AIA cure for Baumol's disease?

Picture William Baumol in the 1960s, sitting with data from orchestras and theater companies across the United States. He notices something that should be obvious, but somehow nobody had formalized: a Beethoven string quartet still requires four musicians and roughly forty minutes — exactly as it did in 1820. No efficiency gain. No shortcut. The music simply is the labor. Now fast-forward sixty years, and ask whether a technology capable of writing scripts, composing music, and generating photorealistic faces has finally broken that centuries-old logic. That is the question Gillian Doyle and Sabine Baumann take seriously in their 2024 paper, "AI: A Cure for Baumol's Disease?" Their answer is careful and a little sobering. Baumol's cost disease starts with a simple model of the economy in two parts. In one part — manufacturing, computing, agriculture — machines and new processes steadily raise productivity. You get more output per hour of labor, which means costs can fall even as wages rise. In the other part — performing arts, live theatre, much of cultural production — the central inputs are specific kinds of human creative labor that resist mechanization. The problem is that wages tend to rise across the whole economy together. So the stagnant-productivity sector must pay those higher wages without the offsetting efficiency gains. The result, as Baumol and Bowen argued in 1966, is an inevitable, persistent escalation of real costs in labor-intensive creative activities. Above-average inflation is built into the structure of the work itself. Doyle and Baumann apply that framework directly to film, television, and news. In those industries, the defining inputs are screenwriting, acting, directing, editing, and composing — work that hinges on singular human judgment. The turn of phrase in a script. The musical cue that reframes a scene. The editorial instinct that shapes pacing. These are exactly the things that make content valuable, and they are exactly the things that resist straightforward automation. As Doyle and Baumann put it, citing Ruth Towse, there is little "flexibility to substitute capital for this input." Where mechanization has been possible — in some rendering tasks and basic transcription — it has not touched the core creative decisions that drive both value and cost. So the question becomes: what exactly can generative AI do, and does it reach those decisions? Quite a lot, it turns out — up to a point. Doyle and Baumann trace a historical arc from algorithmic support in journalism since the 1980s through the release of GPT-1 in 2018 and successive model improvements through 2023. Today, AI is being used across the entire production chain in film and television. In pre-production, algorithms analyze scripts for narrative structure, character development, and projected audience appeal to guide investment decisions. In post-production, AI assists with editing, rotoscoping, and color grading, with some reported reductions in editing time from days to a few hours. Specific examples include automated rotoscoping for "Everything Everywhere All at Once," AI-driven de-aging and face-swapping in "Here," and voice cloning with automated lip-syncing for dubbed versions of content. Studios also use AI to forecast box-office and streaming performance to shape marketing and release strategies. In newsrooms, automated writing is already routine for financial updates, sports results, and local reporting. AI tools handle formatting, distribution, and social media dissemination, and they support investigative data analysis. The paper catalogs all of this in detail, and it adds up to a genuinely significant transformation of workflows. But here is where Doyle and Baumann draw the crucial distinction: these tools are augmenting human labor, not substituting it. The paper describes generative AI as "creative rocket fuel," citing Stuart Connock — something that makes skilled workers more productive rather than rendering them unnecessary. Scripts generated by AI require human writers to ensure quality and coherence. Automated news copy requires extensive human intervention to guarantee factual accuracy. De-aging effects and photorealistic avatars require skilled technicians to supervise and refine. The pattern is consistent: AI supplies drafts, suggestions, and labor-saving artifacts, and then humans take over for the parts that actually matter. The paper does acknowledge real displacement risks. An International Monetary Fund study cited by Doyle and Baumann estimates that AI will affect forty percent of all jobs worldwide, with highly skilled roles particularly exposed. Some roles in media are already being reduced — certain script analysts, basic post-production editors, routine copywriters, and some background performers. At the same time, new roles are emerging: AI programming and management, advanced visual effects work, and data literacy and analysis. Daron Acemoglu and Pascual Restrepo's framework, which the paper invokes, holds that displacement effects can be counterbalanced by technologies that create new tasks. Whether that counterbalancing actually happens at sufficient scale and speed is an open question the paper does not pretend to resolve. What Doyle and Baumann do resolve clearly is the central question. Does generative AI cure Baumol's disease? Their answer: "Alas, not." The reasons are both practical and structural. On the practical side, AI tools introduce errors, incoherence, and bias that require sustained human oversight to catch. The productivity gains are real, but they do not eliminate the need for creative human judgment — they relocate it. On the structural side, significant legal and policy constraints limit how far AI can go. Copyright infringement and unauthorized training of models are a persistent industry complaint. Licensing and intellectual property protection are urgent and unresolved priorities. Regulators face mounting challenges around misinformation, transparency, and the risk of market dominance by a small number of AI providers. These constraints are not temporary friction that will eventually be engineered away — they reflect genuine tensions between how AI learns and what creative industries need to protect. The paper's conclusion is intellectually honest in a way that is easy to appreciate. Generative AI dents Baumol's disease. It automates time-consuming and repetitive tasks, reduces some production timelines dramatically, and frees human workers for higher-order creative decisions. That is not nothing. But the core of creative production — the work that makes film, television, and journalism actually valuable — remains heavily dependent on human labor. And because wages across the economy continue to rise together, the mismatch that Baumol identified in 1966 persists. Cultural industries are still paying higher wages for work that still cannot be fully mechanized. The disease has a new treatment. It does not have a cure. What that means going forward is genuinely consequential. For creative firms, the agenda is how to integrate AI tools while protecting intellectual property, managing new liability risks, and developing the specialist skills — AI management, advanced visual effects, data analysis — that these workflows now require. For policymakers, the priorities are misinformation and transparency, copyright and licensing reform, and preventing market concentration among AI providers. For workers, the transition involves real friction: some routine roles diminish while new specialist roles emerge, and the speed of that shift matters enormously to the people caught in the middle. Baumol's insight, it turns out, was more durable than almost anyone expected. Creative human labor resists full automation — not because we haven't tried hard enough, but because the qualities that make creative work valuable are bound up with the fact that humans made it, with all the judgment, instinct, and irreducible decision-making that entails. That is a policy challenge. But it is also, quietly, a statement about what culture actually is. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.