It is not enough that it speaks

Artificial intelligence has become a presence that is difficult to think about calmly. It appears in headlines, business plans, classrooms, family conversations and institutional speeches. It is spoken of as though it were a historical force on the march. It arrives, transforms, threatens, replaces, creates or saves. The public grammar of AI tends to grant it an agency of its own.

This way of speaking is not innocent. When we say that “AI writes”, “AI paints”, “AI composes”, “AI decides” or “AI replaces”, the set of social relations that makes each of those actions possible is partly hidden from view. Behind a generated image lie visual archives, captured styles, prior artistic labour, training systems, companies, licences and disputes over authorship. Behind a synthetic voice lie recordings, acoustic models, imitated bodies, permissions granted or permissions absent. Behind a generated video lie computing infrastructures, learned visual patterns, attention economies and new forms of credibility. Behind a song created with AI lie musical memories, rights, cultural industries and collective sensibility.

AI does not appear on its own.

The deepest social problem lies in the reification produced around artificial intelligence. In the wake of what Marx called the fetishism of the commodity, a social relation appears as if it were an autonomous thing; and, in terms close to the reification theorised by Lukács, a web of labour, data, infrastructure and power presents itself as an independent technical object. What is infrastructure appears as magic. What is a business decision appears as technical destiny.

In the Spain of mid-2026, this enchantment no longer belongs only to technological culture. It forms part of public conversation. AI appears linked to regulation, employment, education, creativity, security and administration. Spain has a state agency dedicated to overseeing artificial intelligence and takes part in international governance spaces. That institutional existence matters, but it does not exhaust the problem. It also coexists with doubts about the real capacity for oversight in a field that changes faster than its own legal frameworks.

This essay proposes reading that scene through three Spanish reference points. Carmen Victoria Marrero Aguiar, Marta Peirano and Jon Hernández allow us to illuminate different planes of one and the same transformation. They do not form a school, nor do they hold identical positions. That difference is what makes the dialogue fruitful. Marrero leads towards language, voice, perception and evidence. Peirano shifts the gaze towards infrastructure, data, companies and power. Hernández brings in the social experience of the transition, with its mixture of pedagogy, urgency, opportunity and fear.

The underlying question is not about deciding who is most right. The question is another. What happens to us as a society when a technology begins to occupy the symbolic place of intelligence, creativity, authority and the future. What becomes invisible when the mass media and the market turn “AI” into a historical character. What we stop looking at when a social infrastructure presents itself as a mind.

Three trajectories before a single scene

Carmen Victoria Marrero Aguiar is a professor at the UNED and a significant figure in contemporary Spanish linguistics. Her path crosses experimental phonetics, applied linguistics, auditory perception, the clinical study of language, the teaching of pronunciation, voice technologies and forensic phonetics. She also sits close to the current debates on synthetic voices and deepfakes. Her work reminds us that language is not only text, nor the voice a mere channel.

That point is decisive. Many conversations about AI reduce language to verbal output, a prompt or a string of text. Marrero compels us to recover another scale. Language is articulated, heard, learned, interpreted, pathologised, rehabilitated, measured, falsified and certified. The voice involves body, age, accent, timbre, prosody, social memory and perception. In a world of voice cloning, audiovisual avatars and synthetic assistants, the question of “who is speaking” ceases to be obvious.

Marta Peirano occupies a different position. Her journalistic and essayistic work has for years insisted on surveillance, data capitalism, digital infrastructures, corporate concentration, automation, the climate crisis and the political capture of technology. Her view of AI shifts the debate from a fascination with the models towards the conditions that make them possible. The question is not only whether a system understands, reasons or creates. The question is who trains it, with what data, under what business model, with what energy consumption, with what effects on labour and under what regime of responsibility.

That view makes it possible to understand that the mystification of AI is not produced solely by the naivety of the public. It is also a self-interested effect of the technology industry. Presenting AI as an almost inevitable entity helps blur concrete decisions. Automation appears as natural progress. The extraction of data appears as innovation. The concentration of power appears as efficiency. Material consumption appears as the cloud. Technological dependence appears as modernisation.

Jon Hernández represents another dimension of the same moment. His activity as a populariser, trainer and communicator specialising in AI sits at the crossroads between public pedagogy, practical adoption, labour anxiety and the culture of acceleration. In his discourse one finds employment, productivity, deepfakes, geopolitics, companies, medicine, education and the need to learn to live with tools that change too fast.

Hernández helps us observe how AI is socially experienced as a transition. Not only as infrastructure, nor only as a political threat, but as an everyday pressure. Professionals who feel they must adapt. Companies that adopt tools without fully understanding their effects. Students who normalise conversational assistants. Families who hear talk of job replacement. Creators who watch generated images, texts, videos and music alter the social value of their work.

The three trajectories allow us to sustain a broader reading. Marrero situates language in the body, perception and evidence. Peirano situates AI in infrastructure, capital and power. Hernández situates the transformation in the social experience of adoption. But the conversation is not entirely peaceful, and therein lies its interest.

Hernández’s pragmatism, oriented towards learning tools and not being left out of the transition, comes into friction with Peirano’s structural suspicion. Does AI literacy emancipate or manufacture consent? Does teaching people to use corporate tools increase autonomy or normalise dependence? Marrero shifts that tension towards a prior question. Before deciding whether a tool liberates or captures, it is worth attending to the sign it produces, to its materiality and to its conditions of recognition.

That friction makes it possible to dismantle the fetish. AI is not an isolated mind. It is an assemblage of models, media, data, labour, desire, institutions and narratives.

What an LLM is and what is meant by AGI

It is worth pausing on two terms that are often mixed up. An LLM, or large language model, is a machine-learning system trained on enormous quantities of text to process and generate language. Its functioning rests on complex statistical patterns learned during training and subsequently adjusted to produce responses that are useful, plausible or appropriate to an instruction. LLMs can summarise, translate, draft, classify, program, converse and simulate styles. They are not giant dictionaries or mere search engines, but neither are they people who understand the world as those who live in it do.

AGI, or artificial general intelligence, names a far more ambitious and still contested aspiration. In general terms, it refers to a system capable of operating with versatility across a wide range of tasks, adapting to open environments, transferring learning between domains and solving new problems without being confined to a narrow function. There is no full consensus on its definition or on how to measure it. That is why many discussions of AGI oscillate between science, foresight, marketing, philosophy and media imagination.

The difference matters. An LLM can be very powerful without being an AGI. It can also form part of broader architectures that integrate vision, audio, video, action, memory, external tools and planning. But to reduce AGI to linguistic fluency would be a simplification. General intelligence, if the term retains a strong sense, would demand more than producing language. It would demand a situated relationship with environments, robust learning, continuity, perception, action, memory and the correction of errors.

Even so, language is not secondary. In human societies, language does not only communicate. It organises thought, coordinates action, transmits memory, creates institutions, distributes authority and makes it possible to imagine futures. That is why LLMs carry a disproportionate cultural weight. Not because they are necessarily intelligent in a human sense, but because they operate in the medium through which societies make meaning.

It is worth adding an uncomfortable fact about the leading models of mid-2026, such as Claude Opus 4.8 or GPT-5.5. Their makers do not publish, in any verifiable way, the number of parameters, the exact cost of training, the total computing time or the energy consumption required to develop and deploy them. We know the magnitudes are enormous, but we cannot attribute closed figures to those specific models without lapsing into speculation. That absence of data is not a minor technical detail. It forms part of the problem this essay tries to name: systems that present themselves as public intelligence, yet whose material conditions remain protected as a trade secret. Independent estimates of frontier models already speak of investments in the tens or hundreds of millions and of a trend towards training runs on a billion-dollar scale. The question, then, is not only how large a model is, but what kind of society accepts that an infrastructure of such scale should act as a cultural mediator without properly knowing its costs, resources and conditions of production.

The media fetishisation of AI

AI has become a character in the mass media. It is announced as the author of images, the composer of songs, the writer of novels, the replacement for professions, a medical adviser, a personalised tutor, an existential threat or a productive salvation. That narrative tends to concentrate attention on the spectacular result. An image appears in seconds. A song imitates a genre. A video reconstructs an impossible scene. A voice reproduces someone who is not there. A text seems reasoned.

The spectacle produces an inversion. Instead of asking about the conditions of production, we ask about the wonder of the result. Instead of following the data, we celebrate the appearance. Instead of analysing the infrastructure, we speak of magic. Instead of identifying corporate interests, we imagine a technical force advancing on its own.

This mechanism affects not only those who reject AI. It also runs through those who use it intensively. The user may experience a mixture of fascination, relief and dependence. The tool seems to respond, to accompany, to solve and to produce. At times it becomes easier to attribute agency to “AI” than to acknowledge the set of mediations operating behind it. The conversational interface reinforces that illusion. But so too do the generated image, the automatic music, the synthetic video and the promise of frictionless productivity.

The enchantment does not consist simply in believing that AI is alive. It can be more subtle. It consists in treating it as though its technical existence were enough to justify its social effects. If it can be done, it will be done. If it will be done, we had better adapt. If we had better adapt, the political discussion arrives too late. This chain turns debatable decisions into inevitability.

An anthropological reading must interrupt that chain. AI is not destiny. It is a historical form of technical and social organisation. It has actors, costs, territories, beneficiaries, people harmed, imaginaries, resistances and conditions of possibility. Its power lies not only in what it does, but in what it manages to make us stop asking.

Language, thought and world

Linguistics makes it possible to dismantle part of the fascination. Language is not equivalent simply to communication. We can communicate without words and we can use words for something more than informing. Language takes part in reflection, memory, imagination, planning, the promise, the threat, care, the norm and conflict.

This distinction matters because generative systems produce very convincing communicative effects. They answer, summarise, translate, explain, recommend, imitate tones and simulate registers. But that effectiveness does not by itself settle the question of understanding. It is one thing to produce an appropriate response. It is another to share a world, experience, intentionality and responsibility.

The difference between grammatical competence and communicative competence helps to refine the problem. Dell Hymes formulated this second notion to remind us that speaking a language does not consist only in producing formally correct sentences, but in knowing how to use them in a socially appropriate way. A speaker knows when to keep silent, how to repair a misunderstanding, what a promise entails, what a word can wound, what authority a register carries and what consequences an accusation may have.

A model can simulate part of that appropriateness. It can adjust its tone. It can approximate a situation. It can imitate courtesy or empathy. But it has not lived a human socialisation. It does not belong to a speech community in the same sense as a person. It has no vulnerable biography, nor does it answer morally for what it says. This difference does not cancel its effects. It makes them more delicate.

Sapir and Whorf allow us to open up another dimension. The relationship between language and thought should not be framed as a deterministic prison. We do not think only what our language allows. Nevertheless, languages orient ways of classifying, attending, remembering and ordering experience. The words available matter. The categories matter. The contrasts a community stabilises matter.

Generative AI reopens that question on a technical scale. Models do not only produce responses. They also return sedimented categories, dominant associations and cultural hierarchies present in their data and adjustments. When a tool becomes the habitual mediator of searches, summaries, decisions, images, diagnoses, translations or recommendations, it takes part in the organisation of the thinkable. It does not wholly determine thought, but it can orient questions, normalise frameworks and render other possibilities invisible.

Chomsky introduces a different caution. His competence/performance pairing reminds us that observable linguistic performance does not exhaust the question of the language faculty. Producing an impressive verbal performance does not necessarily amount to understanding, acquiring language, inhabiting a world or answering morally for what is said. The apparent naturalness of the output can produce an illusion of mind. But a socially effective illusion remains socially important.

Marrero and the materiality of the sign

At this point, Marrero’s trajectory makes it possible to recover the materiality of language. The voice is a sign, but not just any sign. It carries bodily, social, technical and affective marks. It is shot through with breathing, hearing, the vocal tract, learning, age, accent, prosody, illness, gender, territory and memory.

Generative AI does not alter only writing. It alters a wider set of expressive forms. Synthetic voices, spoken machine translation, generated songs, artificial dubbing, audiovisual avatars, hyper-realistic videos, camera-less images and style simulations shift the relationship between sign and presence. It is no longer only a matter of whether a text was written by a person. It is also a matter of whether a voice corresponds to a body, whether an image comes from a scene, whether a song belongs to a performer, whether a video records an event or fabricates it.

Phonetics, auditory perception and forensic phonetics show that trust in the voice was never simple. Recognising a voice involves complex processes. Attributing an identity to a vocal emission requires listening, context and criteria of validation. AI intensifies that fragility. The generated voice can sound familiar without being tied to the physical presence of the one who seems to speak.

The question extends to image, music and video. Each modality touches a different zone of social life. The image touches visual memory. The video touches the event. The voice touches presence. Music touches shared sensibility. Text touches argumentative authority. Taken together, generative AI does not only produce content. It reorganises the conditions of trust in signs.

Marrero allows us to maintain that language and the voice should be thought of as embodied practices, not as mere outputs. By introducing this materiality, the fetish weakens. AI ceases to appear as a magical mind and begins to be seen as a technology that intervenes in human signs deeply laden with body, trust and social relation.

Peirano and the infrastructure of the fetish

Peirano focuses on what fascination tends to cover up. AI does not operate in thin air. It depends on data centres, energy, water, chips, platforms, contracts, labelling teams, content moderation, data harvesting, financial investment and regulation. The cloud is a misleading metaphor when it makes us forget the ground.

Technological mystification turns that infrastructure into abstraction. The user sees a clean interface. The media show a spectacular demonstration. Companies promise efficiency. Politics speaks of modernisation. But the material chain remains out of sight. That invisibilisation is part of the power of AI. The more natural the tool seems, the less its conditions are discussed.

The institutional dimension is also concealed. A model does not need consciousness in order to take part in decisions with real effects. It can filter, classify, prioritise, recommend, surveil, summarise, translate and order information in labour, administrative, educational or policing processes. Agency lies not only in the model. It lies in the system that decides to use it, in the institution that legitimises it and in the company that designs it.

This is why the question of AGI should not absorb the whole debate. Before any hypothetical general intelligence, there already exist partial intelligences integrated into general systems of power. Systems that do not think like people, yet affect people. Systems that do not understand like a community, yet can alter the categories with which a community understands itself.

The question is displaced. Not only what AI can do, but under what political economy it does what it does. Not only whether it makes mistakes, but who answers when it does. Not only whether it creates, but what it extracts in order to produce. Not only whether it will be conscious, but what forms of social unconsciousness it generates when it is accepted as inevitable.

From here, the friction with Hernández becomes productive. Learning tools may be necessary, but the question does not end there. We must also discuss who defines that literacy, which platforms organise it, what commercial interests run through it and what dependencies it leaves installed. A public culture of AI should not limit itself to teaching people to use already-given systems. It should also teach people to recognise their conditions of production, their limits, their costs and the forms of power they normalise.

Hernández, Bauman and the social experience of acceleration

Hernández allows us to observe another plane. AI is experienced as urgency. His public discourse reflects a widespread sense of accelerated transition. Many people do not ask in the abstract what intelligence is. They ask what will happen to their work, what they must learn, whether they are arriving too late, whether their children will study differently, whether a company will be able to replace tasks, whether an image will be credible, whether an audio recording will be authentic, whether a profession will still have value.

The culture of AI adoption has a pedagogical component, but also an emotional one. It promises control amid uncertainty. Learn the tools, understand the change, do not get left behind. This promise can be mobilising. It can also reinforce anxiety. If the technology appears inevitable, responsibility shifts onto the individual person. Those who fail to adapt seem guilty of their own obsolescence.

Zygmunt Bauman described liquid modernity as a condition in which stable structures dissolve and the task of staying afloat falls increasingly on each individual. The culture of AI adoption prolongs that diagnosis. It presents an environment of permanent change in which adapting ceases to be a collective decision and becomes a private obligation. Each person must update themselves, learn new tools, reconfigure their labour value and prove they have not been left behind. What in Bauman was the dissolution of stable frameworks becomes here the dissolution of professional certainties.

There arises an everyday form of enchantment. AI presents itself as an external force to which one must adjust, not as a social process that can be discussed, regulated, limited, redesigned or reappropriated. The transition becomes a personal destiny. Politics turns into training. Labour conflict is translated as continuous learning.

The experience of use reinforces that sensation. A tool generates a text, an image, a song, a video or a work strategy in seconds. Speed produces authority. What comes out without friction seems more intelligent than it perhaps is. The abundant seems valuable. Yet anthropology reminds us that adoption is never homogeneous. Technologies do not reach all social classes, territories, genders, ages or professions in the same way. Between the demonstration and life there are mediations.

Hernández allows us to situate the debate in that intermediate zone. AI is neither only invisible infrastructure nor only media fetish. It is also everyday practice, learning, fear, opportunity, pressure and social conversation. The point is not to ridicule the urgency, but to politicise it. A society cannot limit itself to adapting to AI. It must discuss which forms of adaptation are just, which are imposed and which it is worth rejecting.

Three ways of de-fetishising AI

Although they do not agree on everything, the three trajectories need one another. Precisely because there is friction between them, they make it possible to think AI without reducing it to a single plane.

Marrero leads towards concrete signs. Voice, hearing, perception, accent, phonetics, evidence. Against the abstraction “AI speaks”, a more precise question appears. What kind of sign is produced, how it is perceived, what body it imitates, what trust it claims and what criteria allow it to be validated.

With Peirano, the gaze shifts towards infrastructure. Data, companies, energy, water, regulation, surveillance, labour chains, contracts. It is no longer enough to say that “AI is advancing”. We must ask who decides, who pays, who benefits, who is exposed and who answers.

Hernández allows us to situate the discussion in social experience. Work, learning, anxiety, adoption, productivity, expectations. Against the abstraction “AI replaces”, a more concrete question appears. Which tasks change, which professions are reconfigured, which inequalities widen, which capabilities are democratised and which discourses turn adaptation into an individual obligation.

These three ways are not mutually exclusive. They need one another. AI as fetish can only be dismantled if it is followed at once in its signs, its infrastructures and its social uses. It is not enough to analyse the model. It is not enough to denounce the company. It is not enough to teach the tools. The phenomenon is broader.

It is not enough that it speaks

The title remains valid, but no longer as a reduction of the problem to spoken or written language. It is not enough that it speaks because AI also looks, draws, sings, composes, animates, summarises, recommends, classifies and simulates. It is not enough that it speaks because speaking does not by itself demonstrate understanding. It is not enough that it speaks because the focus on the machine’s voice can conceal the human voices, the human labour and the human interests that lie behind it.

The decisive question is not whether AI speaks. The question is how a society comes to hear it as authority. How it comes to see its images as evidence. How it comes to accept its videos as event. How it comes to incorporate its music as sensibility. How it comes to take its classifications as knowledge. How it comes to present its decisions as neutrality.

Contemporary artificial intelligence must be taken seriously, but not as a fetish. Not as a magical subject or an autonomous destiny. It must be taken seriously as a total social fact, in Marcel Mauss’s sense. A phenomenon that at once runs through economy, law, politics, technique, symbolism, affects, labour and institutions.

De-fetishising AI means returning it to its real conditions. It means looking at the data centres when an interface seems magical. Listening to a cloned voice while remembering the absent body. Looking at a generated image while remembering the archives, styles and labour that precede it. Listening to a synthetic song while remembering the musical traditions it imitates. Reading an automatic text while remembering the categories it reproduces. Analysing a promise of productivity while remembering the jobs it displaces, devalues or reorders.

A democratic society does not need less technology. It needs less enchantment. It needs better questions, better mediations, better institutions and a public culture capable of distinguishing between technical power and social authority.

AI is not simply a machine that speaks. It is a technical mirror in which a society projects desires for efficiency, fear of replacement, fantasies of control, inequalities and imaginaries of the future. That is why it is not enough to listen to what it says. We must ask who makes it speak, with what materials, in whose interests, for which publics and with what consequences.

Only then does artificial intelligence cease to be enchantment and return to being an object of shared discussion.

References

Agencia Española de Supervisión de la Inteligencia Artificial. (n.d.). Garantizando una IA ética y responsable.

Bauman, Z. (2000). Liquid Modernity. Polity Press.

Cadena SER. (2026, 28 May). Óscar López inaugura en València el Laboratorio de Gobernanza de IA para la Humanidad de la ONU: “Debe servir a la paz”.

Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press.

Cottier, B., Rahman, R., Fattorini, L., Maslej, N., Besiroglu, T. & Owen, D. (2024). The rising costs of training frontier AI models. arXiv.

Hernández, J. (2026). La hostIA que viene. Planeta.

Hymes, D. (1972). On communicative competence. In J. B. Pride & J. Holmes (Eds.), Sociolinguistics: Selected Readings (pp. 269–293). Penguin.

Lukács, G. (1923). History and Class Consciousness. Malik-Verlag.

Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B. & Fedorenko, E. (2023). Dissociating language and thought in large language models. arXiv.

Marrero Aguiar, C. V. (n.d.). Teaching and research profile at the UNED.

Marx, K. (1867). Capital. A Critique of Political Economy. Volume I. Verlag von Otto Meissner.

Mauss, M. (1925). The Gift. The Form and Reason of Exchange in Archaic Societies. L’Année Sociologique.

Sapir, E. (1929). The status of linguistics as a science. Language, 5(4), 207–214.

Triguero, I., Molina, D., Poyatos, J., Del Ser, J. & Herrera, F. (2023). General Purpose Artificial Intelligence Systems (GPAIS): Properties, Definition, Taxonomy, Societal Implications and Responsible Governance. arXiv.

Whorf, B. L. (1956). Language, Thought, and Reality: Selected Writings of Benjamin Lee Whorf. MIT Press.