From Human-Centered AI to Culturally Situated AI

1. Introduction

Human-centered artificial intelligence (HCAI), human-AI interaction (HAI), and hybrid human-artificial intelligence (HHAI) have emerged as necessary responses to technology-centered definitions of AI. Their central thesis is difficult to dispute: AI systems should support, augment, and collaborate with people rather than replace or subordinate them. This premise has shaped design guidelines, evaluation criteria, and research agendas around trust, transparency, user control, accountability, explainability, and human oversight [1,2,3,4]. In this sense, HCAI has successfully treated an important imbalance in AI research by moving attention away from algorithms alone and toward the social contexts in which AI systems are used. Yet, the current framing of human-centered AI remains too narrow for the generative AI moment, since its dominant assumptions were largely formed around AI systems that classify, recommend, predict, optimize, or assist decision-making. In those settings, the human is commonly imagined as a user, supervisor, validator, evaluator, or stakeholder. While these roles still matter, Generative AI (GenAI) systems is now able to produce diverse content, including essays, images, scripts, music, explanations, code, learning materials, design concepts, and public-facing narratives. Their outputs are not merely information products or decision aids, but cultural artifacts: they participate in systems of meaning, representation, authorship, genre, taste, credibility, and social authority. In this paper, we argue that human-centered AI needs a cultural extension. The point is not simply that AI needs the humanities, or that humanities scholars should be invited to comment on AI ethics; that link is now well established and, on its own, no longer constitutes a strong position. The more specific claim advanced here is that HAI and HHAI are still shaped by a predominantly cognitive model of intelligence, focusing on how humans and AI share cognitive work: perceiving, deciding, explaining, recommending, predicting, and solving problems. Generative AI, however, increasingly shares cultural work: producing meanings, representing social worlds, shaping categories, imitating voices, framing evidence, and redistributing authorship.

We, therefore, propose culturally situated AI as a conceptual layer for HCAI and HHAI. Culturally situated AI treats AI systems not only as tools used by people, but as socio-technical systems embedded in cultural practices. As a result, it asks not only whether a system is accurate, useful, transparent, or fair, but also what forms of meaning it creates, which cultural assumptions it normalizes, whose perspectives it amplifies, and how it reshapes relations between creators, audiences, institutions, and knowledge. In this paper, we build on concepts from HCI, hybrid intelligence, media theory, cultural analytics, and recent AI research on cultural representation and alignment to argue for a more precise (and meaningful) research agenda.

2. Human-Centered AI is mostly Cognitive

Human-centered AI has been highly productive because it makes the human consequences of AI design explicit. Shneiderman [4], for example, frames HCAI around systems that are reliable, safe, and trustworthy, with high levels of human control and high levels of automation. Similarly, Amershi et al. [2] formulate design guidelines for human-AI interaction that address user expectations, uncertainty, feedback, correction, and appropriate reliance. Recent work on transparency in Large Language Models likewise emphasizes that transparency should support appropriate human understanding for different stakeholders and contexts [3]. Hybrid intelligence research has extended this logic by defining intelligence as a combination of human and machine capabilities, with the aim of augmenting human intellect, rather than replacing it [1]. This background is essential, but identifies a structural limitation in its dominant thesis: even when HCAI and HHAI are socially aware, they often remain cognitively centered, meaning that the core interaction is usually imagined as a relation between a human agent and a computational system around a task. The system produces a recommendation, classification, explanation, prediction, or generated output, and the human interprets, accepts, rejects, corrects, or acts upon it. The main question becomes whether the interaction produces appropriate understanding, calibrated trust, better decisions, or improved task performance. This framing works reasonably well for many decision-support systems, but is less adequate for AI systems that generate and circulate cultural material. When a language model drafts a policy text, produces feedback on a student essay, summarizes a social controversy, writes a poem, suggests a public communication strategy, or generates an image of a cultural event, the output cannot be evaluated only as a task artifact; it also has cultural meaning. It may reproduce dominant assumptions, erase minority perspectives, flatten historical complexity, imitate styles, encode stereotypes, or present contested interpretations as neutral summaries. The problem is not that HCAI ignores ethics or society. On the contrary, much of the field is explicitly concerned with responsible design. The problem is that culture is often treated as context surrounding the interaction, rather than as constitutive of what the system does. The human is centered as a user, but not always as an interpreter situated within language, history, institutions, communities, and symbolic practices. This distinction matters because generative AI is not only used within culture; it acts upon culture by reorganizing the production and circulation of symbolic forms.

3. Why Generative AI Should Be Treated as a Cultural System

The claim that AI is a cultural system rests on three observations. First, GenAI produces symbolic artifacts at scale. Second, these artifacts circulate through existing institutions and media ecologies. Third, their meaning is not contained in the output itself, but emerges through interpretation. Media theory has long argued that media do not simply transmit messages, but also shape the conditions under which social reality is constructed. Couldry and Hepp [5] describe contemporary social reality as mediated through infrastructures, practices, and institutions. From this perspective, AI systems are not neutral channels through which information passes. They are increasingly part of the infrastructure through which people encounter, produce, and authorize knowledge. Search engines, recommender systems, chatbots, text-to-image generators, and AI writing assistants influence what becomes visible, plausible, memorable, and legitimate. This is also why cultural analytics is relevant. Manovich [6] argues that computational methods have become central to the analysis of large-scale cultural data, particularly visual media. Yet the relationship between AI and culture is not one-directional: AI is used to analyze culture, but it is also trained on cultural artifacts and returns new artifacts into cultural circulation. van Noord et al. [7] describe this as a bidirectional relation: AI models are used to interpret culture, while also embedding cultural expressions, assumptions, and limitations from the data on which they are trained. This bidirectionality is central to the argument of culturally situated AI. GenAI intensifies this relation because it does not merely classify or retrieve existing cultural objects, but produces new ones. Epstein et al. [8] argue that GenAI is changing creative work and the media ecosystem, especially by raising questions about authorship, ownership, and the relation between human and machine creativity. These questions cannot be answered through technical evaluation alone. A generated image or text may be technically impressive and still culturally shallow. It may be coherent but derivative, inclusive in surface features but stereotypical in framing, fluent but epistemically weak, or persuasive while erasing uncertainty. Recent research on cultural representation in AI makes this issue explicit. Prabhakaran et al. [9] argue that AI systems often rely on assumptions about human behavior without sufficiently accounting for the cultural contexts, values, and practices that shape such behavior. Because AI development and training data are concentrated in particular countries, AI systems may embed and export culturally specific assumptions as if they were universal. Tao et al. [10] provide empirical evidence that several Large Language Models express cultural values resembling English-speaking and Protestant European countries, while cultural prompting can reduce but not eliminate cultural bias. Qadri et al. [11] go further by arguing that evaluations of AI-generated cultural representation often rely on thin, reductive measures that neglect how communities interpret and assign meaning to their own representation. Taken together, this literature supports a shift in perspective. Generative AI should be treated as a cultural system because it participates in cultural production, relies on cultural data, mediates cultural interpretation, and shapes the social visibility of meanings and identities. The central issue is, therefore, not only whether AI systems are aligned with human values in the abstract, but how they operate within situated, contested, and unequal cultural worlds. This argument also connects to earlier work that frames cultural bias in GenAI as a problem of distorted representation, using Plato’s Allegory of the Cave to show how AI systems may reproduce partial cultural realities as if they were complete ones [12].

4. What the Humanities Add Beyond Ethics

The humanities are often introduced into AI debates through ethics: bias, fairness, accountability, dignity, autonomy, and harm. While these are important concerns, treating the humanities only as an ethical correction underestimates their relevance. The humanities also provide methods for analyzing meaning, representation, interpretation, narrative, historical context, authorship, power, and cultural memory. This is particularly important for digital humanities, where AI can both extend interpretive practice and reproduce colonial patterns in data, language, visibility, and epistemic authority [13]. These methods are directly relevant to GenAI because it operates through language, images, symbols, genres, and conventions. A basic insight from cultural theory is that representation is not a mirror of reality. It is a practice through which meaning is produced. Hall et al. [14] argues that meaning is constructed through systems of representation, language, and shared codes. This matters for AI because generated outputs do not simply depict people, cultures, or events; they frame them, by selecting categories, stabilizing associations, reproducing genres, and suggesting interpretations. As a result, the question is not only whether an output is accurate, but how it represents a social world and under what assumptions. This also connects to older HCI and science and technology studies. Suchman [15] challenged the idea that human-machine interaction can be fully understood through plans, scripts, or formal models detached from situated action. Dourish [16] similarly argued that interaction is embodied and embedded in social practice. Culturally situated AI extends this line of thinking to generative systems, by arguing that interaction with AI is not simply a user-system exchange, but a culturally mediated event. The same AI-generated text may function differently in a classroom, a newsroom, a public administration office, a museum, or a political campaign. Meaning depends on setting, audience, institution, history, and stakes. The humanities also help clarify the limits of purely metricized evaluation. AI evaluation often privileges what can be counted: accuracy, toxicity scores, benchmark performance, demographic parity, calibration, or user satisfaction. These measures are useful but incomplete: cultural meaning frequently depends on nuance, implication, memory, irony, genre, omission, and context. Qadri et al. [11] describe this as the difference between thin and thick evaluation: thin evaluation checks observable or simplified features, while thick evaluation asks how communities interpret representation within their own social worlds. This is not an argument against measurement, but an argument against pretending that measurement is sufficient when the object being measured is interpretive.

The humanities, therefore, add three things that are especially relevant to HAI and HHAI. First, they provide theories of meaning that prevent AI outputs from being treated as self-contained artifacts. Second, they provide methods of interpretation, such as close reading, discourse analysis, historical contextualization, and critique of representation. Third, they provide a vocabulary for cultural agency, authorship, and responsibility. These contributions are methodological, not merely decorative. They change what counts as a good AI system.

5. Toward Culturally Situated AI

Culturally situated AI is proposed as a conceptual extension of human-centered AI. It does not replace HCAI, HAI, or HHAI; instead, it adds a cultural layer that becomes especially important for generative and interactive AI systems. This layer can be summarized through three shifts. The first shift is from the individual user to the interpretive community. HCAI often asks whether an individual user understands, trusts, controls, or benefits from an AI system. Culturally situated AI asks how groups, communities, and institutions interpret the system and its outputs. This is particularly important when AI systems generate representations of culture, identity, history, or social life. A prompt such as “generate an image of a traditional wedding” or “summarize the causes of a political conflict” does not have a culturally neutral answer: it invokes histories, stereotypes, rituals, absences, and contested narratives. Evaluation should, therefore, involve interpretive communities, not only generic users or crowdworkers.

The second shift is from transparency to interpretive transparency. Standard transparency asks what the model can do, how it behaves, what data it uses, and how uncertain its outputs are [3]. These are necessary questions. Interpretive transparency adds another set: what cultural assumptions does the system appear to rely on? Which categories does it stabilize? What alternative interpretations are missing? How does the interface encourage users to treat outputs: as suggestions, summaries, evidence, creative drafts, or authoritative accounts? This is especially important because generative AI often produces fluent outputs that conceal uncertainty, contestability, and representational choice.

The third shift is from value alignment to cultural situatedness. Value alignment is usually framed as the problem of ensuring that AI systems act according to human values [17]. However, values are not static objects that can be fully specified in advance. They are interpreted, negotiated, and contested within cultural contexts. Culturally situated AI does not abandon alignment; it makes alignment less abstract by asking whose values, expressed in which language, under which institutional conditions, and interpreted by which communities. Recent work on cultural alignment in Large Language Models shows that cultural bias is a structural challenge for globally deployed generative systems [10]. These shifts imply concrete design and evaluation directions, meaning that AI systems that generate cultural content should be evaluated not only for factuality and harmful content, but also for representational depth, interpretive diversity, and cultural specificity. In the same context, participatory design should move beyond requirement gathering and include community-based interpretation of outputs, accommodating plurality by presenting alternatives, uncertainties, and contextual notes rather than a single fluent answer. Documentation should include cultural limitations, not only model capabilities. Educational uses of GenAI should teach students not only to prompt effectively, but to read AI outputs critically as cultural texts, promoting important concepts such as autonomy, ethical judgment, and the changing role of teachers in AI-rich learning environments [18]. This reframing also changes the role of the human in human-AI interaction. The human is not merely a user who gives feedback or a supervisor who corrects errors, but also an interpreter, contextualizer, and co-producer of meaning. The goal of interaction is not only task completion, but critical engagement. A culturally situated system should help users ask better questions about the output: What is being assumed? What is missing? What genre is being imitated? What authority is being claimed? What community might read this differently? These are not secondary questions. In GenAI, they are central to responsible use.

6. Implications for HAI and HHAI

The proposed framing has direct implications for future work in HAI and HHAI. First, it expands the object of evaluation. HAI research has developed strong methods for studying usability, trust, explanation, reliance, and control; culturally situated AI adds cultural impact, representational adequacy, interpretive plurality, and epistemic authority. This does not mean every system requires full cultural analysis, since different cases pose different cultural risks. The point is that systems producing or mediating symbolic content should not be evaluated as if they were culturally neutral tools.

Second, it expands the range of methods: quantitative benchmarks and user studies remain useful, but they should be complemented by qualitative and interpretive methods. Close reading can expose recurring narrative patterns in AI-generated text. Discourse analysis can examine how AI frames social groups or political issues. Historical contextualization can reveal anachronisms or colonial assumptions in generated cultural material. Participatory workshops can identify representational failures that generic metrics would miss. These methods are familiar in the humanities and social sciences, but they remain underused in AI evaluation.

Third, it reframes interdisciplinarity. Many AI projects include humanities or social science perspectives too late, after the system has already been designed. Culturally situated AI requires earlier integration. Humanities scholars should not only comment on harms after deployment; they should help define what the system is for, what counts as an adequate representation, what kinds of ambiguity should be preserved, and what kinds of authority the system should not claim. This is particularly relevant for AI in education, cultural heritage, media, public communication, and policy support, where outputs are deeply entangled with interpretation and public meaning.

Finally, culturally situated AI offers a sharper contribution to the HHAI agenda. Hybrid intelligence should not only describe the combination of human and machine cognition. It should also describe the co-production of cultural meaning between humans, AI systems, institutions, and communities. A hybrid system that helps a teacher generate feedback, a journalist draft an explainer, a museum create visitor narratives, or a public agency communicate policy is not only augmenting cognition. It is shaping culture. HHAI should therefore ask how such systems can support interpretation, plurality, and critical judgment, rather than only efficiency, personalization, and trust.

7. Conclusion

Human-centered AI is necessary but not sufficient. It has helped move AI research away from algorithmic performance alone and toward human needs, values, and contexts. However, the rise of GenAI exposes a gap in its conceptual foundations. Many AI systems are now cultural systems: they generate symbolic artifacts, mediate interpretation, reshape authorship, and influence what societies see as credible, normal, creative, or valuable. This paper has proposed culturally situated AI as a way to name and address that gap. The contribution is not the general claim that AI and the humanities should collaborate: it is the more specific claim that HAI and HHAI require a cultural layer because GenAI participates in cultural production and meaning-making. This layer moves attention from users to interpretive communities, from transparency to interpretive transparency, and from abstract value alignment to culturally situated evaluation. The humanities matter here not as a ceremonial ethical add-on, but as methodological infrastructure, helping AI research understand representation, context, narrative, authorship, ambiguity, and cultural memory. Without this layer, human-centered AI risks becoming technically sophisticated but culturally naive. With it, HAI and HHAI can better address the systems now emerging around us: systems that do not merely assist human thought, but increasingly participate in the making of culture.