Paper deep dive
Supporting and Performing Culture from the Inside
Lea Frermann, Steven Bird
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Anthropology and related disciplines which study culture have often found it useful to consider their epistemological and methodological approaches in terms of emic versus etic. This is the distinction between the insider perspective, how the world is viewed by a member of the culture, versus the outsider perspective, the scientific cataloguing of cultural knowledge and practice. We adopt this framework to systematically analyse assumptions in recent research on culture in NLP along the full pipeline of task selection, data collection, system design and evaluation. We draw attention to the `culture of NLP' which shapes the research focus and approaches of our field, and suggest pathways towards a more culturally attuned AI.
Tags
Links
- Source: https://arxiv.org/abs/2609.32281v1
- Canonical: https://arxiv.org/abs/2609.32281v1
Trouble viewing inline? Open PDF directly â
Full Text
39,751 characters extracted from source content.
Expand or collapse full text
Supporting and Performing Culture from the Inside Lea Frermann Affiliation: The University of Melbourne, Melbourne, Australia Email: lea.frermann@uni-tuebingen.desteven.bird@cdu.edu.au Steven Bird Affiliation: Charles Darwin University, Darwin, Australia The University of TĂŒbingen TĂŒbingen Germany Abstract Anthropology and related disciplines which study culture have often found it useful to consider their epistemological and methodological approaches in terms of emic versus etic. This is the distinction between the insider perspective, how the world is viewed by a member of the culture, versus the outsider perspective, the scientific cataloguing of cultural knowledge and practice. We adopt this framework to systematically analyse assumptions in recent research on culture in NLP along the full pipeline of task selection, data collection, system design and evaluation. We draw attention to the âculture of NLPâ which shapes the research focus and approaches of our field, and suggest pathways towards a more culturally attuned AI. 1 Introduction Human culture can hardly be disentangled from the culture that a researcher brings to an investigation, a phenomenon that has been called Kulturbrille (cultural glasses) by Franz Boas (Engelke, 2018). The same holds true for the flourishing research on culture and cultural alignment in natural language processing (NLP) Pawar et al. (2025). The culture of NLP is so deeply embedded that it can be hard to see (Hutchinson, 2024). Upstream of the data that flows into NLP systems, culture is the invisible filter that selects, shapes and colours what is observed, the meaning that is assigned, and the text or speech that is ultimately produced before it is ingested by an LLM. Downstream of the language technologies we create â like information retrieval (IR), machine translation (MT), automatic speech recognition (ASR), and most recently, chatbots and large language models (LLMs) â are all the diverse ways that information and technology are socialised in the worldâs cultures. Between the upstream and the downstream culture that is invisible to NLP, there is the culture embedded in the representational choices of LLMs and in the interactional style of chatbots. This NLP culture includes evaluations of the capability of NLP artefacts like LLMs and chatbots to reproduce culture, including cultural facts but also deeper notions like cultural sensitivity and cultural competence (Nejadgholi et al., 2026), or situational appropriateness and values alignment (Minggad et al., 2026). NLP culture is present wherever we limit our conception of culture to information that is documented in textual or spoken records, separated from the paralinguistic and the extralinguistic, communicative layers that arise from the embodied nature of communication. This includes evaluations of the capability of LLMs and chatbots to reproduce cultural facts, but goes further to evaluations of cultural sensitivity and cultural competence (Nejadgholi et al., 2026), or situational appropriateness and values alignment (Minggad et al., 2026). Figure 1: Hallâs Cultural Iceberg: forms and behaviours (etic) can be observed by outsiders, while corresponding meanings and beliefs (emic) are only accessible to insiders (Hall, 1976, after) . To systematise epistemological and methodological positions, anthropologists have proposed to distinguish emic versus etic approaches to cultural studies Pike (1967); Sullivan (2009). An etic perspective adopts the outsiderâs view, classifies surface forms, and applies generic measures across cultures in a universal, comparative, and scalable view. The emic perspective adopts the insiderâs view of a member of the culture, with a focus on the underlying cultural system, specific to a single culture. The concepts of etic and emic originate in structuralist linguistics, as cultural generalisations of phonetic and phonemic. Pike (1967) observed that human behaviour can be observed, described, and classified from outside a culture (etic), while having a significance accessible only to insiders (Headland et al., 1990, emic,). A popular analogy for this distinction is the âcultural icebergâ (Fig. 1). The goals of this paper are to challenge seemingly inevitable methods of research and system development in NLP, to put forth an alternative perspective, and to encourage the community to reflect on the culture of NLP and how it impacts our communityâs goals and contributions. This paper contrasts emic with etic perspectives across the Four Stages of AI Development, and on this basis surveys the recent literature on cultural NLP, identifying a dominance of etic approaches. It then applies the emic/etic framework and (1) uncovers hidden assumptions in how our community designs approaches, data sets and model evaluations; (2) points to gaps in research assuming that the ultimate goal is to develop systems which are valuable for the culture in question; and (3) sketches out a path forward towards an emic research agenda. 2 The Etic/Emic Framework In this section we will see that the practice of generative AI is predominantly etic, in all stages of development. Task definitions are understood as general purpose solutions to externally-defined problems, which means that few researchers see any need to engage individual communities about what they need (Sec. 2.1). Primary data about the world is taken to be text, yet this contains many layers of culture, and may effectively be a blind spot for NLP researchers (Sec. 2.2). Systems are designed to replicate or replace humans without due consideration for how this may impact the meaningfulness of lives and livelihoods on the ground (Sec. 2.3). Finally, deployments are optimised to serve global measures of efficiency, productivity, and profitability, and performance benchmarks are applied across the board, regardless of local cultural norms and expectations (Sec. 2.4). In addition, we reconsider these AI development stages from an emic perspective. We incorporate qualitative findings from a review of all papers published in the Findings or Main streams of *ACL and EMNLP conferences (2022â4/2026) containing the word âcultureâ or âculturalâ in the title (N=107). See Appendix A for details of our approach and a link to a GitHub repository containing the full list of papers and labels. 2.1 Tasks In the context of a minority culture, an etic approach usually presumes the presence of the dominant culture, and looks across cultures, seeking opportunities for universal solutions. The most obvious need, at this level, is âovercoming language barriersâ (Pastor et al., 2017). The default task, then, would be translation. We can observe this by submitting a queries to a chatbot in progressively smaller languages and â beyond a certain threshold â notice that the chatbot shifts from answering queries in monolingual mode, to translating queries into the dominant language, usually English. In all such translation processes, minority culture content is shaped along the lines of dominant cultures (Minggad et al., 2026). If the default etic position is to overcome barriers, the default emic position is often to maintain barriers, defending cultural diversity in the face of efforts that seek to flatten it. Community leaders enact their self-determination through intergenerational transmission of cultural practices, strengthening and safeguarding the local lifeworld (Kimmerer, 2013; Harjo, 2019). How this manifests on the ground changes from place to place, but a common linguistic challenge is to transmit ecological knowledge. For example, parents teach their children about food preparation, duties to kin, and the associated knowledge of the country and its seasons. These knowledge practices are increasingly video recorded and shared, to the point where individuals and communities may have sizeable collections of untranscribed audio and video in their local vernacular. This rarely involves universal information access mediated through local language or culture. Those members are usually already bicultural and bilingual owing to the history of contact. This means that they are perfectly capable of navigating the dominant culture using the dominant language. Such is the nature of bilingualism and language shift Fishman (2001); Grosjean (2021). When considering NLP tasks from an emic perspective, we do not seek to âunderstandâ or âtranslateâ or âalignâ, but allow space for diverse local meanings that fall outside the scope of our systems. For example, language technologies can support navigation of untranscribed audio and video through spoken term retrieval, enabling human learning of language and culture directly from nontextual sources (Bird, 2026). Insider-driven considerations like the above are barely mentioned in the NLP literature. The majority of surveyed papers (N=78; 73%) focussed on the intrinsic goal of assessing or augmenting cultural knowledge in LLMs, or so-called âtrivia-centeredâ cultural knowledge representations Zhou et al. (2025); Oh et al. (2025); Alkhamissi et al. (2026). The remaining papers address downstream tasks ranging from culturally appropriate dialogue over culturally attuned translation to culture-aware IR. However, the tasks are selected and formalised by the researchers who are embedded in a dominant culture (by upbringing and/or by academic affiliation), typically with the primary goal of generalizability to many cultures in mind (etic). We did not encounter papers which co-developed a task with a cultural community. 2.2 Data For most of the worldâs cultures, the textual record comes from ethnographic observations that are subject to the cultural conventions of outsiders (Clifford and Marcus, 2023). The record depends on the observable behaviours of insiders, and what insiders may report about the underlying meanings and beliefs, reports which may be unreliable or even deceptive (Borofsky, 2005). For example, Said argues that everything written about the âOrientâ is a projection of western culture Said (1978). This problem is not removed by eliminating outsider texts; there is the culture of writing itself: ⊠writing, wherever it exists, is always only one of several communication channels available to the members of a society. Consequently, the conditions under which it is selected and the purposes to which it is put must be described in relation to those of other channels. (Basso, 1974, p426) There is also the nature of observation, e.g.: there are no pure facts (Popper, 1935); differentiating figure from ground is a perceptual act (Schacter et al., 2009); measurement involves culturally-informed choices (Walter and Andersen, 2013); and âthe world as it isâ differs from âthe world according to dataâ (Mitchell and Fleischman, 2020, p145). Thus texts are separated from ârealityâ by three cultural layers: what is observed; what meaning is attributed; and how this is rendered into text. One response to this situation is to consult with members of the minority culture in the data collection process rather than relying exclusively on existing texts as ground truth. A notable fraction of the surveyed literature (19 papers) drew on community input in concept selection, example creation and/or data validation, where data was typically contributed by members of the target culture which were based at the respective research institution, or via crowd sourcing. However, we also observe a strong trend of creating âculturalâ benchmarks in an at least partially synthetic way (33 papers), where often concepts and/or context are first generated by LLMs and later filtered by humans in a practice that necessarily distorts the emic, intra-cultural perspective (see discussion above). The practice of translating English benchmarks was widely criticised and occurred only three times. 31 papers based their concept samples or examples on existing data sets which often spanned many languages (news, social media, Wikipedia, âŠ) and imply the concerns laid out above. 2.3 System design The vast majority of language technologies is designed to enhance human information access, through IR, MT, ASR, or LLMs in the form of chatbots. The construct of chatbots was founded on the model of a human amateur consulting a human expert (Weizenbaum, 1966; Woods, 1973). Studies of the culture of expertise show that it is not to be understood in terms of information possessed and imparted, but in terms of interaction and ideology. For each domain there is a culture whereby experts are socialised, evaluated, and authorised, and there are particular linguistic performances that produce jargon, structure interactions, and animate evidence (Carr, 2010). Amateur-expert consultations have a shape which is particular to the domain, e.g. in medicine with opening, complaint, examination/test, diagnosis, treatment/advice, and closing (Ten Have, 1989). Across many domains, better outcomes are linked to interpersonal rapport (Wilson et al., 2010). Set against this, we have the agenda of making chatbots into domain experts (Metzler et al., 2021), with the diversion of web queries into chatbot interactions, in the name of retrieving answers rather than documents. However, to make this shift is to âmiss the big picture of why people seek information and how that process contains value beyond simply retrieving relevant informationâ (Shah and Bender, 2022, p221). Moreover, it risks the replacement of diverse cultures of expert-shaped learning interactions with a series of prompt-response dyads shaped by novices. Generative AI is increasingly deployed in agentic systems, designed with the goal of simulating and eventually replacing human experts, from software engineers to academic reviewers. This reflects the culture of efficiency and automation underlying the dominant western view Birhane et al. (2022). An alternative take is to design AI agents to be augmentative Bird (2024), to enhance human expertise rather than seeking to replicate it. This approach bodes better with minority culturesâ matters of concern (Section 2.3), and may well lead to more socially sustainable AI development more broadly as we discuss in Section 3. Most of our surveyed papers (73%) focussed on intrinsic evaluation and did not consider a âdownstream systemâ (cf. Sec. 2.1). Those which were more task oriented did not question the design of a potential system, and adopted the predictable designs for IR, classification, or dialogue as established in the culture of NLP. Task Data System Design Deployment / Evaluation Etic generic solns to external problems data extraction for LLMs simulate or improve on humans efficiency and scalability âł generate facts about any culture âł translated or synthesized common sense benchmarks âł text chat; adapting to user language; flexibility in tone/register âł correctness and completeness of responses Emic local agency in shaping their world, identifying problems, exploring solutions elicit cultural concepts and logics build / augment human capacity in sustaining cultural vitality local decision making, maintaining culture and locally meaningful livelihoods âł understanding western concepts which have no translation (e.g. âriskâ) âł exegesis of western cultural concepts into local languages âł redesign of text interaction to follow local teacher-led pattern âł equipped with actionable information, people make better-informed choices âł transmitting knowledge practices (e.g. food collection/preparation) âł videos of people demonstrating knowledge practices, sharing lore âł collection management and navigation, spoken term retrieval âł measure user-reported utility and safety Table 1: Etic vs Emic Perspectives on Four Stages of AI Development 2.4 Deployment and evaluation How can we evaluate the success of technology developed from an emic perspective? Typical NLP validation paradigms involve benchmarks, baselines, and scores that can be compared universally, e.g., across methods, languages, and cultures. In keeping with the understanding of emic as local and culture-specific, solutions will vary from place to place. However, emic level evaluation will always center culturally-meaningful ends, including cultural survival, intergenerational knowledge transmission, communal thriving and individual wellbeing (Nigatu et al., 2024, cf.). It will be contextualised in concrete application cases and user groups; and it will involve ongoing evaluation monitoring changes in user needs. Of the minority of task-oriented papers in our survey, none deployed a system in a community and attempted an evaluation in an ecologically valid environment. Validation was primarily based on benchmarks with automated scoring, occasionally augmented with a user study. However, human studies typically took the form of comparison of culturally attuned system A with baseline system B on researcher-defined variables, and were thus constrained in focus and context, and insensitive to user considerations beyond the predefined variables. 3 Paths Forward How can we navigate into the emic? One option is to reproduce the four stages of AI development (Sec. 2) within a local culture through monocultural deployments. Another is to reproduce mainstream methods under local control and idiomatic interaction designs. A third begins from a relational understanding of intelligence. Local adaptation. Rather than seeking to contribute cultural data to shape mainstream LLMs, this is a space of deploying language models within local cultures (Mahelona et al., 2023; Biju and Gayathri, 2025; Raj et al., 2025, cf.). Harms that come from outsider views and from homogenisation towards dominant norms (Peterson, 2025, cf.) are mitigated by restricting training data to insider sources. This remains etic because sources are textual, and they have the limitations described in Section 2.2. However, the benefit of conducting the four stages of AI development within a single cultural area is that the cultural assumptions of users are more likely to align with the cultural tropes that appear in texts provided by cultural insiders. Local control. Here we ask about approaches to the design of AI which put the local communityâs aspirations at the centre (Bird and Yibarbuk, 2024; Cooper et al., 2024; Markl et al., 2024, e.g.). These may start with seemingly mundane tasks such as text entry and literacy (Moshagen et al., 2023; Pine et al., 2025). Under local control, the interaction design of a chatbot would be reshaped along the lines of local knowledge transmission. Perhaps the initiative is held by teachers, not by learners (who have no entitlement to knowledge). We provide some suggestions in Table 1. Relational AI. Navigating into the emic involves diverse understandings of intelligence (Bridle, 2022; Lewis et al., 2024), and on approaches not based on the premise of replicating observable human behaviour (Crawford, 2021; Zuboff, 2022). This work exceeds one-size-fits-all thinking (Rehm and Way, 2023; NLLB Team, 2024, cf.), and essentialisations of culture (Lenard, 2020, cf.). We expect this to lead to parochial and hybrid solutions (Srinivasan, 2017; Bird, 2026, cf.). A promising pathway is to develop human resources: increasing the capacity of western technologists to engage respectfully with members of diverse cultures, and increasing the capacity of local communities to conceive language technologies that serve local agendas. This involves rethinking the four stages of AI development under diverse epistemologies (Kimmerer, 2013; Harjo, 2019, e.g.). 4 Discussion In response to constructive feedback, we address the primary counter-arguments below: (a) The etic/emic distinction is presented as near-categorical rather than a spectrum. This suggests intermediate positions between outsider and insider, and that a model could be iterated from etic to emic, or from forms to meanings, or from behaviours to beliefs. This approach imagines that culture is little more than a set of idioms that can be learnt (Kabra et al., 2023; Attia et al., 2026) or values that can be aligned with (Cao et al., 2024; Kabir et al., 2025). See Zhou et al. (2025) for a critique of this perspective. Alternatively, one might wonder about incrementally shifting some of the Four Stages of AI Development (cf. Tab. 1, columns). However, in an emic approach, these stages are tightly coupled, strongly influenced by the same underlying culture. Emic and etic were originally proposed as complementary standpoints on the nature of descriptions of human behaviour Pike (1967). Much of the work to date is etic rather than emic, and we contend that research on culture in NLP would benefit from a greater emphasis on emic approaches. (b) Systems should embed a âmulti-cultural view of knowledgeâ Conceptually, the emic view by definition considers one culture at a time, rejecting the idea of general cross-cultural structures. Practically, large training data sets are always skewed to majority cultures â quantitatively due to a skew towards majority culture data and/or qualitatively due to short-cut data augmentation methods like translation or generation of additional data. This leads to a dominance of majority culture downstream, compromising the goal of developing technology that is maximally attuned to the needs and goals of the target culture. (c) NLP should develop/improve the technical capabilities of a system, leaving its integration into specific scenarios to others (e.g., HCI). This is an articulation of the ideology that technology arises in a cultural vacuum. In fact, technological artefacts are social constructs, a reality that is explored in the discipline of Science and Technology Studies (STS). For case studies and reflections that relate to linguistic and cultural diversity, we refer the reader to (Srinivasan, 2017). (d) Incorporate author demographics into the literature survey as part of determining whether an outsider or insider perspective is applied. Even assuming we could establish author demographics, there remains the effect that researchers from minority cultures whose work reaches ACL venues have already assimilated the culture of NLP with its largely etic approaches. African initiatives such as Masakhane and Lanfrica focus on compiling public datasets and replicating the results of GenAI from dominant languages. 5 Conclusion We have applied the etic/emic framework to notions of culture that arise in NLP, drawing on a systemic review of the cultural NLP literature. We have described the limits to cultural alignment that are intrinsic to etic approaches, given the layers of culture that cannot be removed from â or that fall outside of â the textual record accessible to LLMs. We have shown that cultural alignment of LLMs is not âsolvedâ through etic means, by manipulating the representations of LLMs or the interactional style of chatbots. The problem is âsociety-completeâ (cf. Resnik, 2025, p898), where society differs from culture to culture. We have suggested ways forward, towards a more emic paradigm for research in NLP. 6 Limitations This paper was shaped by the authorsâ own cultural perspectives. While our literature survey is representative of research related to culture in NLP, we do not claim that it is exhaustive and we might have missed a small number of more emic approaches. Still, our main conclusions that etic approaches dominate NLP and that a more emic angle would benefit the utility and value of cultural NLP work still holds. Acknowledgements We are grateful for the reviewersâ constructive feedback. Frermann is supported by the Australian Research Council Discovery Early Career Research Award (Grant No. DE230100761). Bird is supported by Australian Research Council Discovery Projects (DP210100228, DP240101952). The authors made an equal contribution to the research reported here. References Alkhamissi et al. (2026) M. Alkhamissi, Y. Xiao, B. AlKhamissi, and M. T. Diab Hire your anthropologist! rethinking culture benchmarks through an anthropological lens. In Findings of the Association for Computational Linguistics: EACL 2026, p. 1218â1235. External Links: Document Cited by: §2.1. Attia et al. (2026) M. Attia, A. Muhamed, M. Alkhamissi, T. Solorio, and M. T. Diab Beyond understanding: evaluating the pragmatic gap in LLMsâ cultural processing of figurative language. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics, p. 7238â7265. External Links: Document Cited by: §4. Basso (1974) K. H. Basso The ethnography of writing. In Explorations in the Ethnography of Speaking, R. Bauman and J. Sherzer (Eds.), p. 425â32. Cited by: 1st item. Biju and Gayathri (2025) P. R. Biju and O. Gayathri Indic approach to ethical AI in automated decision making system: implications for social, cultural, and linguistic diversity in native population. AI and Society 40 (8), p. 6645â6670. External Links: Document Cited by: §3. Bird and Yibarbuk (2024) S. Bird and D. Yibarbuk Centering the speech community. In Proceedings of the 18th Conference of the European Association for Computational Linguistics, p. 826â839. External Links: Document Cited by: §3. Bird (2024) S. Bird Must NLP be extractive?. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 14915â14929. External Links: Link, Document Cited by: §2.3. Bird (2026) S. Bird Towards a general theory of linguistic diversity. In Towards Inclusivity and Equality: Language Resources and Technologies for Under-Resourced and Endangered Languages, Proceedings of the SIGUL 2026 Workshop, p. 169â182. Cited by: §2.1, §3. Birhane et al. (2022) A. Birhane, P. Kalluri, D. Card, W. Agnew, R. Dotan, and M. Bao The values encoded in machine learning research. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, p. 173â184. External Links: Document Cited by: §2.3. Borofsky (2005) R. Borofsky Yanomami: the fierce controversy and what we can learn from it. University of California Press. Cited by: §2.2. Bridle (2022) J. Bridle Ways of being: animals, plants, machines: the search for a planetary intelligence. Penguin. Cited by: §3. Cao et al. (2024) Y. Cao, M. Chen, and D. Hershcovich Bridging cultural nuances in dialogue agents through cultural value surveys. In Findings of the Association for Computational Linguistics: EACL 2024, p. 929â945. External Links: Document Cited by: §4. Carr (2010) E. S. Carr Enactments of expertise. Annual Review of Anthropology 39, p. 17â32. External Links: Document Cited by: §2.3. Clifford and Marcus (2023) J. Clifford and G. Marcus Writing culture: the poetics and politics of ethnography. University of California Press. Cited by: §2.2. Cooper et al. (2024) N. Cooper, C. Heldreth, and B. Hutchinson âItâs how you do things that mattersâ: attending to process to better serve indigenous communities with language technologies. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, p. 204â211. External Links: Document Cited by: §3. Crawford (2021) K. Crawford The atlas of AI: power, politics, and the planetary costs of artificial intelligence. Yale University Press. External Links: Document Cited by: §3. Engelke (2018) M. Engelke How to think like an anthropologist. Princeton University Press. Cited by: §1. J. A. Fishman (Ed.) (2001) J. A. Fishman (Ed.) Can threatened languages be saved?: reversing language shift, revisited: a 21st century perspective. Multilingual Matters. Cited by: §2.1. Grosjean (2021) F. Grosjean Life as a bilingual: knowing and using two or more languages. Cambridge University Press. Cited by: §2.1. Hall (1976) E. T. Hall Beyond culture. Garden City. Cited by: Figure 1. Harjo (2019) L. Harjo Spiral to the stars: mvskoke tools of futurity. University of Arizona Press. Cited by: §2.1, §3. T. Headland, K. Pike, and M. Harris (Eds.) (1990) T. Headland, K. Pike, and M. Harris (Eds.) Emics and etics: the insider/outsider debate. Sage Publications. Cited by: §1. Hutchinson (2024) B. Hutchinson Modeling the sacred: considerations when using religious texts in natural language processing. In Findings of the Association for Computational Linguistics: NAACL 2024, p. 1029â1043. External Links: Document Cited by: §1. Kabir et al. (2025) M. Kabir, A. Abrar, and S. Ananiadou Break the checkbox: challenging closed-style evaluations of cultural alignment in LLMs. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 24â51. External Links: Document Cited by: §4. Kabra et al. (2023) A. Kabra, E. Liu, S. Khanuja, A. F. Aji, G. Winata, S. Cahyawijaya, A. Aremu, P. Ogayo, and G. Neubig Multi-lingual and multi-cultural figurative language understanding. In Findings of the Association for Computational Linguistics: ACL 2023, p. 8269â8284. External Links: Document Cited by: §4. Kimmerer (2013) R. W. Kimmerer Braiding sweetgrass: indigenous wisdom, scientific knowledge and the teachings of plants. Penguin. Cited by: §2.1, §3. Lenard (2020) P. T. Lenard Culture. In The Stanford Encyclopedia of Philosophy, E. N. Zalta (Ed.), External Links: Link Cited by: §3. Lewis et al. (2024) J. E. Lewis, H. Whaanga, and C. Yolgörmez Abundant intelligences: placing AI within Indigenous knowledge frameworks. AI and Society 40, p. 2141â2157. External Links: Document Cited by: §3. Mahelona et al. (2023) K. Mahelona, G. Leoni, S. Duncan, and M. Thompson OpenAIâs whisper is another case study in colonisation. Papa Reo. Note: Accessed 20250215 External Links: Link Cited by: §3. Markl et al. (2024) N. Markl, L. Hall-Lew, and C. Lai Language technologies as if people mattered: centering communities in language technology development. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, p. 10085â10099. External Links: Link Cited by: §3. Metzler et al. (2021) D. Metzler, Y. Tay, D. Bahri, and M. Najork Rethinking search: making domain experts out of dilettantes. SIGIR Forum 55 (1), p. 1â27. External Links: Document Cited by: §2.3. Minggad et al. (2026) Q. Minggad, X. Zinan, and Y. Sun MonCulture-eval: a hierarchical benchmark for evaluating Mongolian cultural capabilities of large language models across scripts and regions. In Findings of the Association for Computational Linguistics: ACL 2026, p. 28997â29014. External Links: Document Cited by: §1, §2.1. Mitchell and Fleischman (2020) B. Mitchell and W. Fleischman Responsibility in the age of irresponsible speech. In Societal Challenges in the Smart Society, M. Arias-Oliva, J. PelegrĂn-Borondo, K. Murata, and A. M. L. Palma (Eds.), p. 587â592. Cited by: §2.2. Moshagen et al. (2023) S. N. Moshagen, F. Pirinen, L. Antonsen, B. Gaup, I. Mikkelsen, T. Trosterud, L. Wiechetek, and K. Hiovain-Asikainen The GiellaLT infrastructure: a multilingual infrastructure for rule-based NLP. In Rule-Based Language Technology, A. Hurskainen, K. Koskenniemi, and T. Pirinen (Eds.), p. 70â94. External Links: Link Cited by: §3. Nejadgholi et al. (2026) I. Nejadgholi, M. Kianpour, K. Vishnubhotla, and M. Molamohammadi Defining cultural capabilities for AI evaluation: a taxonomy grounded in intercultural communication theory. In Proceedings of the Workshop on Evaluating Evaluations, p. 161â173. External Links: Document Cited by: §1. Nigatu et al. (2024) H. H. Nigatu, A. L. Tonja, B. Rosman, T. Solorio, and M. Choudhury The Zenoâs paradox of âlow-resourceâ languages. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 17753â17774. External Links: Document Cited by: §2.4. NLLB Team (2024) NLLB Team Scaling neural machine translation to 200 languages. Nature 630, p. 841â846. External Links: Document Cited by: §3. Oh et al. (2025) J. Oh, I. Cha, M. Saxon, H. Lim, S. Bhatt, and A. Oh Culture is everywhere: a call for intentionally cultural evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2025, p. 19156â19168. External Links: Document Cited by: §2.1. Pastor et al. (2017) R. R. Pastor, C. T. QuirĂłs, J. P. V. GarcĂa, T. B. CardĂșs, and M. M. NoguĂ©s Language equality in the digital age: towards a human language project. European Parliament. External Links: Document Cited by: §2.1. Pawar et al. (2025) S. Pawar, J. Park, J. Jin, A. Arora, J. Myung, S. Yadav, F. G. Haznitrama, I. Song, A. Oh, and I. Augenstein Survey of cultural awareness in language models: text and beyond. Computational Linguistics 51 (3), p. 907â1004. External Links: Link, Document Cited by: §1. Peterson (2025) A. J. Peterson AI and the problem of knowledge collapse. AI and Society, p. 1â21. External Links: Document Cited by: §3. Pike (1967) K. L. Pike Etic and emic standpoints for the description of behavior. In Language in relation to a unified theory of the structure of human behavior, K. L. Pike (Ed.), p. 37â72. External Links: Document Cited by: §1, §4. Pine et al. (2025) A. Pine, E. Cooper, D. GuzmĂĄn, E. Joanis, A. Kazantseva, R. Krekoski, R. Kuhn, S. Larkin, P. Littell, D. Lothian, A. Martin, K. Richmond, M. Tessier, C. Valentini-Botinhao, D. Wells, and J. Yamagishi Speech generation for indigenous language education. Computer Speech and Language 90, p. 101723. External Links: Document Cited by: §3. Popper (1935) K. Popper The logic of scientific discovery. Routledge. Cited by: §2.2. Raj et al. (2025) A. Raj, H. Singh, and A. Pachouri Harnessing the potential of AI for Indian agriculture: using âBhashiniâ as a tool to deploy responsible AI and increase the uptake of AI applications among farmers. In Advancing Responsible AI in Public Sector Application, B. Ravindran and A. Singh (Eds.), p. 119â128. Cited by: §3. Rehm and Way (2023) G. Rehm and A. Way European language equality: a strategic agenda for digital language equality. Springer Nature. Cited by: §3. Resnik (2025) P. Resnik Large language models are biased because they are large language models. Computational Linguistics 51, p. 885â906. External Links: Document Cited by: §5. Said (1978) E. W. Said Orientalism. Penguin Random House. Cited by: §2.2. Schacter et al. (2009) D. Schacter, D. Gilbert, and D. Wegner Psychology. Macmillan. Cited by: §2.2. Shah and Bender (2022) C. Shah and E. M. Bender Situating search. In Proceedings of the 2022 Conference on Human Information Interaction and Retrieval, p. 221â232. External Links: Document Cited by: §2.3. Srinivasan (2017) R. Srinivasan Whose global village?: rethinking how technology shapes our world. NYU Press. Cited by: §3, §4. Sullivan (2009) L. Sullivan Emic and etic perspectives. In The SAGE Glossary of the Social and Behavioral Sciences, p. 174â175. External Links: Document Cited by: §1. Ten Have (1989) P. Ten Have The consultation as a genre. In Text and Talk as Social Practice, B. Torode (Ed.), p. 115â135. External Links: Link Cited by: §2.3. Walter and Andersen (2013) M. Walter and C. Andersen Indigenous statistics: a quantitative research methodology. Taylor and Francis. Cited by: §2.2. Weizenbaum (1966) J. Weizenbaum ELIZA: a computer program for the study of natural language communication between man and machine. Communications of the ACM 9, p. 36â45. Cited by: §2.3. Wilson et al. (2010) J. Wilson, R. Ryan, and J. Pugh Professorâstudent rapport scale predicts student outcomes. Teaching of Psychology 37 (4), p. 246â251. External Links: Document Cited by: §2.3. Woods (1973) W. A. Woods Progress in natural language understanding: an application to lunar geology. In Proceedings of the National Computer Conference and Exposition, p. 441â450. Cited by: §2.3. Zhou et al. (2025) N. Zhou, D. Bamman, and I. L. Bleaman Culture is not trivia: sociocultural theory for cultural NLP. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, p. 25869â25886. External Links: Document Cited by: §2.1, §4. Zuboff (2022) S. Zuboff Surveillance capitalism or democracy? The death match of institutional orders and the politics of knowledge in our information civilization. Organization Theory 3 (3), p. 1â79. External Links: Document Cited by: §3. Appendix A Literature Survey Our main paper results are supported by a survey of the recent NLP literature on culture, with the aim to obtain a representative picture of the treatment of culture in recent NLP research. Our literature is systematic, but not exhaustive. We discuss a broader follow-up survey in Section A.1. We started by obtaining all papers published between 2022 and April 2026 in the main or findings track of a top-tier NLP conference (*ACL, EMNLP) which mention âcultureâ or âculturalâ in their title. To keep the scope for manual paper inspection manageable, we removed multimodal approaches, and papers whose primary focus is not culture. We reviewed a total of 107 papers. We manually labelled these papers for the following information: Primary contribution (resource, method, analysis, survey) Data creation (predominantly manual, at least partially synthetic, based on a broad background corpus, âŠ). Only for empirical papers (i.e., excluding surveys) Task Assessment/improvement of cultural knowledge in LLMs (upstream or intrinsic) vs. task-oriented papers proposing culturally sensitive applications (downstream or extrinsic). Only for empirical papers (i.e., excluding surveys). The full list of reviewed papers and labels is provided on GitHub.11 1 https://github.com/ColiLea/emic-culture-nlp/ Summary statistics underpinning our analyses in Section 2 are provided in Tables A.1 â A.3. A.1 Expanded Survey We complemented our survey, retrieving all papers in the ACL Anthology which mentioned the word âcultureâ or âculturalâ in their abstract or title (N=1536). These were infeasible to manually analyse so we instead searched for an expanded list of keywords indicating an emic approach.22 2 emic, etic, community-centric, community centric, community-centered, community centered, community-focussed, community focussed, community-focussed, community focussed, anthropological, anthropology, insider, outsider, intercultural 108 papers mentioned at least one of these terms. Even with the very optimistic assumption that all these papers take an emic approach, they would make up for 7% of our sample. We note that again this is a crude approximation because both the denominator (there will be other papers about culture which did not meet our initial keyword filter) and the numerator (there may be emic papers not mentioning any of our expanded keywords, and many matched papers wonât be emic) are noisy. We provide this complement survey as an additional data point, complementary to our more focussed manual analysis. Category N %\% Resource 50 0.467 Survey 8 0.075 Analysis 15 0.140 Method 31 0.290 Table A.1: Distribution over main contributions in papers in our survey. Category N %\% Assess / augment cultural knowledge in LLMs 78 0.729 Table A.2: Fraction of paper with a focus on assessing or augmenting cultural knowledge in LLMs. Category N %\% culture annotators 19 0.178 broad data 41 0.383 synthetic 33 0.308 translation 3 0.028 â / LLM output analysis 4 0.037 â / survey 8 0.075 Table A.3: Distribution over data set creation approaches in the papers in our survey. Papers of the last two categories were not included, as they did not contribute data sets or investigated culture in LLM output (rather than human artifacts).