AI governance concerns who decides how AI is developed, deployed, disseminated, used, whose interests it serves, who can challenge its decisions, and how. Liberal democracies generally distribute power among corporations, states, and to a certain extent also the people. Democratic governance requires public participation, enforceable accountability, and meaningful control over the infrastructures shaping collective life, including AI systems.
Here we will examine this issue through two influential French thinkers, Louis Althusser and Michel Foucault, both of whom have an extensive body of work focused on power and domination in capitalist societies. Our goal is to understand the relevance of their ideas regarding AI governance. Do AI systems maintain or amplify the capitalist order? Do they contribute to, or as some say, enable emancipation?
Althusser, a Marxist philosopher, examined how capitalist societies reproduce the class relations on which they depend. In his view, apart from repressive state apparatuses, such as the police, courts and army, domination is sustained through ideological state apparatuses, such as schools, religious institutions and the media. People’s beliefs and social roles are shaped within institutions and practices that help reproduce existing class structures. Somewhat differently, Foucault, a philosopher and historian, examined how power operates throughout society through institutions, bodies of knowledge and everyday practices i.e. not just the state. His studies of prisons, schools and hospitals explored how surveillance, discipline and norms shape people’s conduct and understanding of themselves. In his view, governance encompasses both how populations are governed and the ways individuals are encouraged/taught to govern themselves. Read together, their approaches raise a question central to AI governance: What is the role of AI in reinforcing or mitigating institutional power and the everyday shaping of human behaviour?
Althusser would position AI as yet another tool at the disposal of states and corporations to expand their ideological and repressive apparatuses: Algorithms embedded in platforms, education, hiring, and media profile, score and/or nudge individuals in desirable directions while enabling intensified surveillance, thus stabilizing class relations reproduced through digital capitalism.
Drawing on Foucault’s account of power, we can view AI not necessarily as an expansion of the ideological/repressive apparatus but as an extension of disciplinary power across state and private institutions through data collection, prediction and the shaping of conduct. In this scenario, populations are not necessarily governed by overt state coercion but through diffuse systems that produce self-regulating subjects (through rankings, recommendation systems etc).
The AI power landscape today is a combination of concentrated ownership and distributed influence. As such, each of these two perspectives offers insights while also falling short of seeing the new challenge posed by AI systems. The fact that state-capital blocs in varying degrees control models, data and compute is in line with Althusser’s political analysis. At the same time, the decentralized, distributed and internalized aspect of AI is reminiscent of Foucault’s view in that consumer conduct is shaped through decentralized interfaces making domination less visible or indirect, not by command, but through more effective automated norm-setting and behaviour steering.
So where will power actually reside once computation becomes infrastructural, global or partially autonomous? Althusser’s state-centered scheme of coherence and intentionality does not paint the full picture of the contemporary AI power which is not cleanly orchestrated by a unified state apparatus but is fragmented across transnational firms, open-source ecosystems, cloud infrastructures, and opaque supply chains. As we have seen in the surge of recent years’ AI models, these systems often outpace state control rather than serving it. Also since AI systems continuously reconfigure subjects through real-time, adaptive feedback loops, the concept of state ideology is less a structure than a moving target.
Foucault captures this diffusion of power but risks misidentifying the stark asymmetries introduced by concentrated compute, proprietary models, and capital-intensive scaling; his analytics of discipline and governmentality struggle to account for platform monopolies and infrastructural chokepoints that reintroduce quasi-sovereign control.
But perhaps more fundamentally, since both thinkers have built their theories on human-centered power relations, they fall short of addressing the non-human decision layers whose operations (optimization objectives, emergent behaviors) cannot directly be reduced to human intention or discourse. This complicates the classical notions of domination, discipline and governance, all of which need further elaboration. As decisions become harder to trace, establishing the responsibilities of those who build, own and deploy AI systems, while protecting the rights of those affected by them, becomes increasingly urgent.
CounterCritical