Making AI Work: A Critical Theory of AI Production

In 1930, critical theorist Max Horkheimer described modern society as a skyscraper, in which each level represents a different social layer (1978, 66–67). Horkheimer explained that the top of the skyscraper is occupied by the capitalists, right below them are wealthy landowners and those with powerful jobs, followed by professionals like professors and engineers. Further down the tower are craftsmen, grocers, and farmers. Down the bottom, “the unskilled and the permanently unemployed, the poor, the aged and the sick.” Even below that, “we encounter the actual foundation of misery on which this structure arises,” Horkheimer writes, arguing that the existence of modern capitalist societies is based entirely on the exploitation of this majority world. In “this house of present-day mankind,” he concludes, its basement “is a slaughterhouse, its roof a cathedral, but from the windows of the upper floors, it affords a really beautiful view of the starry heavens.” Through an analysis of the ways in which AI workers are misrecognized, this article shows that the AI industry resembles the skyscraper that Horkheimer described. Horkheimer wrote that those on the top floors of the skyscraper would be “dizzy” if they were to look down and see the foundation of misery on which their comfort zones are built. By mapping the entire production chain of human labor that is needed to make AI work and by highlighting all the harms and wrongs that arise along that production chain, the article aims to make the reader dizzy. Recent research and reports have pointed to various forms of hidden human labor involved in AI development—such as data annotation, content moderation, or other types of “microwork.” This research stems predominantly from the fields of law, sociology, and economic geography. However, so far, few scholars have approached the topic of AI labor in an all-encompassing manner, by considering all the different types of human labor needed to make AI work, that is, AI's entire production chain (noteworthy exceptions being Crawford and Joler 2018; Fuchs 2014; Muldoon, Graham, and Cant 2024). Furthermore, while plenty have already emphasized that certain forms of AI labor are performed under precarious and exploitative conditions, the literature on this topic still lacks in-depth normative theorization. Under what conditions a labor practice can be called precarious or exploitative, and when and why these issues can be deemed unjust, are not self-evident matters from the viewpoint of ethics and social philosophy. This article therefore seeks to contribute to existing literature on AI labor, not only by drawing a comprehensive map of AI labor and its associated harms and wrongs but also by bringing forward a critical theory approach that helps us to better understand the injustices of the different harms and wrongs occurring along AI's long production chain. The critique of AI labor presented in this article is based on one of the most elaborate recent models of critical theory, namely, Axel Honneth's theory of recognition (Honneth 1992/1995). The core of the critical model this article puts forward is that the different harms and wrongs of AI labor are instances of misrecognition of people as workers and inhabitants of the majority world.1 One may object that Honneth's work on recognition is not the most obvious choice for a critical theory perspective on AI labor. Other valuable approaches would build on the notions of exploitation or alienation or draw analogies with the critiques of capitalism developed by the first generation of Frankfurt School critical theorists.2 We agree that the approach taken in this article is not the only possible one for a critical theory of AI labor—which is why the article received the humble subtitle “a critical theory of AI production.” That said, we propose that Honneth's theory of recognition is particularly valuable in capturing the variety of injustices and harms many workers experience in the AI industry, as these injustices and harms stem not only from structural economic asymmetries but also from social, cultural, and geopolitical power dynamics. In our discussion of AI labor, we choose the term “production chain” as it that human is needed to and and to the of all the human involved in AI one also of the production the and chain” and or economic like chain” and However, these not the social and the of AI labor The article of The first a comprehensive map of AI labor that the entire production chain of This map not only the variety of human labor needed to make AI work, but also the of the AI industry and the of the different forms of labor The of the article of a of the harms and wrongs associated with AI labor. This is by the of AI labor, its and the power asymmetries AI's production chain. In the and of the we these types of harms and wrongs the of Axel Honneth's critical theory of We that this critical perspective is particularly to and what it is that labor and In to the different of the as Horkheimer for capitalist society in the this the production chain of human labor needed to make AI We the production chain of AI data model and each we the different in the of this This is based on research on and the Muldoon, Graham, and Cant and as as research the and of AI 2024). The map presented in this to the literature by the of the AI production chain. The article to research on the of AI by highlighting the that each of AI's not only a and but also a social and human the production chain of AI shows that AI labor is performed all of the the are in of the Crawford AI are a are on the top of the skyscraper, people to a view of the starry In the it is the precarious labor of those in developed of the that the on which the AI skyscraper is being built. we mapping the AI production chain, a few on are We the term to to all the different types of involved in AI work along the chain of The term forms a of AI labor, which the to with the and of data 2024). the of we the term to to all types of to workers labor or called or for as and for AI and AI The production of this various of various forms of human labor. all with the of and other needed in the production of which and other forms of labor involved in and 2024). of an for AI is which is predominantly in and and to be in to be in and other types of and Joler of all in AI are in while other are in and the are needed to 2024). to be to house and Graham, and Cant 2024). labor is also needed to and and to from their to and from to and Joler and are in the as as in the of AI's production chain the human labor involved in and is a of AI's production chain One and AI models variety of are involved in the data of the production chain of to all types of labor involved in data data as as or existing data is a of and to make them for model 2024). to a of for each in the to a that is, a or what is in the and is for to data and as or as or and by the on these of with a of the work a data be and to those the data the work of data by and data analysis data have and and on 2024). are different types of a in them to AI many other forms of labor are needed to make AI work, it is the involved in the model that are most associated with are also the types of AI labor. on an in to other forms of the of a data in the is first of AI labor in this is AI AI research forms a and industry and is involved in approaches to AI of and research on the topic 2024). is that of the AI to the data the AI on the and needed for AI model 2024). AI “a of and to build AI can be as or a term in to “the people understand to to from of and are to have particularly in and a in and and also data but the of a data data Furthermore, workers are not involved in or models but contribute to AI models by them on This of labor can be called AI are also various forms of human labor that an in the model but not One of labor is AI AI a and that of model is experience which that different of an AI as the are and to the 2024). and is the and of AI or in the 2024). but not AI or to the AI models are to various types of workers that the models as also make that the involved models to and to or issues 2024). a and a is in practice 2024). to and is an as which a can an or other forms of In the of for a and by the AI also and critique the and normative of a and research to the normative that AI can also be for that AI are with Furthermore, they are based in or in industry, AI Other human labor involved in the of AI's production chain to with the and of AI human labor to the of AI's In the this the of by human workers workers are involved in a to the on their are under the that recognition is the needed to data and models is for or different are this and of the but not to in like or and In these many work it is that in in and in in involved in of and of and of the labor involved in in the This that it is to of this labor and also that it is to workers from the many that they are to when the of and the conditions of those involved in is needed 2024). the production chain of various harms and wrongs This these harms and wrongs of AI labor, them precarious conditions, and power The not issues that would have not in on forms of AI labor. it to the existing literature by a of the different harms and wrongs associated with AI labor. different forms of AI labor the conditions of AI However, conditions can be deemed or precarious on various the discussion and understand the the different forms of that different AI workers we and the of AI labor based on different labor work and to with the of to an of the production chain of the AI industry of the labor and in certain by AI labor to these AI labor can be the or only for workers in or are and for labor is and This is, for the for workers in workers data workers on and or those that are in Graham, and Cant 2024). is the of that is, or it would be for workers to their In the of AI labor, it is on or like is In many not is in as workers are by the as workers to this of workers are from their when are few on or to the 2024). The of those data work for is but to from a few to a and are the AI labor on labor, labor, as in the industry 2024). those for in are not by but to make a from valuable from the that were and is the to and that one to have a to is the to in a and an in to but also the involved in the data are by the that are This it for workers to the AI labor but in many it also that is not That is not to that these workers not have valuable or that their work not make a In the of work, in and from workers to in the and which from data work by being can workers with other types of data as data or AI is have in to and with which associated with a better and However, to the of in the of data workers to their and and in of their work in to to their to the in to that conditions not the and of workers forms of labor involved in AI's development—such as and the of be and their on the However, recent research that as as also in on to the content that data workers are to and to the of Furthermore, the that work in of and can and and workers to be and their and most is the labor involved of the AI production and in the conditions in these can be and if one can and in the or if the the to This is the of the AI skyscraper few the top and of the skyscraper would for many AI jobs, in core and in would By the of AI labor of with for or to are the or like and In all of these is workers that an is in and many workers in the AI industry their from work or work, a is from for many AI Furthermore, one that the that AI workers in is not or the by the industry that they are a of 2024). The by the AI would not be possible the labor of and in to the that these AI workers for the they to the and of the AI production chain, the of labor and the of labor like work, as a or social are or to Muldoon, Graham, and Cant and to the of workers to conditions, for by or of AI labor this 2024). on it to issues the and and the of the different on the AI production chain it for workers to or other of and to Muldoon, Graham, and Cant 2024). the and of the AI production chain make it for AI workers to understand the of their labor recognition for it issues are in the The human workers are needed to make AI work are Graham, and Cant or and which is why they can also be deemed and In this we the different of the of AI labor. AI labor is from society in a by being from The many data workers their is their and are to the that labor those with or to an a of from the is that the of their labor its precarious conditions are by AI workers are also hidden they a This to with the of the AI industry and the of labor the production chain of in are in or and work in and data and are performed and workers are when we different types of workers as to the namely, the we that this is a the human labor that AI work is the of and that are to be they are presented as are in still on human to in the of human to and AI are hidden the of on human labor is and by the the of modern can be a in the of that to the and of human labor to AI is also an that the industry to the of human labor and labor the term to this practice of of term in this is and which all the human work involved in and In to being hidden to are various of AI labor that are hidden to workers of in many of AI labor, it is to workers their are or what they to of the types of data work involved in AI work are by like or which can the of their and and social like and not only to each other different and to the but also to the actual they are Other types of AI labor are of like or In these as the is to but not the and down the production chain, which their labor is the and the of to workers they on and to other workers involved in other of the production chain of AI and workers not the or are not in the not that their labor is not and a those AI models on those involved in and but also on those or those and the of their labor or the it is hidden to This to with the of labor. and their labor only a of a and chain that an with many forms of AI labor therefore that are workers not have a view of the they contribute to a and and in data workers not that they are labor. are various ways in which as social the and of the or of their Fuchs the that that a is not a to recognition the and that social can be as to AI models have to to these of data and and and 2014; and to a as are various forms of and are also different power along the AI production chain. The AI production chain not along a but a to as workers along that chain are in understand and critique the AI industry, and to all the different cultural, and and each a in the AI industry and In this we of the most power to AI labor, a exploitation and to structure the to is not only an economic but also a social in which the of on power different social which those to those only is their labor are a few of AI labor that forms of exploitation and the in the power of labor. of the involved in AI is by the which to the labor and The to labor to in a to the bottom, the of workers while their and Furthermore, AI are to the AI to better practice also as a comprehensive see The power of data forms of a and and all-encompassing and of the work from the which the and content of of and by workers from each other and which and and by workers to a of the asymmetries workers and which are particularly in the of and also contribute to the exploitative of AI of on the of workers their can all of while workers the they are for or the their labor is are to not have this but also they are on the they as as the they the of that are and modern capitalism and present-day and so it also the AI AI's have called it and power but it also and and In the AI production chain the of AI and data to the In this are on and that to this as the that is still an in The of the AI industry on human labor also to geopolitical dynamics. on the labor of workers in and and on the in other of the world. these economic be the of labor and other from and and with to is this can be described as practice and and and The and many AI workers are based are the of and exploitation of labor and under still the present-day Furthermore, the AI industry on the of or that and shows that a on exploitative workers and in core not the In to labor capitalism also on the of labor, and other from the of capitalist that exploitative may be unjust, but workers can still be deemed and on the other and labor. exploitation while is “the of those is a hidden of for the of those it an of the we understand the AI labor is not described in the The production chain of AI also on the of labor as the labor of or the labor of and have to other the AI industry is not only by a and but it also these power dynamics. Through the of the industry and it and the it AI in the and and other and 2024). on the production that make AI work stems predominantly from like law, sociology, and economic geography. a this research The harms and wrongs of AI labor, in the have by various However, in the existing literature on the a normative analysis of AI labor and of the harms and wrongs associated with it is and to the normative of AI and other the topic of AI labor is in these that to for the and of the AI industry and its normative being and The of this is to these and the of the harms and wrongs that along the production chain of We propose the issues in as injustices different forms of misrecognition and of In we draw on the model of critical theory presented by Axel and and The the of recognition theory is the that human are social for on they are by This social by and each other (Honneth this is not an of not being with or not being by but a a their with namely, when the they have with in society that their and are a is in for the to their By when a is recognition or is misrecognized, this to their for to recognition theory, first and the of social which that them to to their in to that a of as social or and is not it but it namely, the of social and and which in and for to this as the of the that people and it (Honneth of and social particularly of and normative associated with forms of (Honneth The first of that is, is the recognition of a and these are the can have to with the world. is the recognition of the as a of the to and in their This that the is as a in This of recognition is which a and to a the recognition of the as a to the of the their are a can Honneth's on which a of However, work in and research to that recognition that recognition theory puts forward an normative core that in a namely, “the of which can as a normative in a of recognition theory in to other is its to the of as pointed by the that it the of and in an of We agree with these and propose that recognition theory can be to production it in the of social In production like the one the of and in one of the are to and have a on the of and in other This to social with its normative recognition theory a for a critique of AI labor. of recognition theory as a for AI labor is that it not in the of as other of but it to the experience of that from social is as the is the an social on and for theory to approaches to in that it to the experience of In the of AI labor, this that a recognition approach us to critique the social that is the AI industry by the of AI workers as its The core of recognition theory it particularly to the of and of that AI labor. a of the AI production chain is the of the different and that in the of and on the work of of and hidden many of are in the theory is to a normative on these AI workers AI's production chain. the of not on a social or but on all the or from the work of as an of different social and it the many have the to existing of and so This as as in considering what would The of many forms of AI labor, in core and in in the to a from the of to and is it would workers a better and so This of an is a of of on the AI production chain from the instances of are as to of is can be and their to be for the it to the AI In many this of is as by they is not the to the different forms of work but on a a of that would to them that they as to the AI industry in which they are in the of AI labor is the and of the production chain of By to workers in other of the and the of labor or and that and to the and of AI's production chain, it is which and are for the of AI labor Muldoon, Graham, and Cant and these of the AI industry also to the recognition they forms of as and the of recognition and have different and when it to considering them on a In the of a is the of many forms of and social of and that However, in its and are from or the of for AI workers in the of and production of misrecognition a in the of the human labor needed to make AI That AI labor is and in various ways hidden is a in and In its most AI workers are hidden from the one in a of only the or in a in from the to and that AI workers are the level of This misrecognition is by the of recognition in the of and the of on the of and AI labor to that work hidden is a to and Furthermore, in to a of for their and and a of and hidden AI workers for the and of their to the of AI and the in this article is that all types of human labor in are in the production chain of each of these forms of labor, the existence and of AI would not be However, the that many forms of AI labor hidden that this labor is not as to make AI work or that it is not possible for workers to be as In the of AI labor is the of the of for labor. In other it is the of AI labor that a of By certain types of labor as data or AI as and in the of while or the existence and of other are based in developed the workers in the that is, the majority world. different types of labor and their are other of the which and may obvious that the of of the human labor needed to make AI work is a of but not that is hidden or is in a to recognition as social is the of and the that human labor is needed to make AI work, as as the precarious conditions involved in AI labor, a of of the in the majority and world. Furthermore, the of workers to each other also to a of AI workers may be and different but are of the far, we have AI workers are in their as We that their be that to be in and that societies to the of their labor. However, the make AI work are also on other their labor. The misrecognition of AI workers is not to a of for their work, of labor and for certain types of labor, but it also to with and and power The misrecognition of AI workers social and of The exploitation and of labor power by the industry not only people as workers but also as to certain or AI workers are in and forms of and majority which the for they or that be in their to the injustices they we in the AI production chain not only asymmetries of power but in many also them or the that AI for or the of approaches to AI it is to be that the and the of its and the and of people in the majority world. The power asymmetries and the that are only the of the when it to injustices that in the of In skyscraper, from the are in the upper the floors the top may be of but the the of the tower also of In this we have that normative of AI that only on the of AI on many of the injustices in the AI AI is a with a production chain. like or data not be to AI but these are in the of labor the on which the AI skyscraper is built. the hidden and precarious labor of workers in the the modern of a society would like a tower are In this we have that, while a variety of human labor is to make AI work, only a of those labor recognition in the of conditions, and and a of the by the AI Honneth's theory of we that the AI industry the recognition of the and of all types of workers involved along the AI production chain. The of recognition is of and the injustices involved in AI labor. recognition theory to understand the variety of harms and wrongs involved in AI production a This research by the of and by The of