AI Art's Real Threat Isn't the Algorithm. It's the Compute Monopoly.

The most revealing line in the old Open to Debate newsletter on whether AI will kill the creative arts is not one of the pros or cons. It is Jonathan Taplin’s warning, buried in his “yes” argument: “Only the biggest of big tech players will dominate generative AI because it requires massive amounts of computing power.” That was December 2023. Nearly three years later, the claim reads less like a prediction and more like a description of the present.

The debate itself, between Taplin, director emeritus of the Annenberg Innovation Lab at USC, and Rebecca Fiebrink, a professor at the University of the Arts London’s Creative Computing Institute, framed the question around authorship and imitation. The cons list in the newsletter is tidy: AI produces derivative art, it is an algorithm not an innovator, it jeopardizes creative livelihoods, and it can never express the human experience. The pros are equally tidy: technology democratizes artmaking, disrupts mediums, and pushes boundaries. The newsletter even cites the canonical examples: the Drake and The Weeknd deepfake song that went viral, the AI-generated Portrait of Edmond de Belamy that sold at Christie’s New York for $432,500, and the Hollywood writers and actors strikes over digital likenesses.

That framing was always incomplete. The real question was never whether a statistical model could produce something a human would call art. It is who gets to run the models at all. Taplin’s throwaway line about computing power was the actual thesis, and the intervening years have proven him right in a way that makes the rest of the debate feel almost quaint.

The compute chokepoint

Generative AI’s creative ceiling is not set by model architecture or training data quality. It is set by access to GPUs. Training a frontier-class image or music model requires clusters of thousands of accelerators, each costing tens of thousands of dollars, plus the data-center infrastructure, cooling, and power to run them. That is not a garage operation. It is a balance-sheet operation.

The economics have concentrated accordingly. The companies that dominate generative AI today are the ones that own or control the compute: OpenAI, Anthropic, Google DeepMind, and Microsoft, which has invested tens of billions in OpenAI and built out its own Azure AI infrastructure. Meta has its own clusters. Amazon has Trainium and its AWS footprint. The list of independent labs that can train frontier models without a hyperscaler partner is short and getting shorter.

Fiebrink’s argument in the debate was that AI can help more people participate in creative activities and enable new kinds of creative practices. That is true at the margin. A musician can use a free tool to separate stems or generate a drum loop. A writer can use a language model to brainstorm. But the tools themselves are rented, not owned. The creative act is mediated by a platform that controls the model, the pricing, the terms of service, and the data. The democratization Fiebrink describes is real, but it is a democratization of consumption, not of production. Anyone can prompt. Almost no one can train.

The derivative problem is a data problem

The other major claim in the newsletter’s cons list, that AI produces derivative art, has also aged into something more specific: a copyright crisis. The derivative nature of AI output is not a philosophical flaw, it is a legal liability. The training data for these models is scraped from the internet, largely without permission from the artists, photographers, illustrators, and writers whose work forms the statistical substrate. The Getty Images lawsuit against Stability AI, the class actions against OpenAI and Anthropic from authors and artists, and the ongoing fights over fair use in AI training are all downstream of this.

The debate in 2023 treated derivation as an aesthetic critique. In 2026, it is a regulatory and commercial one. The European Union’s AI Act, which entered into force in stages beginning in 2024, requires transparency about training data. The US Copyright Office has issued guidance that AI-generated works without substantial human authorship are not copyrightable. The result is a strange inversion: the more derivative an AI work is, the harder it is to own, and the harder it is to own, the less economic value it has for the artist who prompted it.

That is the actual mechanism by which AI threatens the creative economy. It is not that AI makes bad art. It is that AI makes art that cannot be cleanly owned, and the only entities with the legal teams and compute budgets to navigate that ambiguity are the platforms themselves. The individual artist is left with a tool that produces work they cannot fully claim, on infrastructure they do not control, under terms they cannot negotiate.

What the debate missed

Neither debater addressed the labor market mechanics that have actually played out. The Hollywood strikes of 2023 won some protections around digital likenesses, but the broader creative workforce has seen a different pattern. Illustration and stock photography markets have contracted as companies substitute AI generation for commissioned work. Game studios have laid off concept artists. Localization and dubbing work is increasingly automated. The newsletter’s cons list said AI “jeopardizes the livelihood of creatives,” and that has happened, but not because AI makes better art. It happened because AI makes cheaper art, and the buyers of art, the studios and publishers and ad agencies, are rational actors who optimize for cost.

Taplin’s deeper point, that the creative economy would be captured by the same platform giants that captured the music and news industries, has also played out. The streaming economy already decimated per-unit revenue for musicians. AI adds a layer of abstraction where the model itself becomes the distribution channel. An artist who uses a platform’s AI tools is feeding the model that competes with them. The platform gets the usage data, the engagement metrics, and the subscription revenue. The artist gets a byline that the copyright office may not even recognize.

What AI builders should take from this

For the engineers and researchers building creative AI, the lesson is not to stop building. It is to recognize that the bottleneck is not capability, it is concentration. The most valuable work in creative AI over the next few years will not be a bigger model. It will be infrastructure that lowers the compute barrier for independent creators: efficient fine-tuning, on-device inference, open-weight models that run on consumer hardware, and licensing frameworks that let artists train on their own work without surrendering it to a platform.

The question Open to Debate posed in 2023, whether AI will kill the creative arts, has a more precise answer now. AI will not kill art. It will kill the independent economics of artmaking unless the compute and the data are distributed more broadly than they are today. Taplin was right about the mechanism. Fiebrink was right about the possibility. The future of the creative arts depends on which of those two forces wins, and that fight is happening in data centers, not in galleries.