The Rent of Taste: Generative Systems and the Reproduction of the Artistic New

The question of whether generative systems will replace screenwriters, editors, and directors is usually discussed as a question of employment. On closer examination it turns out to be far broader. It concerns how a culture discovers the new, passes on mastery, and decides what deserves attention. If art is treated as an industry in which some performers are swapped for cheaper ones, the main point is lost: in the creative professions, production, training, and the formation of standards of quality are interwoven more tightly than in any industrial manufacturing. Automation therefore affects not only jobs but the entire environment in which people capable of creating something unpredictable come into being.

The initial situation looks almost absurd. Professionals who spent years learning to understand why a line sounds alive, where a scene loses its rhythm, and at what moment an emotion turns false are left without their usual work and begin to evaluate texts written by models. They assign scores, explain mistakes, and compare variants. By doing this well, they improve the very systems that are shrinking their own market. From the outside it is easy to call such a decision a betrayal of guild solidarity. But for a person who has nothing to pay the rent with this month, the choice is a matter not of philosophical principle but of whether there will be anything to pay with at all. If moral judgment is placed above this circumstance, the heaviest reproach falls on the most vulnerable participant in the chain. Those who gathered the training data without the authors’ consent, who organized the market for expert annotation and profit from it, meanwhile almost escape moral condemnation.

The mechanism for which such specialists are needed is called reinforcement learning from human feedback. A pretrained language model is good at predicting the continuation of a text, but it does not itself know what makes an answer accurate, witty, appropriate, or truthful. People rank and correct its answers. Their preferences are turned into a reward model, and the system gradually shifts toward what people like. At this stage the machine’s fluency becomes what is perceived as human. For creative tasks, it is precisely people of developed taste who handle this best. The system cannot create taste on its own, so it rents it, and usually at a price below what that taste cost when a normal market for authors’ labor existed.

Here a second layer of the problem appears. At first the models were trained on the past work of screenwriters, artists, and journalists, collected from publicly accessible sources without their permission. Now they require the current judgment of those same professionals. This time consent is sought and even payment is offered, but the terms strip the work of everything that made creative labor meaningful. It has no authorship, no credit, no royalties, and no right to the improvement it creates, and a non-disclosure agreement does not even allow the person to say what they were doing. The contribution dissolves into a model that bears no trace of its author, even though the model competes with that author in the author’s own profession.

This arrangement has a predecessor. Around automated systems there has long existed an invisible human labor: people label images, moderate content, and correct errors so that the client sees only a smooth result and does not know about the human involvement. Such work is usually poorly paid, offers no social guarantees, and can be terminated at any moment without explanation. What is new about the present situation is that this niche is now occupied not by anonymous workers from poor regions but by credentialed specialists from one of the most visible industries in the world. The conditions, however, remain the same: projects are closed without warning, access to working systems vanishes silently, and qualified experts are supervised by people who have barely finished their studies. Only the level of skill being consumed in this work has changed. For the most complex creative decisions the system needs a highly trained person, and a market has arisen in response to this need in which such a person’s judgment is bought on the terms of invisible labor.

From the old practice in which laid-off specialists were obliged to pass on their knowledge to their cheaper successors, the present situation is distinguished by two circumstances that work in opposite directions. The person who trains the model is usually not compelled to do so by an employer: many come on their own, because the familiar market has already closed to them. But the one they are training will not become a colleague, will not build a career, and will not one day pass experience on to the next generation. It is a system that can be copied without limit, that does not tire, does not join a union, and does not age. The specialist is preparing not a successor but a replacement that absorbs the specialist’s function without continuing the specialist’s professional path.

This also reveals the limits of the comparison with crossing a picket line. The protection that creative unions won from the studios regulates relations between authors and their employers, but not relations between a freelance author and an annotation platform, which does not fall under any collective agreement at all. The comparison is right in the sense that the sum of individual decisions does indeed strengthen the technology that organized labor tried to restrain. But it is wrong because there is here neither a strike, nor a specific employer against whom the struggle is waged, nor a structure of solidarity that could be broken. The work takes place in a legal vacuum, and it is this vacuum, rather than the personal choices of individual people, that constitutes the essence of the problem.

The deepest contradiction concerns where what is now being transferred to the models comes from. Taste and professional judgment are not acquired by instruction. They take shape over years in writers’ rooms, on film sets, in editing suites, in work with mentors, through mistakes and through the experience of observing what works before a real audience. Today’s professionals went through this path inside a functioning industry. Now they are passing what they have accumulated to systems whose development coincides with the disappearance of the very conditions in which this experience could be acquired. If performing work continues to shrink, it is unclear where the next generation will learn to feel dialogue, rhythm, and composition. One cannot learn this by evaluating other people’s attempts under a non-disclosure agreement. A feel for a line of dialogue appears when you spend years writing lines that live actors speak before live audiences.

Labor-market data show that the creative professions are not disappearing wholesale but are stratifying. Strategic positions such as creative directors and producers are so far holding up noticeably better than performing roles, which can be described, evaluated, and reproduced. The likely outcome therefore looks not like the disappearance of human creativity but like its concentration. A small group of people at the top of the value chain will manage systems trained on the judgment of a far larger number of people pushed out of the middle of the profession. The models, in essence, are spending a reserve of judgment accumulated by an industry that they are themselves undermining, and there is as yet no mechanism for replenishing this reserve. A culture can live for quite a long time on accumulated competence, but it cannot live indefinitely on a reserve that is no longer replenished. If models increasingly learn from their own outputs instead of from fresh human judgment, then the quality of the systems themselves will probably suffer in time as well.

Late antiquity is often recalled here, when craft production supposedly held back technological and economic development for a long time. Historians dispute the thesis of Roman technological stagnation: Rome created concrete, aqueducts, water mills, and advanced mining technology. The comparison is therefore useful not as a general diagnosis of an epoch but as a description of a specific mechanism. It is best illustrated by the molded pottery known as terra sigillata. It was made in stamped molds, marked with the workshop’s stamp, and shipped throughout the empire. The stamp guaranteed uniform quality and appearance, and the customer cared more about recognizability than about novelty. A major film company turned into such a workshop would work in a similar way, but with one essential difference. A Roman workshop depended on materials, the number of craftsmen, and transport capacity, whereas a generative system spends almost nothing on producing a new variation. Uniform production ceases to look uniform, because inside an unchanging template the outward design can be varied endlessly.

From this follows a refinement of the common formula according to which the model simply averages. Probabilistic learning does not mean that every result will be banal, and learning from human preferences does not mean that the literally average rating always wins. A more accurate hypothesis is a different one: the system can produce unusual forms while retaining an unchanging requirement to adjust itself to the recipient. In ten years millions of feature-length films may differ in their actors, visual style, era, and genre. But if they are united by one logic of conflicts, one trajectory of character development, the same distribution of emotional scenes, and similar ways of resolving contradictions, then only the outward design will be diverse, while the artistic thinking will remain homogeneous.

Personalization of a work appeared long before generative systems, but its earlier forms are arranged differently. Blade Runner has a theatrical cut, a director’s cut, and a Final Cut, and they differ not only in running time but in how the viewer perceives the main character. The theatrical and director’s cuts of Donnie Darko arrange the semantic emphases differently. The 1985 film Clue was shown in different theaters with three different endings. Earlier still, in 2000, Stephen King published his novel The Plant in installments online and promised to continue the work if a sufficient share of readers paid for each release. When the share of paying readers fell, publication was halted, and the novel remained unfinished. The reader by then already governed the fate of the work economically, but did not receive a version of the novel of their own: they decided whether there would be a continuation, not what it would be like.

These examples show a sequence of stages. First the author creates a single work and receives the public’s reaction. Then the industry permits several authorial interpretations of a single conception. After that interactive forms appear, in which the viewer chooses the development of events. The next possible stage is arranged differently: the work itself decides for the viewer what it should be, and continually refines this decision using accumulated data about the viewer’s reactions. The director’s cut remains a choice among several authorial statements. The adaptive version fits the statement itself to the consumer. What is new here is not the aspiration to take the audience’s wishes into account, since art has always existed in a relationship with patrons and audiences. What is new is that feedback becomes continuous, measurable, and built directly into the creation of artistic content. The long-standing conflict between authorial intention and market demand is turned into an automated procedure in which demand determines not only which works survive but also what they will be like.

Such a system should not be described as a machine for soothing, set against truly disturbing art. The opposition looks too simple. If the system discover