How AI Transforms Collective Hermeneutical Resources

This paper examines the impact of artificial intelligence (AI) on collective hermeneutical resources (CHR), the shared tools and concepts through which societies understand themselves and the world. It argues that AI, through algorithmic profiling and datafication, is fundamentally altering the way CHR are formed, stabilized, and accessed. This shift raises concerns about epistemic injustice, particularly hermeneutical injustice, as individuals are increasingly subject to AI-generated meanings and categories that they cannot fully understand or control. The paper explores the implications of this new reality, highlighting the problems of accessibility, privacy, and negotiation in relation to AI-driven CHR. It concludes by proposing the development of structured meta-CHR to interpret and evaluate AI-generated meanings, emphasizing the need for human collaboration and control in the face of AI’s growing influence on our epistemic lives.

Introduction 

Epistemic injustice affects both our digital and physical worlds. Recent research in social epistemology has shown that “there is no clear distinction between offline and online knowledge production and dissemination. . . . We cannot fully understand how knowledge is produced and circulated in our era without understanding the role of the Internet.”¹ Additionally, “epistemic resources are being located in technological infrastructures, many of which individuals do not have the capacity to manipulate” or understand.²

Algorithmic profiling, which characterizes and unifies our digital lives, exacerbates the lack of epistemic friction—exposure to perspectives and experiences different from our own—thereby contributing to multiple forms of epistemic injustice.³ Stewart et al. describe algorithmic sorting as “the increasing separation of people on social media into different epistemic worlds which have little overlap in informational content . . . the sorting of social media users into increasingly different, increasingly closed-off, and deeply biased informational spheres as a result of algorithms and algorithmically constrained human decisions.”⁴ This exacerbates forms of hermeneutical marginalization⁵ and willful hermeneutical ignorance.⁶ Automated inequalities caused by ranking algorithms in eligibility systems and predictive risk models in finance, healthcare, and social services demonstrate that datafication and algorithmic profiling are shifting the credibility economy in favor of machines, undermining individuals’ credibility and narratives, and resulting in new forms of testimonial and hermeneutical injustice.⁷

Beyond social media, the increasing integration of algorithms into all aspects of life and the outsourcing of institutional decision-making to them generate significant hermeneutical injustice toward citizens. The opacity of these systems, the lack of transparency, and the institutional structures in which they are embedded make it difficult for citizens to identify or describe the extent of the epistemic wrongs done to them. We often lack the hermeneutical resources needed to identify when, why, and how a wrong is being done to us, and when our epistemic agency is being diminished.⁸

Recent scholarship on epistemic injustice and algorithms suggests that the relationship between collective hermeneutical resources and algorithmic systems remains insufficiently understood. Several key issues emerge from this literature:

1. Algorithms and technological infrastructures increasingly shape the formation of collective hermeneutical resources by curating the content to which individuals are exposed.
2. Collective hermeneutical resources appear to exhibit a bias in favor of algorithmic outputs, much as they have historically privileged dominant epistemic agents. This phenomenon is often described as automation bias. Can algorithms therefore be understood, at least to some degree, as epistemic agents?
3. Algorithms may themselves function as hermeneutical resources, accessible only to a limited group of actors—technocrats, corporate owners, and technical workers—while nevertheless shaping broader social understandings and shared interpretations of the world.
4. Collective hermeneutical resources traditionally depend upon a public sphere capable of generating and circulating shared meanings. In the digital era, however, that public sphere appears increasingly fragmented: opinions are shaped through data-driven processes, and social interactions acquire meanings that may differ from those understood by the individuals participating in them.
5. Individuals are continuously observed through algorithmic systems across a wide range of activities. These systems generate meanings, classifications, and profiles that are subsequently applied to individuals, often in ways that remain opaque to those being classified.

This new situation gives rise to new forms of epistemic injustice, particularly hermeneutical injustice. Algorithms construct individuals through hermeneutical resources that remain unavailable to them, thereby producing forms of alienation. Big Data society is generating collective hermeneutical resources on a massive scale that are not genuinely collective: they are not readily accessible to those whose beliefs, feelings, and behaviors they shape.

This paper examines the concept of collective hermeneutical resources as it has been developed in the literature on epistemic injustice and seeks to reconceptualize it in light of the datafication of the public sphere.

What are Collective Hermeneutical Resources?

According to Miranda Fricker, collective hermeneutical resources are the shared interpretive tools through which a society understands itself, others, and the world, including concepts, discourses, and historical narratives. These resources are formed, stabilized, and influenced by institutions such as the media, law, and academia, and circulate through the public sphere as a kind of “semantic commons” that enables individuals to make sense of themselves and others. Fricker also refers to a “theoretical atmosphere” in which social interactions occur and to the “intellectual gear” through which individuals navigate the social world epistemically.⁹ The two forms of epistemic injustice identified by Fricker can be traced to collective hermeneutical resources. Testimonial injustice arises from stereotypes and prejudices embedded within these resources that diminish the credibility of members of particular social groups. Hermeneutical injustice, by contrast, occurs when collective hermeneutical resources lack the concepts necessary to make sense of a particular experience. As Fricker defines it, hermeneutical injustice is “having some significant area of one’s social experience obscured from collective understanding owing to a structural identity prejudice in the collective hermeneutical resource.”¹⁰

Critics of Fricker’s concept of collective hermeneutical resources, including Rebecca Mason, Kristie Dotson, and José Medina, argue that it does not adequately account for the epistemic agency of marginalized groups. Marginalized communities have always developed concepts and interpretive frameworks through which to make sense of their experiences; they possess their own hermeneutical resources, even when these are ignored, excluded, or delegitimized by dominant groups.¹¹ Ignorance, moreover, may be as significant as a lack of resources. Ignorance, as Charles Mills describes it, is “not knowing what we do not know” and may itself produce hermeneutical injustice, rather than a simple lack of hermeneutical resources.¹²

Another possible criticism is that Fricker’s concept of collective hermeneutical resources appears overly universalistic. In her account, CHR are shared through a “global” mainstream narrative that combines commonsense understandings of social reality with institutionally produced classifications and meanings. Yet Fricker does not fully explain how this immaterial public sphere is organized or hierarchized. In politically charged periods within Western democracies, CHR are not a “given” set of resources readily available to everyone. Rather, they are continuously negotiated within public discourse and are often highly politicized. Institutions constantly produce new social kinds that challenge traditional or intuitive categorizations—as in debates surrounding gender roles—and these classifications frequently become sites of political conflict and polarization.¹³ CHR may therefore be publicly acknowledged without being deeply accepted as shared meanings. Individuals may privately maintain beliefs or prejudices that differ from publicly expressed views, even while outwardly conforming to dominant hermeneutical frameworks. This can generate phenomena of pluralistic ignorance, in which members of a group mistakenly believe that their private thoughts, attitudes, or behaviors differ from those of the majority when, in fact, they are widely shared.¹⁴ Pluralistic ignorance is thus a hermeneutical phenomenon with significant political consequences.

CHR are therefore not a homogeneous body of commonsense knowledge shared consensually by a group. Rather, they are the locus of violent power struggles for the control of public meanings and categorizations that shape behavior, rights, and societal practices. A compelling example is the classification of mental disorders. The Diagnostic and Statistical Manual of Mental Disorders (DSM) is often regarded as the “bible” of psychiatry and serves numerous bureaucratic functions. As Ian Hacking notes, individuals seeking insurance coverage for mental health treatment in North America must generally receive a diagnosis that conforms to DSM classifications and coding systems.¹⁶ Naming and classifying nature and society are among the most fundamental meaning-making activities of humans; through them, people negotiate the world they inhabit together. What is named, left unnamed, or wrongly named causes most epistemic injustices.

To address critiques that the concept of CHR is overly general, Fricker distinguishes between maximal, midway, and minimal forms of hermeneutical injustice.¹⁷ Tristan Goetze similarly argues that the expression “collective hermeneutical resources” should be retained at a macro-level of analysis:

The central bank of the hermeneutical system is what Fricker calls the collective hermeneutical resource. This is the set of concepts and conceptions which are shared all but universally. . . . It is a common interpretation to read the idea of a collective hermeneutical resource as being specific to a culture or society but the notion could be extended to include all of humanity—the extent of our shared concepts is an empirical matter.¹⁸

Goetze also introduces two more scales of hermeneutical resources: communal hermeneutical resources, which are generated and maintained by particular communities or social groups (for example, feminist hermeneutical resources), and personal hermeneutical resources, which belong to individual epistemic agents and are shaped by their participation in different social and epistemic communities as well as by their subjective experiences. Furthermore, communal hermeneutical resources are defined relationally, that is, in relation to a particular epistemic agent. Goetze distinguishes between local hermeneutical resources, which circulate within communities to which an epistemic agent belongs, and distal hermeneutical resources, which circulate within communities of which the agent is not a member.¹⁹

Collective hermeneutical resources are thus a complex concept that cannot be uniformly applied across all communities or distributed homogeneously among them. At a minimum, three levels of hermeneutical resources can be distinguished: the personal level, through which individuals make sense of their lived experiences; the communal level, through which local communities interpret and negotiate shared experiences; and the institutional level, through which mainstream universal concepts and categorizations make sense of people’s experiences.

Do We Have Access to Collective Hermeneutical Resources Online?

Collective hermeneutical resources are the symbolic materials, both implicit and explicit, through which individuals make sense of their experiences in society. They shape social interactions, guide the acquisition and interpretation of information, and influence behavior. They also play an important role in the expression of emotions and in making sense of affective experience.²⁰ Although many of these resources remain only partially explicit—including stereotypes, values, implicit biases, and linguistic expressions charged with unjust or derogatory meanings—they nevertheless function as a background framework through which individuals make sense of their feelings, experiences, and actions in the world.

Today, most collective hermeneutical resources are privately produced by algorithms and generative AI that interpret our online behavior and regulate our offline behavior. Algorithms and generative AI use Big Data to learn how to think and conceptualize in a “human-like” way. They label and classify user behavior to generate commercial recommendations, improve medical diagnoses, identify military targets, predict events such as the spread of viruses, and more. Through these processes, they contribute to the production of collective hermeneutical resources by labeling, categorizing, and classifying objects, behaviors, and language.

Several types of data are used to train AI today: structured data (such as databases, spreadsheets, and system logs) and unstructured data (such as texts and images). Unstructured data is typically produced spontaneously by users and media (such as through social media posts, image archives, and newspaper collections). Images are used to train AI for facial recognition, robotics, warfare technologies, and medical diagnostics. Textual data are used to train large language models (LLMs) to perform sentiment analysis and extract information about people and places. Real-time data (such as location data generated by mobile devices) generate a constant stream that allows AI to predict behaviors. There is an important distinction between spontaneously produced user data and data collected by professionals with informed consent (such as medical data).

The processing and interpretation of data constitute a new mode of constructing and stabilizing CHR. Data analysis involves processing raw data and assigning “meaning,” done manually by experts or automatically through various AI applications. This labeling activity tends to rigidify CHR by transforming them into a unique set of less negotiable labels. A fundamental human activity—the ongoing negotiation of meanings and classifications through which individuals make sense of the world—is increasingly externalized to technical systems operating according to opaque and often biased forms of logic. As a result, artificial CHR freezes the negotiated meanings we create as humans, leading to new forms of hermeneutical injustice. For example, when AI training data systematically exclude or misrepresent the experiences of particular groups, it leads to a gap in CHR available to those groups to make sense of their realities. Similarly, data-labeling practices that reproduce stereotypes or dominant narratives further obscure the perspectives of marginalized communities, exacerbating hermeneutical marginalization. The lack of diverse and inclusive labeling teams results in hermeneutical blind spots, where labeling fails to capture the nuances of different social experiences. Inadequate access to data labeling processes perpetuates epistemic exclusion and injustice for certain groups.

Three Problems: Accessibility, Privacy, and Negotiation

People’s lives are increasingly inseparable from Big Data. Whether checking the morning weather, planning a route, booking tickets, or engaging in leisure activities, we rely on Big Data. Even work and air quality indicators are dependent on data. Data is omnipresent and governs the world, whether metaphysical or physical. This permanent and dynamic flow of data is becoming a new foundation of our CHR.

Three main problems arise from the automation of CHR: accessibility, privacy, and negotiation. Ordinary individuals often do not know how their data is extracted, by whom, or how frequently. Nor do they know who owns that data, given the fuzzy boundaries of data ownership.²¹ Given that Big Data feeds algorithms, data ownership is a false problem: once shared in the data flow, data contributes to CHR construction. Even if you withdraw your data, it has already helped create the automated meanings and categories applied in interactions with institutions using the data. What people lack is access to knowledge about the processes through which data is assembled, circulated, and organized. Privacy concerns arise from the perceived loss of control over one’s digital life. Yet privacy alone does not prevent the generation of automated CHR. First, it is virtually impossible not to generate data in the digital era. Every banking transaction, phone movement, and purchase generates data. Second, even if you withdraw your data from databases—an increasingly difficult task—your data may already have contributed to the algorithms that shape future interactions with institutions. Concealing your salary, for example, will not necessarily prevent you from being classified as a “middle-class Italian academic woman,” a category that may influence insurance premiums or loan costs. That label is generated from billions of aggregated data points. Consequently, even without sharing your data, the system will know you as soon as you interact with it.

In Automating Inequality (2018), Virginia Eubanks offers a critical analysis of data-driven technologies and automated decision-making systems designed to improve public services, showing how they often reinforce existing social inequalities.²² Eubanks develops her argument through three case studies. First, in Indiana, an automated welfare eligibility system led to widespread benefit denials and increased hardship. In Los Angeles, the Homeless Management Information System, intended to streamline service provision, contributed to intensified surveillance and punitive measures. Third, in Allegheny County, Pennsylvania, predictive analytics assessing child abuse risk disproportionately targeting low-income families, raising privacy and discrimination concerns. Eubanks highlights how automated systems often neglect the human element, leading to dehumanization that exacerbates the struggles of already marginalized individuals. These cases illustrate the potential injustices produced by artificial CHR, which do not depend primarily on breaches of privacy but on processes of profiling made possible through data collection and sharing. Such profiling frequently reproduces preexisting social biases, reinforcing unjust CHR that reflect and perpetuate unacceptable visions of society.

The third issue is negotiation. We live within a dynamic flow of data characterized by complex and often uncontrollable structures of organization and hierarchy. Big Data is commonly defined as “high-volume, -velocity, and -variety information assets that demand cost-effective, innovative forms of information processing for enhanced insight and decision making.” According to IBM, businesses generate nearly 2.5 quintillion bytes of data each day. The analysis of Big Data is a lengthy and complex process, further complicated by the volume, velocity, and variability of contemporary data streams. Real-time data generated by Internet of Things (IoT) devices, sensors, and machines is produced at such scale and speed that meaningful analysis becomes increasingly difficult.

Companies undertake extensive processes to analyze data, classify it, assign meaning to it, and build predictive models for commercial applications. This process of data-driven sense-making contributes to the production of CHR outside the direct control of those whom the data describes. The resulting inability to negotiate the meanings generated from data may constitute a new form of epistemic injustice produced by AI. Humans are deprived of their socio-cognitive capacity to participate in the production of CHR, and individual participation in the generation of shared meanings becomes negligible.

Is There a Remedy?

How can epistemology help remedy this new form of epistemic injustice—the deprivation of human participation in the production of CHR? This is a crucial issue for future philosophical debate. It is difficult to imagine a human society without the means to produce its own meanings and narratives. For many scholars and commentators, a solution appears impossible as long as the production of CHR remains controlled by a small number of companies that dominate the data and AI industries. No single agent has the political power to stop the AI market, and doing so would mean sacrificing the advantages associated with technological innovation.

A possible solution is the development of new epistemic competencies capable of evaluating automated CHR, identifying algorithmically generated biases, and constructing a meta-CHR through which automated meanings and classifications can be critically assessed. The creation of a structured meta-CHR for interpreting AI-generated CHR involves several steps requiring human collaboration and negotiation:

1. Definition and Scope. Clearly define the purpose and scope of the meta-resource. Specify the types of AI-generated CHR it is intended to interpret and the objectives it seeks to achieve, such as promoting epistemic justice, reducing inequality, or identifying gender and racial biases.
2. Framework Development. Develop a framework outlining the key components of interpretation and evaluation, including methods for analyzing AI-generated content, criteria for assessment, and guidelines for interpretation. Define key terms and concepts related to both collective hermeneutics and AI, ensuring a common understanding.
3. Collaborative Platform. Establish a platform through which users can submit AI-generated CHR for interpretation and analysis. Implement features (like AI metadata) for user discussions, insights, and feedback on interpretations.
4. Integration of AI Tools. Integrate AI-based analytical tools to assist in the interpretation of AI-generated content. Implement natural language processing and sentiment analysis tools to understand underlying meanings and emotions in CHR.
5. Peer Review Mechanism. Implement a peer review mechanism for expert-vetted interpretations, ensuring accuracy and reliability. Encourage collaboration and constructive feedback among users to enhance interpretation quality.
6. Ethical Considerations. Establish ethical guidelines for the generation and interpretation of AI-generated CHR. Such guidelines should address the potential biases and limitations of AI systems and provide mechanisms for identifying, evaluating, and mitigating these harms.

By following these steps and incorporating the principles of collaboration, transparency, and continuous improvement, a structured meta-CHR for interpreting AI-generated CHR could provide valuable human insight and help maintain human control over the production and negotiation of CHR.

Conclusion

This paper has examined a new form of epistemic injustice arising from the machine-controlled production and evolution of collective hermeneutical resources. This injustice deprives human beings of one of their most fundamental cultural and political activities: the negotiation of meanings and categories through which they make sense of the social world. A world in which CHR are entirely shaped by AI is a world in which our epistemic lives are reduced to that of passive recipients of labels and categories negotiated without our consent and, increasingly, within private corporations. Philosophy and epistemology should be at the forefront of efforts to develop methods for resisting this “freezing” of CHR and the production of meanings. Our humanity is grounded in the capacity to assign meaning to our own experiences and to those of others, to negotiate those meanings within a shared social space, and to understand perspectives different from our own. To remain human in an AI-mediated world requires keeping the production of CHR active, even if at a different level: the level of organizing, interpreting, and governing AI-generated CHR by ensuring public oversight of their development. This is an epistemological as well as a political and institutional effort—one that we must learn to undertake together.

1. Karen Frost-Arnold, Who Should We Be Online? A Social Epistemology for the Internet (New York: Oxford University Press, 2023), 4.
2. Sara Haarman, “The Data Should Not Speak for Itself: Epistemic Injustice and Data as Rhetoric,” Philosophical Studies in Education 52 (2021): 33.
3. José Medina, The Epistemology of Resistance: Gender and Racial Oppression, Epistemic Injustice, and the Social Imagination (New York: Oxford University Press, 2013).
4. Heather Stewart, Erica Cichocki, and Carolyn McLeod, “A Perfect Storm for Epistemic Injustice: Algorithmic Targeting and Sorting on Social Media,” Feminist Philosophy Quarterly 8 (3/4). Article 8.
5. Miranda Fricker, Epistemic Injustice: Power and the Ethics of Knowing (Oxford: Oxford University Press, 2007).
6. Gaile Pohlhaus Jr., “Relational Knowing and Epistemic Injustice: Toward a Theory of ‘Willful Hermeneutical Ignorance,’” Hypatia 27, no. 4 (2012): 715–35.
7. Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor (New York: St. Martin’s Press, 2018); Gloria Origgi and Serena Ciranna, “Epistemic Injustice: The Case of Digital Environments,” in The Routledge Handbook of Epistemic Injustice, ed. Ian James Kidd, José Medina, and Gaile Pohlhaus Jr. (London: Routledge, 2017); Salvatore C. Scotto, “Digital Identities and Epistemic Injustices,” Humana.Mente Journal of Philosophical Studies 13, no. 37 (2020); Gordon Hull, “Dirty Data Labeled Dirt Cheap: Epistemic Injustice in Machine Learning Systems,” SSRN Scholarly Paper 4137697 (2022), https://doi.org/10.2139/ssrn.4137697.
8. James Symons and Rafael Alvarado, “Epistemic Injustice and Data Science Technologies,” Synthese 200, no. 2 (2022): 1–26.
9. Fricker, Epistemic Injustice, 99–100.
10. Fricker, Epistemic Injustice, 155.
11. Rebecca Mason, “Two Kinds of Unknowing,” Hypatia 26, no. 2 (2011): 294–307; Kristie Dotson, “Conceptualizing Epistemic Oppression,” Social Epistemology 28, no. 2 (2014): 115–38; Medina, Epistemology of Resistance.).
12. Charles W. Mills, “White Ignorance,” in Race and Epistemologies of Ignorance, ed. Shannon Sullivan and Nancy Tuana (Albany: State University of New York Press, 2007), 13–38.
13. Ian Hacking, The Social Construction of What? (Cambridge, MA: Harvard University Press, 1999).
14. Leaf Van Boven, “Pluralistic Ignorance and Political Correctness: The Case of Affirmative Action,” Political Psychology 21, no. 2 (2000): 267–76.
16. Ian Hacking, “Lost in the Forest,” London Review of Books 35, no. 15 (August 8, 2013): 7–8.[/efn_nite] Yet these classifications are continually challenged by patients’ experiences and ongoing debates concerning the authority of the DSM itself. Produced by the American Psychiatric Association, the DSM reflects the biases and culture of that community. Nevertheless, it has become the international standard for mental health diagnosis, shaping how patients understand themselves and make sense of their experiences, even though it is not issued by an international body such as the World Health Organization. Are DSM classifications part of CHR? They are embedded within institutions—including insurance companies, schools, and healthcare systems—as criteria for distributing or denying benefits. However, they do not always correspond to patients’ lived experiences, although the existence of these labels shapes how patients behave and perceive their illness. Collective hermeneutical resources are the locus of hermeneutical struggles to establish what Michel Foucault called an “order of things.”¹⁵Michel Foucault, The Order of Things: An Archaeology of the Human Sciences (Paris: Gallimard, 1966).
17. Miranda Fricker, “Epistemic Injustice and the Preservation of Ignorance,” in The Epistemic Dimensions of Ignorance, ed. Rik Peels and Martijn Blaauw (Cambridge: Cambridge University Press, 2016).
18. Tristan S. Goetze, “Hermeneutical Dissent and the Species of Hermeneutical Injustice,” Hypatia 33, no. 1 (2018), 14.
19. Goetze, “Hermeneutical Dissent,” 16–17.
20. A. Pismenny, G. Eickers, and J. Prinz, “Emotional Injustice,” Ergo: An Open Access Journal of Philosophy 11 (2024).
21. Wendy Lipworth, Paul Mason, Ian Kerridge, and John P. A. Ioannidis, “Ethics and Epistemology in Big Data Research,” Journal of Bioethical Inquiry 14 (2017).
22. Eubanks, Automating Inequality.