess consciousness?" but "Can the machine convince a human observer that it thinks?" The Turing Test measures persuasiveness, not intelligence. What matters is the capacity to produce belief.
At the economic level, a similar logic applies. The operative question is not whether AI constitutes a genuine technological revolution, but whether sufficient actors believe it does. Collective belief has material effects: investment flows, organizational structures change, labor practices are reorganized. If enough executives believe that AI will reduce costs, they will invest in AI, regardless of whether the evidence supports this belief. The METR study finding that experienced developers work 19% slower with AI tools has not significantly altered adoption patterns. Social institutions generally rest on such collective beliefs. Money functions because participants believe it will be accepted. Financial markets respond to expectations rather than present conditions. Political power depends on perceived legitimacy. The AI industry operates within this same structure: adoption is driven by narrative.
The decision to automate, to precarize labor, to substitute machines for workers is therefore not a technical necessity determined by the capabilities of technology. It is a political decision, grounded in beliefs about productivity, progress, and the relative value of human labor. And beliefs, unlike physical laws, are contingent on our values and discursive positions.
V. The Wrong Question
The conventional response to these developments takes the form: "Will humans be replaced?" This framing is inadequate, not because the question lacks importance, but because it individualizes and technologizes what is a structural problem.
A more productive question is: Why is work organized the way it is? For whom? Toward what ends? If work consists in selling time and ability to those who profit from them, whether corporations, shareholders, or abstract capital structures, then the problem is not the technology that displaces workers. The problem is the structure that constitutes labor as a commodity.
Within this structure, all workers, regardless of qualification, function as inputs producing outputs. The relevant variables are efficiency, cost, and substitutability. AI does not introduce this logic; it renders it visible. The software developer experiences, perhaps for the first time, what the warehouse worker or the assembly-line operator has always known. Under capitalist organization, where work takes the form of wage labor in the service of private capital accumulation, any technology that increases productivity will tend to deepen alienation rather than alleviate it. This applies to the assembly line, to the computer, and to AI. The tool itself is indifferent; the structure in which it is deployed determines its effects.
VI. Some Better Question To Ask
There is an alternative framing. Rather than asking "How can I retain my position?" one might ask: "What work is worth doing?"
Marx's conception of non-alienated labor was not labor without effort but labor in which humans "find themselves in a world they have themselves created." Work as expression of human capacity rather than as means of survival. Work oriented toward community rather than profit. Work in which the product does not confront its producer as an alien power.
This conception may appear utopian, but it describes experiences that are not unfamiliar: projects pursued out of intrinsic motivation, assistance given without expectation of return, creative work undertaken for its own sake, engagement with causes one judges worthwhile. In such moments, work is not alienated, not because it is easy, but because its meaning exceeds its exchange value.
The current disruption, affecting software developers, artists, and all whose work is transformed by AI, might therefore constitute an opportunity. Not for defending existing arrangements, but for asking what forms of work are worth preserving, and what forms were never worth doing in the first place.
References
Berardi, F. (2017). *Futurability: The age of impotence and the horizon of possibility*. Verso.
Dasari, P. (2025). AI anxiety among creative professionals: Global stress and future outlook. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5384712
Gent, E. (2025, December 15). AI coding is now everywhere. But not everyone is convinced. *MIT Technology Review*. https://www.technologyreview.com/2025/12/15/1128352/rise-of-ai-coding-developers-2026/
JetBrains. (2025). The state of developer ecosystem 2025. https://www.jetbrains.com/lp/devecosystem-2025/
Marx, K. (1844). *Economic and philosophic manuscripts*. In Marx-Engels-Werke, Ergänzungsband I. Dietz Verlag (1968).
METR. (2025, July 10). Measuring the impact of early-2025 AI on experienced open-source developer productivity. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
Sonar. (2025). State of Code Developer Survey Report 2025. https://www.sonarsource.com/state-of-code-developer-survey-report.pdf
Stack Overflow. (2025). Developer Survey 2025. https://survey.stackoverflow.co/2025/
Stack Overflow. (2025, December 29). Developers remain willing but reluctant to use AI: The 2025 Developer Survey results are here. https://stackoverflow.blog/2025/12/29/developers-remain-willing-but-reluctant-to-use-ai/
Stanford Digital Economy Lab. (2025). Canaries in the coal mine? Entry-level employment and AI exposure. https://digitaleconomy.stanford.edu/