AIthropology Lab
ESENHI

Visual anthropology and AI

Cinema ParAIso

Artificial intelligence is reshaping audiovisual production, but the debate is not only about generated images. It affects labour, archives, languages, rights, representation and the power to decide which stories circulate.

·Martín González Senosiain
Film set with crew, costume, sound equipment and visual layers of artificial intelligence.

When Netflix confirmed that it had used generative artificial intelligence in El eternauta, it was not talking about a Silicon Valley experiment or a United States superproduction. It was talking about an Argentine series. The platform explained that a shot showing a building collapsing in Buenos Aires was completed ten times faster than with conventional visual effects workflows and that, without that tool, the scene would have fallen outside the available budget. Reuters reported the announcement.

The case does not prove that AI is a general solution for audiovisual production. It shows a concrete tension. A technology can expand the repertoire of a production while also changing working conditions by demanding more results with less time, budget or crew.

The answer depends on where one looks from. In India, where cinema operates within a multilingual ecosystem of enormous scale, AI is being used for dubbing, lip synchronisation, mythology-based content and new versions of catalogue films. Reuters documented the use of NeuralGarage technology to dub War 2 from Hindi into Telugu and the conflict caused by a version of Raanjhanaa whose ending was altered with AI for its re-release. The rights owner defended that version as a reinterpretation, while the director Aanand L. Rai and the actor Dhanush publicly rejected it. Reuters’ investigation and The Guardian’s coverage show that the dispute is not only about a tool. It is about authorship, the memory of a work, performance rights and control over catalogues.

Argentina and India are not interchangeable examples of a supposed Global South. They are industries with different histories, languages, scales, working conditions and relations with platforms. That is precisely why they help prevent this transformation from being reduced to Hollywood, Brussels or California-based companies.

AI arrives in an industry that matters on several levels. It is collective and highly specialised labour. It is a network of studios, cinemas, festivals, platforms, schools, unions and suppliers. It is an economy of rights and catalogues. It is also a technology of memory, because a film preserves gestures, voices, landscapes, ways of dressing and ways of imagining who belongs to a society. Changing how images are made means intervening in all those layers at once.

Cinema does not begin with the prompt

A generated image may look like a film, at least for a few seconds. That resemblance, however, says little about the process that disappears when the result is mistaken for the whole work.

Visual anthropology has spent decades insisting that images are not transparent windows. A camera never records from a neutral place. It selects, frames, excludes, orders and gives a particular form to what appears. Elisenda Ardèvol showed that audiovisual media participate in the circulation of representations of cultures and ways of life, and that an image alone does not easily distinguish between performed behaviour, genre convention and social reality. José C. Lisón offered an equally simple warning: the audiovisual tool does not replace the anthropological question, the context, the interpretation or the responsibility of whoever looks.

AI does not break that problem. It amplifies it. A model does not make decisions in a vacuum. It translates an instruction through statistical models, visual repertoires, safety filters, interface designs, commercial conditions and training data that other people selected or circulated before. It may propose a shot, a texture or a plausible voice. It does not decide by itself what gaze sustains a scene, what emotional continuity a cut requires or what silence alters the meaning of a performance. Calling it simply a tool may therefore be insufficient. It is an infrastructure that organises possibilities and limits.

It is also worth remembering that cinema is not only image. Audiovisual anthropology has understood the medium as a sensory and situated experience, where sound, body, rhythm, space, silence, gesture and memory matter as much as framing. That perspective helps explain why a spectacular shot is not yet a mise-en-scène, and why a cloned voice or a retouched face are not purely technical changes. They alter the relation between performance, editing and presence.

The distinction does not require idealising the analogue. Throughout film history, every innovation has changed trades, languages and ways of working. But it avoids another simplification. This is not a choice between human creativity and technology. It is about asking which tasks are automated, who makes that decision, which forms of knowledge are preserved and which forms of collaboration become possible or expendable.

An efficiency that also reorganises labour

AI already runs through pre-production, visual references, previsualisation, audio cleaning, localisation, lip synchronisation, material classification, image corrections and visual effects. For small teams, it can reduce testing time and open access to resources once reserved for large-budget productions. For cinematography, it can help test an atmosphere. For sound teams, it can help explore a correction before a definitive process. For multilingual production, it can help a work travel between audiences with less friction.

The actress Diana Palazón places the issue in less spectacular scenes, and perhaps for that reason more revealing ones. Before moving into concrete examples, she warns of something that runs through the whole debate: the changes feel too vast, too fast and too overwhelming for the industry to know clearly yet what it is talking about. That uncertainty does not weaken her testimony. On the contrary, it shows an experience shared by many people in the industry: the feeling that the technology is already entering the craft before there are enough words, rules and agreements to understand it.

In self-tapes, AI can help change the background of a test recorded at home when a plain wall, clean lighting and a professional appearance are required. That help exists, but it does not remove a prior loss. The audition is made alone, without a casting direction present, without a voice that can accompany, correct or propose another reading.

On set, she observes something similar with sound. In the past, an aeroplane, a motorbike or an external noise could force a scene to stop. Now another response is beginning to settle in. It will be cleaned later with AI. That can save a take and protect an emotional continuity that used to be lost through an outside interruption. It can also reinforce a logic of lower costs and greater speed. If everything can be fixed later, there is less time to repeat, rehearse or hold the scene in another way.

None of that is minor. The mistake would be to call every reduction in costs democratisation. A cheap tool can move the cost elsewhere in the process. Many tests must be generated, continuity checked, errors corrected, rights verified, consent guaranteed, versions documented and possible legal conflicts answered. The efficiency of one task can coexist with compressed schedules and with the expectation that a team will do more without more time, pay or staff.

Here a concern appears that is not nostalgic. The British Film Institute warned that automation may especially affect entry-level positions. Not because a trade disappears overnight, but because support, documentation, translation, editing assistance or previsualisation tasks are often places of situated learning. On a set, much knowledge is transmitted by sharing time, materials, decisions and problems with other people. If that first layer is thinned out, the transmission of knowledge that later leads to editing, camera, costume, sound, production design or visual effects is weakened.

Audiovisual work does not leave only a finished film. It produces tests, discarded materials, guide voices, references, metadata, scans and continuity materials that may later become assets, archives or data for new systems. That is why the labour debate and the debate about training cannot be fully separated.

United States union negotiations show that this is not an abstract discussion. The Writers Guild of America establishes that AI-generated content is not literary material and that a company cannot require a screenwriter to use AI in their work. It also requires disclosure when material handed over includes AI-generated content. The WGA summarises its protections here. The IATSE agreement for technical crews introduced limits on demands for prompts that might displace people covered by the agreement. Reuters reported those limits.

Those agreements cannot simply be transferred to every country as if they were a universal recipe. Audiovisual industries work with very different union frameworks, property regimes and public protections. Their logic, however, matters. The discussion should include negotiation, information, training for crews, a path into the profession and a distribution of savings that reaches the people who make the work possible.

Scale has a geography

AI is often presented as a global and placeless force. It is not. It depends on data centres, energy, connectivity, licences, dominant languages, cloud providers and capital for experimentation. It also depends on who has time to learn a tool while still paying rent, who can afford an enterprise subscription and who has advice to understand what happens to the materials they upload.

In El eternauta, the argument about budget viability matters. A tool can expand the repertoire of an Argentine production and allow a result that might otherwise not reach the screen. In India, AI-assisted dubbing responds to a specific problem of circulation between languages and audiences. That is not the same as generating a full film from a textual description. Placing both uses under a single label erases decisive technical, labour and cultural differences.

The opposite gesture should also be avoided. It is not enough to add a reference to Argentina, India, Nigeria, Korea or Mexico to a note otherwise organised around Hollywood’s conflicts. A global approach has to ask how production conditions change in each case. Who owns the catalogue. In which languages a model is trained. Which contracts protect performers and crews. What role public policies play. Which infrastructures reach schools, cooperatives, independent studios and communities that do not negotiate with a global platform as equals.

The difference between a production backed by Netflix and an independent crew is not only a matter of budget. It is a matter of capacity to decide over data, review contracts, store materials, challenge a claim and sustain a conflict. AI may lower some barriers to entry and raise other, less visible ones.

It is therefore useful to think of models as sociotechnical assemblages, not as black boxes. In anthropological research on algorithms, this view shifts attention from isolated software to the people, rules, data, infrastructures and expectations that make a system operate in a particular way. In cinema, it requires looking not only at the final shot, but at the set of relations that made it possible.

Archives also decide who appears

Generative systems do not learn from culture in the abstract. They learn from archives. And archives are not neutral repositories. They are accumulations of decisions about what is preserved, what is digitised, what is labelled, what becomes profitable and what falls out of circulation.

This question is central to cinema because audiovisual images build collective imaginaries. The reports of the Observatory of Diversity in Audiovisual Media offer a localised radiography, not a world sample. That is precisely why they are useful. They show how inequality operates within a concrete archive, that of Spanish fiction.

The 2026 ODA report, centred on films and series released in 2025, places racialised characters at 9.98 per cent of its sample of 2,015 characters. The same report observes that poverty often functions as a plot trigger, that luxury is normalised and that the middle class dominates the screen. On disability, it identifies increased presence, but also persistent associations with dependency, punishment or conflict. The 2025 report on fatphobia shows that 90.8 per cent of the characters analysed had normative bodies and that many characters with non-hegemonic bodies lacked their own storylines. ODA brings together its research and publications here.

It cannot be concluded that a model will mechanically repeat each of these patterns. But it can be said that, when a system is fed and evaluated through an unequal archive, it has incentives to present as normal what was statistically dominant. The risk is not only that an image generator produces an obvious stereotype. It is that a chain of decisions about casting, references, costume, character design, advertising or recommendation turns inherited biases into an efficient and apparently realistic aesthetic. In doing so, it can freeze historical exclusions, give them the appearance of statistical normality and make them seem like an objective description of the world.

This is where contributions from digital anthropology are especially necessary. Gisela Cánepa and Elisenda Ardèvol proposed studying visual technologies by attending to the ways in which they intervene in cultural difference, inequalities and possibilities for change. The point is not to ask AI to add diversity as if it were a colour filter. It is to ask who participates in the repertoires from which it learns, who defines what is plausible and what capacity represented groups have to intervene in the construction of their own image.

An archive can also be a place of resistance. Community projects, family collections, local film libraries and the archives of social movements contain experiences that large datasets often ignore or decontextualise. The problem is not only that they might be absorbed without permission. It is that, when they become data without relation, they may lose the names, production conditions, memories and bonds that gave them meaning.

A voice is not a database

The discussion about voices, faces and performances shows clearly how far the change reaches. A performance cannot be reduced to the sum of facial movements, vocal timbre and metadata. It is situated bodily work. It has time, training, working conditions, rights and a relation to the person who performs it.

The United States Copyright Office concluded in 2025 that copyright protection may cover human contributions in works incorporating AI, but not purely AI-generated material or material in which there is not enough human control over expressive elements. It also argued that, with currently available technology, prompts alone do not provide the control needed to attribute authorship. The full report is available here.

In performance, SAG-AFTRA agreements introduced requirements for informed consent and compensation for uses of digital replicas. They do not resolve every grey area, especially around synthetic performers and later uses of materials, but they establish an important reference. The WIPO Beijing Treaty, in force since 2020, recognises intellectual property rights for performers in audiovisual performances at international level.

A public SAG-AFTRA testimony on non-consensual digital replicas placed the problem in an even broader field than employment. For centuries, recognising a face or a voice was a basic way of recognising a person. Technologies capable of fabricating a plausible appearance break that elementary trust. A performer may see a version of themselves circulate saying something they never said, endorsing something they do not endorse or damaging in minutes a reputation built over years.

That warning connects with the concern Diana Palazón formulates from the craft. When voice, face, gesture and presence begin to be treated as reusable materials, the discussion about intellectual property also becomes a discussion about identity. Who can dispose of a person when their image can be separated from their will.

The conflict is not exhausted by law. Written consent within an unequal employment relation may not be free consent in a strong sense. When renewing a contract, being called back to a shoot or keeping a source of income depends on accepting a clause, consent needs additional guarantees so that it does not become a mere formality. That is why specific clauses are needed, limited to concrete uses, separated from ordinary hiring and understandable to the people who sign them. Authorisation to correct a dub should not automatically become permission to train a model, create a future campaign or reconstruct a performance in another work.

Whoever controls the archive controls part of the future

The lawsuits filed by Disney and Universal against Midjourney in 2025, and Warner Bros. Discovery’s later action, show another side of the conflict. The studios argue that the service allows protected characters to be generated without authorisation. Midjourney replies that training with copyrighted works may be protected by fair use. Reuters summarised Disney and Universal’s lawsuit and the later Warner Bros. Discovery lawsuit.

The dispute does not mean that major studios reject AI. Rather, it signals a struggle over who will be able to train, generate, license and commercialise catalogues. Those who already own franchises, characters, designs and archives with a consolidated chain of rights are in a privileged position to turn them into closed and licensable ecosystems.

That logic became explicit in the agreement announced between Disney and OpenAI at the end of 2025. It envisaged users generating short videos with more than two hundred animated, masked or creature characters from Disney, Pixar, Marvel and Star Wars within Sora. It did not include the faces or voices of real performers. The commercial idea was to move into a licensed environment something that had previously circulated informally: the possibility of inventing and starring in a short scene inside someone else’s fictional universe. Whether a person could appear with their own face depended on additional permissions and the platform’s specific design, but the characters remained under the control of whoever owned the franchise. Sora’s later closure ended the agreement before that model was consolidated, although the project remains revealing.

This can reduce some non-consensual uses. But it can also concentrate more cultural power. Fan art, parody and remix are part of the social conversation around works. When that conversation passes through a platform that fixes canon, permissions, moderation and terms of use, the creativity of audiences can become value for a brand without those audiences sharing ownership or editorial decision-making.

The case of Artificial, Luca Guadagnino’s film about the OpenAI crisis, points to a different tension. Amazon MGM Studios let it go a few months after announcing a major alliance with OpenAI. Amazon did not say the two decisions were related and there is no basis for presenting it as proven censorship. Even so, the episode exposed a problem that goes beyond the film: the same corporation may participate in technological infrastructure, funding, distribution and the critical circulation of stories about that infrastructure. Finally, Neon acquired the worldwide rights to the work and announced a 2026 release. Associated Press reported the acquisition.

There is no single regulation and no single answer

Rules are not only a technical manual. They also express different ways of distributing responsibilities between platforms, companies, performers, states and audiences. Regulation is moving unevenly. The European Union has approved a broad framework that includes transparency obligations. Article 50 of the AI Act requires certain generative systems to produce outputs detectable as artificially generated or manipulated. When a work is clearly artistic, creative or fictional, the information must be offered appropriately without preventing its display or enjoyment. The Regulation will apply generally from 2 August 2026. The official text is available on EUR-Lex.

The Chinese framework responds to different priorities. Its provisional measures on public generative AI services cover systems that produce text, image, audio or video for the population within mainland China. They include obligations on security, data quality and additional procedures for services with capacity to influence public opinion or mobilise people socially. The United States Library of Congress summarises the measures.

None of these rules is a global solution. United States collective agreements protect only those covered by them. The AI Act does not replace labour negotiation or resolve intellectual property conflicts on its own. The Beijing Treaty depends on national implementation. And many of the decisive choices continue to happen in private contracts, closed interfaces and terms of use that can change without public deliberation.

Cinema is still a way of making worlds

AI can help plan, test, translate, restore, improve accessibility and create scenes that previously would not fit within a budget. Denying that would not be useful. Nor would it be useful to accept that every saving of time automatically equals progress.

Cinema is a collective practice of attention. Building a set, testing a costume, waiting for a light, refining a sound or deciding when not to cut are technical actions, but they are also ways of producing worlds. Knowledge, affects, hierarchies, rights and possibilities of representation circulate there.

A costume professional with experience on major productions formulates this from another place on set. AI can allow large sets, impossible sequences or worlds that a production could not build with conventional human and economic means. But that expansion does not exhaust the value of a work. Art does not speak only through the scenic capacity it reaches, but through the soul that runs through it. Her concern points to a central tension. If production is governed by economic structures of power, stories will tend to be built according to what the market considers recognisable, saleable or efficient.

She also recalls that many substitutions announced as inevitable did not fully erase previous practices. Digital cinema did not eliminate photochemical shooting. Platforms did not end cinemas. Digital music did not make vinyl disappear. There may be a part of audiences that continues to seek works made by people, with a human presence that cannot be reduced to fast entertainment consumption.

Cinema with AI already exists. What matters is which conditions will become consolidated as it spreads. Clear contracts are needed for voice, face and training. Accessible training for temporary and freelance crews. Protection for entry-level positions. Traceability of changes. More diverse reference libraries. Tools that do not turn representation into a decorative add-on. And public policies that understand audiovisual media as culture, labour, memory and the right to imagine.

A generated screen can be spectacular. But an industry capable of protecting those who make it, listening to those who appear in it and leaving room for stories that do not fit the norm can be something more important. It can remain cinema when a costume decision, a silence in the mix or a sustained shot responds to a relation of work and care, not only to the speed with which a machine can produce its image.

Sources and further reading