When fast-fashionistas scrolling through Shein’s website came across a shirt seemingly modeled by Luigi Mangione, the alleged shooter of UnitedHealthcare CEO Brian Thompson, confusion, outrage, and maybe a little bit of swooning ensued. Mangione’s low-resolution doppelgänger had all the hallmarks of generative artificial intelligence—odd textures, strange lighting, and unnatural fingers. Researchers at BBC Verify found a 99.9 percent similarity between the image and a courtroom photo of Mangione. The clothing company swiftly removed the image, while blaming a third-party vendor and promising tighter oversight. But since we lack control over the ways in which technology is reshaping our world, there’s nothing to guarantee that AI won’t continue to cause such problems.
Shein generated headlines because of the person whose image was being replicated, but the incident is symptomatic of the ongoing datafication of the uncoded world, which echoes past periods of expropriation. Data is produced through human activity and subsequent machine surveillance. These digital hoards become capital as companies build intangible assets, such as the profit-generating AI tools trained on vast quantities of images, text, and data from our digital footprints, often without consent. We lack control over our data and receive no monetary compensation for it.
Recently, however, there have been a growing number of high-profile legal challenges to AI companies, including one case between Anthropic, the maker of the chatbot Claude, and a coalition of authors and publishers. In a June 2025 summary judgment, Judge William Alsup of the U.S. District Court for the Northern District of California ruled that legally acquired books could be used to train AI under “fair use,” and that the “technology at issue was among the most transformative many of us will see in our lifetimes.” However, he also found that Anthropic had knowingly and illegally downloaded millions of pirated ebooks from sites like Library Genesis and Pirate Library Mirror. Instead of purchasing the texts, Anthropic chose to “steal” them, exposing itself to damages of up to $150,000 per pirated work.
Anthropic recently agreed to a $1.5 billion copyright settlement—the largest in U.S. history. For half a million authors, the $3,000 payout per book might look like a victory, but for Anthropic, it was simply the cost of doing business. While lawyers dubbed it the “Napster moment” of the AI industry, they seem unaware of the difference between Napster (a peer-to-peer file-sharing network) and a venture-capitalist unicorn that tripled its valuation from $61.5 billion to $183 billion in 2025.
The Anthropic case is indicative of a systemic issue: AI is essentially theft. The data that AI systems need is often not acquired by entirely lawful, let alone fair, means. AI companies expropriate human knowledge, automate labor processes, patent them, and then attempt to sell all this back to us. Governments pander to tech oligarchs for fear of losing the technological advantage, despite widespread public hostility to AI systems. Yet labor unions and social movements are increasingly resistant to Big Tech. What lessons can we draw from these efforts?
In acknowledging the illegality of mass-pirated datasets, the judge in the Anthropic case recognized that the raw material of AI systems is the product of human labor whose ownership and consent for use rights cannot be ignored. This conception resonates with what legal theorist Salomé Viljoen calls a “propertarian” view, which frames data as a personal asset and grants individuals certain rights over their information. Prominent figures have advanced the propertarian conception. One example is Tim Berners-Lee, inventor of the World Wide Web, whose latest project seeks to decentralize the internet to restore data ownership and privacy. Glen Weyl and Eric A. Posner’s Radical Markets proposes treating data as labor, compensating users for their contributions to challenge so-called technofeudalism. A few politicians have echoed these ideas. Alexandria Ocasio-Cortez has called for stronger data protection and regulation. Andrew Yang has campaigned for data property rights and launched the Data Dividend Project. But most stop short of claiming that those who produce the data should have property rights. Why? Because this could expose contradictions in the liberal conception of property itself.
John Locke, writing in the late seventeenth century, claimed that “every man has a property in his own person.” By mixing one’s labor with nature, one acquires ownership of the product of that mixing. Farm a field and the vegetables are yours to harvest; chop down some trees and you may take the wood. Denis Diderot extended this idea in his Letter on the Book Trade, arguing that if the products of the mind are not treated as property, then “no one in society is master of his property.” More recently, Frank Easterbrook, writing in the Harvard Journal of Law & Public Policy, gave it a modern gloss: “intellectual property is no less the fruit of one’s labor than is physical property.” There is a certain intuitive appeal here. These arguments affirm the dignity of labor and the justice of distribution according to contribution.
Turning data into property, however, is more complicated than using patents and trademarks to establish intellectual property claims. It first requires the transformation of labor into machines able to collect data. Then it requires the transformation of the data—often through human moderator or programmers—into the clean, standardized, and translatable forms that other machines can use to in turn surveil human behavior and automate processes. For example, the microtask platform Amazon Mechanical Turk states in its terms and conditions that it reserves the right to use worker data to “improve the Site and other machine learning related products and services offered.” Consumers of any range of Amazon products, from Alexa to Amazon Web Services, also agree to allow the company to use their data as the company sees fit.
The propertarian view has some drawbacks. It ignores the historically specific conditions under which labor is organized through the various technological and political systems we have encountered under capitalism. In his recent book, The Eye of the Master: A Social History of Artificial Intelligence, the philosopher Matteo Pasquinelli notes that much of the discourse on AI obscures the fact that its “intelligence” is nothing more than the expropriation of collective knowledge rendered into commodified form. This perspective has older antecedents. In the “Fragment on Machines” in the Grundrisse, Karl Marx developed the idea of the “general intellect,” arguing that knowledge embedded even in the humblest forms of work is later appropriated by machines and turned against the workers from whom it was alienated. The Italian revolutionary Antonio Gramsci seized upon on this idea in his Prison Notebooks, noting that “all men are intellectuals. . . but not all men have in society the function of intellectuals.” In other words, society holds some types of knowledge in higher regard than others, yet everyone is capable and contributes to the collective knowledge of humanity. Reconstructing this genealogy in the era of large language models clarifies that capitalist production has always hinged on the alienation of knowledge, with the mechanization of intelligence rooted in the division of labor.
Channeling Marx, Viljoen is critical of proposals to make all data personal property, claiming they ignore the collective nature of data production and potentially block its socially valuable uses. The collective nature of data creation means that the solutions to its theft must also be collective. Yet there are potential advantages to treating information as an asset belonging to the individual from whom it originates: it would allow individuals to rent usage rights or to sell the data itself, transferring full ownership. Treating data production as a form of labor could also entitle individuals to receive wages in a data-labor market. Whether an asset or a wage, the aim is to redirect the wealth generated by data away from large technology firms and toward the individuals who produce it.
Data is already implicitly coded as capital through contractual agreements and trade secrecy, but this legal framework primarily benefits corporate holders and excludes individuals. Contractual terms in platform and data‑sharing agreements give corporations asymmetrical power to appropriate, aggregate, and monetize data, while trade secret law further entrenches data as a proprietary, revenue‑generating asset. Data commodification disproportionately benefits data‑collecting corporations rather than data-producing individuals. The alternative to data as property is what Viljoen calls the “dignitarian” view, which focuses on individual privacy. Such governance models afford workers and consumers limited privacy controls or access rights rather than any share in data‑derived value. The issue is not the classification of data as capital, but the allocation of rights to that capital at the expense of labor. Contract, property, and employment law must adapt to the datafication of the uncoded world and recognize that data extraction is an exchange without equivalent. Such recognition would challenge the quasi-enclosure of data by firms like Google, Meta, and Amazon and seek to alter who benefits from its circulation.
This redistributional demand is not unlike historical struggles over wages under capitalism: labor power remained commodified, but workers fought to capture a larger portion of the value they generated. The danger is that laying claim to data as property concedes ground to the commodification of everyday life, rather than resisting it. Extending the socioeconomic models of platform capitalism could deepen the asymmetries of ownership, power, and control that monopolistic firms thrive on. What is gained in individual remuneration may be offset by the reproduction of broader systemic inequalities. The alternative, of course, is public and collective ownership.
Unions have already caught onto the issues surrounding data production and ownership, pushing for industrial agreements and regulations more in line with the propertarian view. Equity, a trade union representing 50,000 performers and creatives in the United Kingdom, has taken one of the strongest stances against AI companies’ appropriation of data. “The industrial-scale theft of our members’ data must end and performers’ rights must be respected in relation to AI use, whether past performances, or work being shot and recorded now and in the future,” said Liam Budd, a union official. Equity has highlighted concerns about recorded material being used to unlawfully train AI systems or generate digital replicas. While existing protections covering intellectual property, performers’ rights, and data protection should already apply, there is a lack of clarity from industry officials and the government about how they interpret the law. An open letter sent by Equity’s General Secretary Paul W. Fleming to the BBC, Disney, ITV, Audible, Pact, and UK Interactive Entertainment says there is an urgent need to “establish fair and legally compliant licensing and data protection frameworks for AI” and ensure performers are compensated. Negotiations are ongoing.
In the United States, SAG-AFTRA’s former president Fran Drescher championed what she called “the most progressive AI protections ever written.” The union’s new contract, won through a disciplined and powerful strike in 2023, requires studios to secure consent and provide compensation before creating digital doubles. The contract mandates minimum day-rate payment for scanning an actor’s likeness and extends protections to deceased performers, requiring approval from estates or, failing that, from the union itself. It also obliges producers to seek union consent before substituting human roles with “synthetic performers” designed to resemble real people. However, studios may bypass consent if the likeness use falls under First Amendment exceptions, and they can compel actors to sign away AI rights as a condition of employment. There is also the risk that studios could generate derivative lookalikes without triggering these protections. Furthermore, the agreement individualizes negotiating, putting labor at a disadvantage. Dissenters in the union argue that these measures could normalize, rather than resist, the automation of performance. “Consent, in the context of this agreement, is tyranny,” said board member Matthew Modine.
The United States has far weaker statutory protections from data theft than the UK and the European Union. There are relatively strong copyright measures and laws that establish some data privacy, particularly around financial data, health data, and data generated by children. Yet there are limited enforcement mechanisms and no comprehensive framework to protect workers from the risks of digital datafication at scale. Recent concerns have spurred state lawmakers in California, Colorado, Connecticut, Utah, and Virginia to pass restrictions on the collection, transfer, and sale of location data. These are modeled on dignitarian protections like the EU’s 2018 General Data Protection Regulation (GDPR), which is designed to limit private data collection by default, requiring users to “opt in” rather than “opt out” of collection. Meanwhile, Donald Trump has rammed through a barrage of efforts to deregulate AI technologies during his second term, starting with his January 2025 Executive Order “Removing Barriers to American Leadership in Artificial Intelligence,” which rescinded what his administration labeled “burdensome” regulations like Joe Biden’s executive order on creating “Safe, Secure, and Trustworthy” AI. The administration has developed a plan for “unquestioned and unchallenged global technological dominance” that aims to ensure limited federal regulation. It’s been reported that Trump and other members of the GOP want to add an amendment to the National Defense Authorization Act that would block states from implementing their own safeguards.
Social protections for workers and consumers against AI tech do exist in Europe, though U.S. pressure is wearing them down. The EU’s AI Act aims to set binding rules on transparency. It builds on GDPR protections in the workplace: employers cannot feed your voice or image into training datasets for surveillance or productivity systems without permission. Article 50 of the act aims to establish machine-readable watermarking for generative AI outputs. And Article 53 requires firms to disclose their training data for general-purpose AI models. Fines for violations also scale with the size of the business: up to €35 million, or 7 percent of global turnover. While controversial for some who advocate for sex worker rights and more targeted safeguarding, the United Kingdom’s Online Safety Act empowers the state to fine firms up to 10 percent of worldwide revenue for failing to proactively remove nonconsensual sexual images, including AI deepfakes. France’s 2024 “digital spaces” law criminalizes distribution of AI-created content without consent, with prison terms and fines up to €75,000, and gives regulators power to order removal. Denmark recently introduced a bill expected to pass in 2026 that would grant every person copyright over their own likeness, meaning that nonconsensual “digital imitations” would be illegal to produce and to share.
Many of these developments are steps toward a better system of rights around worker and consumer data. At the same time, the overwhelming majority of AI models and cloud services are controlled by U.S. firms. As Mario Draghi argued in a 2024 report for the European Union, digital sovereignty—the ability of societies to control their digital infrastructures—is now key to political autonomy. Placing limits on datafication has geopolitical importance in countering the hegemony of U.S. tech monopolies. Europe is currently building a scaffolding of digital sovereignty. The European Union’s Digital Markets Act, for example, establishes the right of EU citizens to data portability, or the capacity to download individual data and move it to another platform. This, in theory, undermines the power of U.S. platforms by allowing users to transfer their data. It means that end users and authorized third parties must have access to these essential tools. The act also acknowledges that individuals have some right of ownership over their own data. But this is not enough.
AI is turning the online world into a sloppy simulacrum, where prompts and inputs render human experience as a specter in the digital machine—all under the guise of progress. Like with previous generations of skilled craft workers, intellectual and creative skill is being expropriated unpaid, injected into systems that obscure the origins of their own acumen. As put plainly above, AI is another opportunity for capitalist theft.
Given the state of digital capitalism today, it may be that individuals and collectives can only contest the theft of data by formalizing property rights over data as a step toward social democratic data governance. While earlier generations of the labor movement fought for the eight-hour work day, a wage floor, and health and safety protections, such bargaining is no longer enough. Unions could seize the opportunity to grow again with public support, if they set the agenda now.
If labor wants to slow AI’s theft of wages and data, they should push for collective agreements with data rights: limits on biometric capture, use-rights for workers’ likenesses and voices, protections against algorithmic management, and guarantees that AI systems cannot be trained on employees’ outputs without compensation. Statutory penalties for violations (including criminal) must be large enough to change corporate behavior. If capital is rewriting the rules of work, then workers must rewrite the terms of resistance. In seizing data rights, workers can turn AI from a tool of extraction into a tool for democratic power. Our choice is between a future in which we own and control the datafied infrastructures that shape our lives, or one in which capital continues to steal from us.
