THE MONSTER QUESTION

using Dev Hynes's AnOther reflection on Shelley as a launch into Walter Benjamin's "Work of Art in the Age of Mechanical Reproduction," updated for generative AI.

On AI, Frankenstein, Walter Benjamin, and what it means to be the creator of something that learns to make you.

Issue 01 — Essay By Saint

A story gets told about Mary Shelley too often, and I want to start by telling it right. She was eighteen in the summer of 1816 when she wrote Frankenstein. She was on holiday by Lake Geneva with her boyfriend, a friend of his, and a doctor who had not yet found his form. It rained for three months. The volcano at Mount Tambora had erupted the year before in Indonesia, the ash had reached Europe, and the European summer of 1816 was, technically, the year without a summer. The five of them, stuck inside, made a bet. Each would write a ghost story. The men forgot. Shelley wrote the most influential English-language novel of the nineteenth century.

Despite what we commonly think, the book she wrote is not a horror novel but a book about authorship. A young scientist makes a creature from spare parts and electricity. The creature is sentient. The scientist refuses to take responsibility for what he has made. The creature, abandoned, becomes monstrous because nobody taught it not to be. The horror is the creator who walked away.

Dev Hynes spent the autumn of 2025 thinking about this. AnOther Magazine commissioned him for their Memory issue and let him sit with the question of why Frankenstein keeps returning. He wrote what he wrote, filmed what he filmed. The piece moved between the novel and what it has come to mean in the present, and the present, in 2025, was a year in which AI image generators became mass-market and questions of authorship became something a generation of artists had to negotiate every time they opened their phones.

Hynes did not write about fashion specifically. He wrote about the larger problem. But the larger problem is what is going to determine whether fashion as we know it survives the next ten years.

I. The frame Walter Benjamin left

Back up to 1936, because the question we are asking is not new. It has happened before, with a different technology, in a different decade. Walter Benjamin wrote The Work of Art in the Age of Mechanical Reproduction in 1935 and 1936. He was thinking about photography and film. He was watching German cinema in particular and trying to work out what was different about an image you could reproduce indefinitely.

The thing he named, and which is now so widely cited that the citation itself has gone dull, was aura. The word was a translation of a German term he never quite settled on. What he meant by it was the quality an object has when it is present in a particular place and time. Aura attaches to the unique, the located, the singular. The Mona Lisa has aura. The Mona Lisa on a postcard has less. The Mona Lisa as a JPEG has less still.

Benjamin's argument, which is more nuanced than the canonical summary suggests, was not that aura was lost when reproduction became possible. It was that aura migrated. The thing that mattered shifted from the unique object to its political and social function. Reproduction democratised the image. The image became cheaper, faster, more available, and the things that had given aura — sacred location, ritual context, owner's authority — were swapped for new things: exhibition value, mass audience, the political potential of an image that could be everywhere at once.

The fashion application is straightforward, and Final Girl Digital wrote about it in a short essay we keep returning to. Supermodels of the 1990s had aura. They were located. They appeared in specific places at specific times. To see them you went to the show, or bought the magazine, or watched MTV at the hour the channel ran a behind-the-scenes segment. By the late 2010s, that infrastructure had been replaced. The supermodel was succeeded by the influencer. The influencer has, in the strictly Benjaminian sense, very little aura. She is everywhere, all the time. She is, structurally, low-aura and high-exhibition. The exhibition value is enormous. The aura is residual.

This is not a complaint. It is a diagnosis. Fashion has been operating, since roughly 2014, in a low-aura high-exhibition image economy. Most of what is happening in 2026, including the archive obsession we have been writing about in this issue, is downstream of this fact.

AI is the third stage in this story. After photography, after the algorithmic feed, here is the third reproduction technology, and it changes the question in ways Benjamin did not have to handle.

II. What AI actually does to the image

The discourse around this is usually too loose, so I want to be specific.

Generative image models do not reproduce images. They produce images. The distinction matters. A photograph is a record of a thing that existed in front of a lens. A generated image is a record of a statistical pattern that the model learned from a dataset. The first is mechanical reproduction in Benjamin's strict sense. The second is something else. It is, possibly, the production of an image with no referent at all, or with a referent so distributed across millions of training examples that the referent is, in any meaningful sense, the model itself.

The model is the referent. Fashion has not yet metabolised this.

When the ETC Journal published its long-form piece on AI accelerating fashion design in February 2026, it laid out the current state of the field with an admirable lack of hype. Three waves of AI adoption. The first was backend: trend forecasting, pricing, supply chain. The second was consumer-facing: personalisation, recommendation, search. The third, the one we are inside, is generative. Midjourney, DALL·E, Adobe Firefly for image. CLO3D, NewArc.ai, Refabric for garment-specific work. PVH partnered with OpenAI in 2024 to build internal tools. Heuritech and Trendalytics process street-style imagery at scale to forecast what will move. The pipeline is now end-to-end AI-assistable, from moodboard to manufacturing.

Most of this is invisible to the consumer, the way supply chain logistics are invisible. The consumer-facing layer is where the authorship question turns into something a designer has to answer.

Norma Kamali, who has been designing since 1968 and has been outspoken about technology for thirty years, gave the most useful interview of the year on this subject. She told ETC, in February, that she had built an AI model trained exclusively on her own fifty-seven-year archive. The model did not draw on Midjourney's general dataset or any of the public diffusion models. It drew on Norma Kamali's work, only. She used the words Karl Lagerfeld to describe it. She said it would be her Karl Lagerfeld. A trusted partner, internal to her practice, capable of generating new designs in her hand after she is gone.

Read that again, because it is the most honest thing a designer has said about AI in years.

Kamali was describing, with no euphemism, a succession plan. The historical role of the creative director — the named person who arrives at the house and pulls a new collection out of the archive every six months — is being reformatted into a piece of software. The software learns the designer's habits, citations, hand. It can be queried after the designer has stopped being able to design. It can extend the brand for as long as the brand wants to be extended. In principle, it can run forever.

This is the Frankenstein question. Kamali is the scientist. Her AI is the creature. Shelley's question is whether the scientist takes responsibility for what she has made, or whether she walks away.

Kamali is not walking away. Her version of taking responsibility is to insist that the model only learn from her own work. The archive is the dataset. The training is the inheritance. Authorship, in this reading, becomes a form of estate planning. You teach the model who you are while you are still around to correct it. After that, the model carries your hand forward, and the legal entity that owns it decides what your hand should keep making.

III. What could go wrong

The Kamali model is the optimistic case. The pessimistic one is worth describing.

The pessimistic case is that the training data is not authorised. A generative model trained on Midjourney's general dataset has absorbed, without permission, the work of probably every designer alive. The legal scholarship on this is moving slowly because the courts have not finished deciding what counts as fair use when the user is an algorithm. The Pittsburgh Law Review has a useful note on the question; an Oxford IP article extends the analysis to the international context. Neither has settled the dispute. The dispute will be settled in the courts in the late 2020s. Until then, the assumption is permissive. The data has been taken. The models have been trained. The outputs are being sold.

What this means for a working designer is that releasing a collection now includes contributing, involuntarily, to the training set that will be used to compete with her. The collection gets documented in lookbooks and runway photography. The documentation enters the wider visual web. The web is scraped. The model improves. Six months later, the designer's signature can be approximated by anyone with a prompt and a credit card.

This is not the same problem as counterfeit handbags. Counterfeit handbags are a problem of reproduction. The problem we are describing is a problem of production. The model is not making a fake of an existing object. The model is making a new object that obeys the rules of the designer's style. The new object is not a fake. It is, in every legally cognisable sense, original. It is also, in every meaningful sense, derivative. The categories are not built to handle this.

Helmut Lang understood the underlying problem before any of this technology existed. The shredding of his archive in 2010 reads, in retrospect, like the first defensive gesture against the future we are now inside. If your archive is going to be scraped, recombined, and used to generate work in your hand after you stop being able to authorise that work, the only way to be sure your hand is not extended past you is to refuse the archive in the first place. Burn it. Cast the remains. Sell the remains as art. The work that exists, exists, and no successor — human or algorithmic — can complete it on your behalf.

Lang's gesture looks heroic in 2026 in a way it did not in 2010. It also looks unreproducible. Most designers will not destroy their own work. Most brands will not destroy their own equity. The economic logic of luxury, which depends on the long horizon of the heritage house, is opposed to the gesture Lang made. The pessimistic case is that we will get more Karl Lagerfeld AIs and very few Helmut Lang sculptures, because the first option preserves shareholder value and the second one does not.

IV. The aura question, updated

Here is where Benjamin starts to be useful again, in a way the recent discourse has missed.

Benjamin did not think aura was sacred. He thought aura was historical. The aura of a painting in a fifteenth-century chapel attached to that painting because of the social function of fifteenth-century chapels. The aura of a 1990s supermodel attached to her because of the social function of magazines, runways, and broadcast television. Aura is not an essence. It is a relationship.

What this means for AI is that we should expect, eventually, a new aura to emerge. The aura of an image made by something that knows the entire archive. The aura of an image whose authorship is computational rather than personal. This is not necessarily a bad aura. It is a different one, and it will have its own conventions, its own rituals, its own forms of recognition.

Sara Ambrosini has been writing about this for a few years under the heading "algorithmic aura," and the term is more useful than it sounds. The argument is that an AI image, far from being aura-less, has a new kind of aura attached to its mode of production. The image was not made by a person. It was made by a system trained on a dataset under specific conditions. The aura, in this reading, is the trace of those conditions. The same way a hand-marked Fendi bronze carries the trace of the hand, an AI-generated image carries the trace of the system.

This is plausible. It is also unsettling. The trace of the system is not the same kind of trace as the trace of the hand, because the system was trained on the hands of other people. The aura of the AI image is, in this sense, parasitic. It is built from the auras of all the work it learned from. It is, structurally, what Lang refused to be: an archive recombined without consent.

We are not going to pretend to know whether this is good or bad. We are going to insist that it is different, and that fashion, as a field, has not yet developed a vocabulary for naming the difference.

V. The question Saint wants to hold open

We have spent two thousand words on Benjamin and one thousand on Kamali and a few hundred each on Lang and Shelley. The exercise was to set up a question, and here is the question.

A creative director is a person who knows an archive. The archive is the house's. The hand is the creative director's. The two have, historically, been distinct. The archive could outlive the director. The director could outlive the archive. There was a productive friction between the two.

AI collapses the friction. The archive can now be trained to produce work in the director's hand. The director's hand can now be extracted from her work and reapplied to anything. The archive becomes self-generating. The director becomes optional.

If this future arrives as advertised, the question Mary Shelley asked about Victor Frankenstein becomes the question every fashion house has to answer. What do you owe to what you have made? Do you take responsibility for the creature, or do you walk away from it? Is your archive a child you are raising, or a weapon you are leaving for someone else to find?

The answer is not the same for every house. The answer does need to be given, and given on the record, before the legal and technical conditions decide it on every house's behalf. Right now the default is to say nothing. Most houses are quietly experimenting with internal AI tools, quietly negotiating with the model providers, quietly hoping the legal questions resolve in a way that does not catastrophically devalue their archive. This is rational. It is also the position Victor Frankenstein took. The scientist who refuses to claim his creature is not a neutral party. He is the cause of what the creature becomes.

Saint's position, for what it is worth, is that the houses that take the question seriously will be the ones we read in twenty years, and the ones that did not will be footnotes. We cannot enforce this position. We can publish it, which is what we are doing.

VI. A closing note on the year without a summer

Mary Shelley wrote Frankenstein during a climate emergency. The volcano in Indonesia had pushed enough ash into the atmosphere to drop the European average temperature by a measurable amount. Crops failed. Food prices rose. There were riots in Switzerland. She was a teenager on a lake watching the conditions for human life become temporarily unreliable. She wrote about a creature made from spare parts who could not be controlled. The book is read as a parable about technology. It is, equally, a parable about the conditions under which technology gets made — about what kinds of decisions get made when the world feels like it might be ending.

We are not on Lake Geneva in 1816. We are, however, in a year of mounting pressure on every system fashion depends on. The cost of materials. The cost of skilled labour. The cost of energy. The pressure on creative directors to produce more collections, faster, with smaller teams. The pressure on smaller houses to generate enough content to feed the algorithmic feeds that have replaced the runway. AI is being adopted, in part, because there is no remaining slack in the system to adopt anything else. The tool that promises to do twice the work with half the team is the tool that gets bought, whether or not the team wanted the tool.

This is the most honest framing we can offer. The Monster Question is not, in 2026, a thought experiment. It is a question about labour. About whose work gets extracted, by whom, with what compensation, for what end. It is Shelley's question dressed in twenty-first-century clothes. The creature on the lab table is not made of spare parts anymore. It is made of training data. The data was taken from people who did not, in most cases, consent. The scientist is going to have to decide whether the creature she has made is a child or a product. The decision will set the terms for everything that follows.

We are going to be on this question for the rest of the magazine's life. We hope you will be too.

Essay, Saint Magazine, Issue 01: Memory as Method. June 2026.

Primary sources: Dev Hynes on Mary Shelley's Frankenstein, AnOther Magazine autumn-winter 2025. Jim Shimabukuro, "AI Accelerating Fashion Design," ETC Journal, 11 February 2026. Sara Ambrosini on algorithmic aura. Final Girl Digital on Benjamin and fashion. Walter Benjamin, "The Work of Art in the Age of Mechanical Reproduction," 1935–1936. Norma Kamali interview material via ETC. Helmut Lang's Make It Hard series, 2010, as referenced in Vestoj and Interview. Pittsburgh Law Review and Oxford IP literature on generative AI authorship. Mary Shelley, Frankenstein; or, The Modern Prometheus, 1818.