AI Does Not Evolve Alone

Anthropological approaches to reading a technology in dispute

Artificial intelligence is often presented as an absolute rupture: a technology capable of transforming work, science, medicine, education, security, war and perhaps the very definition of the human. Yet many of the questions it raises are not new. For more than a century, social and cultural anthropology has studied how hierarchies are produced, how institutions are organised, how difference is constructed and how particular relations of power come to seem inevitable.

This essay does not seek to decide which anthropological school AI “belongs” to. That would be too narrow a question. AI is, at once, a narrative of progress, a machine for classification, a material infrastructure, a commodity, a cultural practice and a device of governance. Each tradition brings one part of the problem into view. None explains it in full.

Emily Chang’s interview with Dario Amodei, Anthropic’s chief executive, offers a revealing corporate scene. Amodei describes AI development as a “smooth exponential”: for a time, nothing seems to happen and then, suddenly, everything accelerates. Against either panic or denial, he advocates a rational management of risk while arguing that advanced models can compress decades of scientific, medical and economic progress (Bloomberg Originals, n.d.).

The interview does not provide transparent access to an individual consciousness. It is a piece of corporate communication in which several voices intervene: that of the engineer, the company leader, the promoter of an innovation, the claimant of regulatory capacity and the person seeking to reassure customers, shareholders and public authorities.

The case of Claude Mythos Preview makes it possible to follow this thread. Anthropic introduced the model as unusually capable of finding software vulnerabilities and launched Project Glasswing to restrict its initial access to defensive organisations and maintainers of critical software (Anthropic, 2026b). Weeks later, the company argued that the constraint no longer lay only in finding flaws, but in verifying, reporting, prioritising and deploying fixes for them (Anthropic, 2026c). Intelligence is not exhausted by identifying a pattern. It depends on those who interpret it, debate it and act upon it.

The anthropological task is to examine which idea of progress organises this narrative, who defines responsible uses, which bodies and territories sustain the infrastructure, who receives the benefits and who is exposed to harm.

The frontier produces its outside

Classical evolutionism, developed by Tylor (1871) and Morgan (1877), imagined human history as an ascending sequence. Morgan organised that sequence through the familiar triad of “savagery, barbarism and civilisation”, understood as a universal journey towards higher forms of technology, institution and social life.

That theory was discredited for its Eurocentrism and for turning historical and cultural differences into a single scale of development. Yet part of its logic reappears in contemporary AI vocabulary.

Amodei speaks of “frontier” models and an economic premium attached to intelligence. The sequence no longer classifies peoples as primitive or civilised. It ranks models, firms, infrastructures and states according to computing capacity, access to data, energy, chips and capital. Some actors move ahead, others fall behind, and history seems to have a direction marked by technology (Bloomberg Originals, n.d.).

The resemblance is not literal, but it is structural. Those who control the most capable models appear as the vanguard. Those unable to access them occupy a subordinate or dependent position.

Boas (1911) offers a decisive corrective. Against unilinear evolution, historical particularism requires attention to concrete trajectories. There is no single AI as a universal stage. A model trained in English and financed from Silicon Valley is not equivalent to a language tool developed by an Andean university, an African public institution or a community that defines its technological ends for itself.

Speaking of AI in the singular can erase local conditions of production, use, refusal and reappropriation. It can also conceal a foundational linguistic inequality. Models that concentrate their strongest capacities in heavily represented languages may treat local registers, idioms and oral memories as noise, deficit or mere data scarcity.

Escobar (1999) showed that modern development recognised and denied difference at the same time. It acknowledged that certain peoples were different, but turned that difference into a problem to be solved through modernisation. The technological frontier works in a similar way when it produces “lagging” subjects who must be brought into the technical present.

Luxemburg (1913) provides an unsettling extension. In The accumulation of capital, she argued that capitalist expansion needed to relate continually to non-capitalist domains, from which it obtained markets, raw materials and labour power. Her thesis has been debated and should not be treated as a universal law. It nevertheless helps formulate a precise problem. AI’s frontier requires a material outside that makes its advance possible.

That outside includes extractive territories, electrical grids, water, data-labelling work, content moderation, public archives, minoritised languages and communities whose data can be turned into raw material without effective participation in the decision.

In the Mythos case, that outside also includes free-software maintainers. Their repositories, reports and repair practices are part of the technical environment in which the model operates, even when they do not necessarily participate in corporate decisions about its deployment. The frontier does not merely leave people behind. It can produce dependency as a condition of its own advance.

Structuralism does not reveal the world’s source code

Lévi-Strauss’s structuralism studied the relations that organise classifications, myths, kinship and symbolic oppositions. It did not seek a hidden technical code inside cultures. It sought to understand how people order experience through differences, rules and relations (Lévi-Strauss, 1958).

AI can be read through that tradition, with a crucial caveat. A model does not discover the world’s source code. It learns regularities from data, categories, archives, tests and design decisions. What appears to be a deep structure may be the accumulated outcome of earlier human classifications.

Mythos makes this tension visible. The system may reconstruct relations among dependencies, permissions, functions, errors and entry points that a person may not find as quickly. Yet a vulnerability is never merely a technical fact. It is also a relation among design, maintenance, budgets, work organisation and response capacity.

Douglas (1966) helps deepen this point. Classification is not simply a way of ordering information. It also marks what is pure and dangerous, acceptable and contaminating. A company’s AI-security discourse establishes comparable boundaries when it decides which use counts as legitimate, which risk is tolerable and which collective is defined as a potential threat.

In Race and history, Lévi-Strauss (1952) offers a further warning. No culture, institution or technology can present itself as a universal formula for progress. AI risks doing so when it turns one historical form of knowledge production—textual, statistical, scalable and predominantly Anglophone—into the general measure of intelligence.

There is no universal syntax of social life waiting to be deciphered. There are relations, memories, differences and conflicts that a system may partially register but can never exhaust.

A change of employment is not a ladder

Malinowski’s functionalism examined how practices such as exchange, kinship, economy and magic sustained particular forms of life (Malinowski, 1922). AI is now presented in a functionalist register. It is meant to cure diseases, improve education, make energy cheaper, accelerate science, strengthen security and increase growth.

Its usefulness may be real. Yet an AI system that filters CVs may serve a firm and harm those seeking work. A tool that drafts reports may speed up an administration while making its procedure more opaque to the public. A system that detects patterns may assist a hospital, insurer or police force, with markedly different consequences for the people it classifies.

Radcliffe-Brown (1952) shifts attention from individual utility towards the function a practice fulfils for a social structure. From this perspective, it matters which order AI helps sustain. It may reduce costs and increase efficiency, but it may also concentrate decisions, degrade professions, intensify surveillance and transfer risk to those with less bargaining power.

Amodei has pointed to the emergence of new jobs, such as the forward deployed engineer. He also acknowledges that emerging roles do not automatically absorb people who lose previous positions (Bloomberg Originals, n.d.). This matters. New occupations do not necessarily appear in the same numbers, in the same places or for the people whose work disappears.

The move from an automated occupation to an emerging one depends on education, income, language, mobility, professional networks, residency rights, care work, property and social position. Between a Silicon Valley engineer and a person working in a call centre in Manila, Nairobi or Bogotá, there is not only a difference of skills. There is a difference of position within an unequal structure of ownership, access and decision-making power.

Wolf (1982) situates that distance within a global history. Apparently separate spaces are constituted through relations of exchange, extraction and domination. Automation does not connect worlds that were previously isolated. It reorganises hierarchies that had already linked territories, bodies and labour markets.

Butler (1872) anticipated, with irony and acuity, that machine evolution does not unfold apart from human evolution. Artefacts and people develop in a relation of mutual dependence. Stiegler (1994) radicalised this intuition by thinking of technics as an exteriorisation of memory and knowledge. The problem arises when this exteriorisation becomes cognitive proletarianisation, when people lose the capacity to understand and practise what they systematically delegate.

Algorithmic acceleration does not only automate tasks. It can destabilise the relationship among work, wages, social protection and the everyday reproduction of life. That disruption introduces entropy into socioeconomic systems that still depend, in order to persist, on income, care and employment-linked rights.

Scott (1998) showed that modern institutions tend to simplify complex realities in order to make them legible and governable. AI expands that capacity for simplification. It optimises processes, but may eliminate what does not fit neatly into a category: an irregular work history, a family context, a local language, a non-standard way of writing or a care experience that is difficult to quantify.

Attention is therefore needed to which lives become more administrable and which lose the capacity to become intelligible on their own terms.

Actor-network, controversies and corporate governance

Functionalism helps explain what a technology does within an institution. On its own, it does not explain how it is built, who participates in its assemblage or how agency is distributed among people, rules, infrastructures, documents and technical objects. Another perspective is needed for that.

Actor-network theory, developed by Latour (2005), helps avoid two errors. The first is to treat AI as an autonomous agent arriving from the future. The second is to reduce it to a passive tool without effects of its own.

An AI system is an assemblage of code, data centres, electricity, water, minerals, chips, cloud contracts, moderation labour, labelling, export rules, interfaces, repositories, metrics, military institutions, security discourses and everyday habits. No model acts alone. It acts because a network of people, objects and organisations makes its action possible.

Mythos is a clear example. Its effectiveness depends on the model, but also on software repositories, maintainers, validation teams, responsible-disclosure procedures, companies able to deploy patches and organisations that decide who may use it. Its intelligence does not reside only in its parameters. It is distributed across that network.

Actor-network theory does not claim that everything is indistinguishably an actor. It proposes following how agency is distributed, concentrated and transformed within heterogeneous associations. Data centres, encryption standards, evaluation metrics and responsible-disclosure protocols do not replace human decisions. They participate in the forms those decisions can take.

Descola (2005) helps situate even this framework. Latour challenged the modern separation between nature and society, but the very distinction between humans and non-humans cannot be assumed to be universal. Descola reminds us that the nature–culture division belongs to a historically specific way of ordering the world. His contribution does not invalidate actor-network theory. It prevents its vocabulary from becoming a universal grammar of agency.

The same perspective reaches corporate governance. Anthropic has a Long-Term Benefit Trust, a body whose members have no financial stake in the company. The Trust can elect and remove a growing portion of the board and, since April 2026, Trust-appointed directors have formed a majority of Anthropic’s board (Anthropic, 2023, 2026a).

This is a form of private constitutionalism. It may introduce meaningful counterweights to immediate financial logics. Yet it is not an automatic guarantee of public good. The Trust exists through documents, appointments, internal rules, reports, metrics, voting systems, contracts and legal mechanisms.

Latour and Woolgar (1979) showed that scientific facts stabilise through chains of inscription. Reports, graphs, articles, protocols and measurements do not simply describe a reality. They participate in producing and consolidating it. Something similar may occur with the “long-term public benefit” invoked by a corporate trust.

It matters who defines its indicators, who audits its outcomes, which controversies remain outside its documents and who can contest its decisions from outside the company.

The Trust may be a relevant institutional limit. It is not democracy. People affected by the models, those who work within its supply chain or the communities sustaining its infrastructure do not necessarily elect those who define its mission.

Responsible governance requires mechanisms of public contestation, independent audits, labour rights, transparency and effective participation. A corporate mission only acquires public legitimacy when those who bear its consequences have a real capacity to debate it.

Coloniality, labour and the capacity to decide

AI is often imagined as immaterial. An answer appears on a screen and seems to have no past, territory or cost. Yet it depends on minerals, energy, water, data centres, low-paid human labour, extracted data and global logistics chains.

Crawford (2021) has shown that AI should be studied as a planetary infrastructure before it is treated as an isolated intelligence. Couldry and Mejias (2019) propose the term data colonialism to name the systematic appropriation of social life as a resource for accumulation.

The decolonial critique goes beyond unjust data extraction. It also examines the claim of an abstract, universalizable intelligence capable of translating any experience into computational patterns. That claim can render invisible ways of knowing, feeling and inhabiting the world that do not need to be scalable in order to be valuable.

Quijano’s (2000) formulation of the coloniality of knowledge helps name the problem. Not all knowledge is produced from the same place or possesses the same capacity to impose itself as universal. When a company defines from California what counts as security, intelligence, harm, evidence or alignment, it does not apply neutral categories. It projects a situated way of knowing that can become global infrastructure.

This does not require the conclusion that every AI system is colonial by definition. It requires recognising a tension. A system can reproduce coloniality when it turns languages, memories, affects and relationships into inputs without consent, reciprocity or shared decision-making power. It can also create spaces for reappropriation when communities control the ends, infrastructure, conditions of use and right to refuse.

The Global South cannot appear only as a site of extraction or harm. It is also a space of technical work, linguistic defence, labour organisation, tactical appropriation and resistance.

That claim has concrete scenes. Posada (2021) documented the coloniality of data work in Latin America through interviews with Venezuelan workers on labelling platforms. Tubaro et al. (2025) compared conditions of AI-linked digital labour in Argentina, Brazil and Venezuela, showing how precarity articulates with specific geographies.

In another register, Huaman et al. (2026) presented automatic speech-recognition resources for Puno Quechua, developed through a participatory design campaign with its speakers. That case shows that resistance is not limited to rejecting an imposed technology. It can also mean building tools from and for specific communities, challenging the idea that innovation can only come from the corporate centre.

Taussig (1980/2010) showed that workers do not experience extraction only as an economic relation. They also produce their own languages for naming and criticising it. His analysis of the devil and commodity fetishism invites attention to the categories developed by those who label data, moderate violent material, extract minerals or sustain energy infrastructures for AI. Those voices rarely occupy the centre of corporate debates on responsible innovation.

Pérez Orozco (2011) provides another necessary shift. Automation is often discussed through productivity, growth, retraining and taxation. Yet the sustainability of life also depends on care, reproduction and maintenance work that does not disappear because a company automates cognitive tasks.

Redistributing the surplus generated by this transformation matters, but democratising the capacity to decide is more far-reaching. It requires affected people to intervene before ownership, data and infrastructure become concentrated.

How to study a planetary and situated technology

Seaver (2017) has shown that technical settings themselves do not maintain a stable definition of “algorithm”. Depending on the context, the term may refer to code, work teams, classification procedures, emergent effects or a convenient explanation for decisions that are difficult to attribute.

That instability is not a methodological failure. It is an ethnographic finding. Algorithms are multiple objects constituted through diverse practices. Studying them requires tracing discourses, controversies, everyday uses, conditions of access and relations of power without waiting for total transparency from corporate black boxes.

Amodei’s interview should be read from that perspective. His statements about risk, employment, the frontier and security are not only data. They also organise expectations, distribute responsibility and delimit which debates appear reasonable. The acknowledgement that new jobs will not automatically absorb people who lose earlier ones humanises the discourse without necessarily altering the general direction of change that is presented as inevitable.

AI also poses a problem of scale. It is planetary infrastructure and situated practice at once. Marcus (1995) proposed multi-sited ethnography as a way to follow objects, metaphors, conflicts and relations across different territories. An anthropology of AI needs a comparable movement.

It can follow a model trained in California as it reaches other languages, observe how a corporate decision materialises in particular bodies and jobs, trace the way a classification alters a right or accompany forms of resistance that reconfigure global systems from specific places.

Opening the black box is not enough. We also need to follow the effects of what the black box makes possible.

AI is not the next stage

Anthropic’s vision assumes that technical acceleration can radically expand human possibilities. It is not a trivial position. It recognises risks of unemployment, wealth concentration, surveillance, military use and cybersecurity, and it proposes taxation, controls and corporate-governance mechanisms (Amodei, 2026; Bloomberg Originals, n.d.).

Yet it retains a deeply modern confidence in scientific acceleration, rational planning and the possibility that sufficiently responsible institutions can steer a technical transformation of immense scale.

Anthropology introduces a necessary discomfort. No innovation arrives in an empty world. Every technology enters societies shaped by colonial memories, linguistic hierarchies, territorial disputes, labour inequalities and pre-existing concentrations of power.

Evolutionism helps identify the narrative of inevitable progress. Boas requires suspicion of any universal AI. Luxemburg helps reveal the material outside the frontier requires. Structuralism shows that classification is not innocent description. Douglas reminds us that every classification organises boundaries. Functionalism helps reveal which order institutions sustain. Latour and actor-network theory locate the model within material and social networks. Descola prevents that very way of ordering agency from being universalised. The decolonial critique requires knowledge to be located and decision-making capacity to be distributed.

What matters is who can decide what AI should serve, who can refuse it, who answers when it fails and which forms of life deserve protection even when they are not the fastest, most profitable or most scalable.

AI does not evolve alone.

It is designed, financed, trained, extracted, repaired, imposed, debated and resisted.

Its direction is not written into its parameters. It is contested.

Anthropology no longer needs to prove that it has something to say about artificial intelligence. The challenge is to build organisations, alliances across disciplines, public infrastructures and collective rights capable of preventing this dispute from being resolved, once again, in favour of those who already concentrate power.

References

Amodei, D. (2026, January). The adolescence of technology: Confronting and overcoming the risks of powerful AI. https://darioamodei.com/essay/the-adolescence-of-technology

Anthropic. (2023, September 19). The Long-Term Benefit Trust. https://www.anthropic.com/news/the-long-term-benefit-trust

Anthropic. (2026a, April 14). Anthropic’s Long-Term Benefit Trust appoints Vas Narasimhan to Board of Directors. https://www.anthropic.com/news/narasimhan-board

Anthropic. (2026b, April 7). Assessing Claude Mythos Preview’s cybersecurity capabilities. https://www.anthropic.com/research/mythos-preview

Anthropic. (2026c, May 22). Project Glasswing: An initial update. https://www.anthropic.com/research/glasswing-initial-update

Bloomberg Originals. (n.d.). Inside the mind of Anthropic CEO Dario Amodei | The Circuit | Extended interview [Video]. YouTube. https://www.youtube.com/watch?v=x2VHFgyawPE

Boas, F. (1911). The mind of primitive man. Macmillan.

Butler, S. (1872). Erewhon: Or, over the range. Trübner & Co.

Couldry, N., & Mejias, U. A. (2019). Data colonialism: Rethinking big data’s relation to the contemporary subject. Television & New Media, 20(4), 336–349. https://doi.org/10.1177/1527476418796632

Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press. https://yalebooks.yale.edu/book/9780300252392/atlas-of-ai/

Descola, P. (2005). Par-delà nature et culture. Gallimard.

Douglas, M. (1966). Purity and danger: An analysis of concepts of pollution and taboo. Routledge & Kegan Paul.

Escobar, A. (1999). Antropología y desarrollo. Maguaré, (14), 42–73. https://revistas.unal.edu.co/index.php/maguare/article/view/11135

Huaman, E., Gamarra Lafuente, A., Cordova, J., & Korhonen, A. (2026). Building community-centred NLP resources for Puno Quechua [Preprint]. arXiv. https://arxiv.org/abs/2605.28253

Latour, B. (2005). Reassembling the social: An introduction to actor-network-theory. Oxford University Press. https://academic.oup.com/book/52349

Latour, B., & Woolgar, S. (1979). Laboratory life: The social construction of scientific facts. Sage.

Lévi-Strauss, C. (1952). Race et histoire. UNESCO.

Lévi-Strauss, C. (1958). Anthropologie structurale. Plon.

Luxemburg, R. (1913). Die Akkumulation des Kapitals: Ein Beitrag zur ökonomischen Erklärung des Imperialismus. Buchhandlung Vorwärts Paul Singer.

Malinowski, B. (1922). Argonauts of the Western Pacific: An account of native enterprise and adventure in the archipelagoes of Melanesian New Guinea. George Routledge & Sons. https://www.gutenberg.org/ebooks/55822

Marcus, G. E. (1995). Ethnography in/of the world system: The emergence of multi-sited ethnography. Annual Review of Anthropology, 24, 95–117. https://doi.org/10.1146/annurev.an.24.100195.000523

Morgan, L. H. (1877). Ancient society: Researches in the lines of human progress from savagery through barbarism to civilization. Henry Holt and Company. https://www.gutenberg.org/ebooks/45950

Pérez Orozco, A. (2011). Crisis multidimensional y sostenibilidad de la vida. Investigaciones Feministas, 2, 29–53. https://revistas.ucm.es/index.php/infe/article/view/38603

Posada, J. (2021). The coloniality of data work in Latin America [Doctoral consortium paper]. AIES 2021 Doctoral Consortium. https://arxiv.org/abs/2105.06262

Quijano, A. (2000). Colonialidad del poder, eurocentrismo y América Latina. In E. Lander (Ed.), La colonialidad del saber: Eurocentrismo y ciencias sociales. Perspectivas latinoamericanas (pp. 201–246). CLACSO.

Radcliffe-Brown, A. R. (1952). Structure and function in primitive society: Essays and addresses. Cohen & West.

Scott, J. C. (1998). Seeing like a state: How certain schemes to improve the human condition have failed. Yale University Press.

Seaver, N. (2017). Algorithms as culture: Some tactics for the ethnography of algorithmic systems. Big Data & Society, 4(2), Article 2053951717738104. https://doi.org/10.1177/2053951717738104

Stiegler, B. (1994). La technique et le temps, 1: La faute d’Épiméthée. Galilée.

Taussig, M. T. (2010). The devil and commodity fetishism in South America (30th anniversary ed.). University of North Carolina Press. (Original work published 1980). https://uncpress.org/9780807871331/the-devil-and-commodity-fetishism-in-south-america/

Tubaro, P., Casilli, A. A., Fernández Massi, M., Longo, J., Torres-Cierpe, J., & Viana Braz, M. (2025). The digital labour of artificial intelligence in Latin America: A comparison of Argentina, Brazil, and Venezuela [Preprint]. arXiv. https://arxiv.org/abs/2502.06317

Tylor, E. B. (1871). Primitive culture: Researches into the development of mythology, philosophy, religion, language, art, and custom (Vol. 1). John Murray. https://www.gutenberg.org/ebooks/70458

Wolf, E. R. (1982). Europe and the people without history. University of California Press.