Cory Doctorow argues that today’s AI boom is driven less by proven profitability than by a compelling story for investors and executives: that companies can replace workers, overcome the limits of mature markets, and keep growing. He does not dismiss machine learning itself. He sees statistical inference as a genuinely useful technology, but contends that large language models lack understanding, face diminishing technical returns, and are being applied far beyond the tasks for which they are economical or reliable. His greatest concern is that companies will eliminate skilled jobs before the economics unravel, destroying institutional knowledge that will be difficult to rebuild.
Doctorow separates AI as a technology from AI as a social and financial narrative. He says the narrative attracts some wealthy executives because it imagines a world with fewer troublesome human beings, while attracting more cynical investors because it is an effective product pitch whether or not the technology fulfills its promises. In both cases, AI appears to reduce the power workers derive from knowing how to do their jobs.
He connects this with the incentives of mature technology firms. Growth companies receive high valuations based on expected future earnings and can use their stock to acquire companies and recruit employees. Once growth stops, those valuations become difficult to justify. After companies exhaust market expansion and the ability to squeeze existing users, they need a new growth story. Doctorow presents AI as the latest and most persuasive of these stories.
Doctorow accepts that a conscious machine may someday be physically possible, but argues that current systems are not on that path. Modern generative AI finds statistical relationships in enormous datasets and predicts likely outputs without constructing an explanatory model of the world.
He illustrates the distinction with human conversation. A person and a phone’s autocomplete may both predict a familiar sentence, but the person can use an understanding of the speaker when confronted with something unprecedented. A statistical system can only search for patterns resembling what it has already encountered. This limits its reliability in unfamiliar, high-stakes situations.
He also argues that model development has entered diminishing returns: newer models require much more money and computation while delivering smaller improvements than earlier generations.
Doctorow portrays AI as a business with extraordinary capital and operating costs but modest revenue relative to spending. Unlike rail infrastructure, which can continue producing value after construction, AI providers must keep paying substantial inference costs and repeatedly train new models because customers can switch quickly to a better competitor.
He questions claims that inference is becoming cheap. Drawing on journalist Ed Zitron’s interpretation of leaked OpenAI financial information, he suggests that free or subsidized usage may be classified as marketing. If true, moving a cost to another accounting category would not improve the underlying economics.
The financing structure also appears brittle to him. Nvidia and a small group of model and platform companies account for an unusually large share of the stock market, while suppliers may finance customers that then spend the money on those suppliers’ products. Data-center projects depend on layered financing whose later, larger investors often demand escape clauses. Delays, energy-price changes, connection problems, or equipment shortages could therefore cause funding to vanish across many projects at once.
Doctorow says bubbles usually look inevitable only in hindsight. The immediate trigger could be a financial margin call, higher interest rates, a failed data-center project, an investor redirecting money to a more urgent national need, or turmoil inside a leading AI company.
Whatever the trigger, the more important weakness is structural: the industry requires continuing investment and replacement of expensive, rapidly obsolete hardware while lacking a demonstrated route to profits large enough to support that spending. Specialized chips and frequently changing infrastructure requirements may also limit the reuse value of failed data centers.
A crash would not eliminate statistical machine learning. Doctorow expects useful systems, open-weight models, and efficiency research to continue. He points to DeepSeek as an example of optimization that reportedly achieved strong results with far less money and lower-end hardware, briefly challenging assumptions about demand for top-tier chips.
He expects the technology to settle into appropriate niches rather than dominate every activity. His chess example illustrates the distinction: an LLM asked to play chess may make illegal moves because it lacks a chess-world model, while a conventional chess program can enforce the rules efficiently. An LLM may nevertheless help a skilled programmer build that conventional program. The lesson is to use each technology for the task it suits.
Many current AI products, he predicts, will disappear when subsidies end. A service can be popular at a nominal price while being uneconomical at its full cost. As users leave, fixed data-center costs would be spread across fewer customers, potentially creating a further price-and-demand collapse.
Doctorow recommends describing automation precisely: the danger is not that a capable AI inevitably takes a job, but that an employer assigns the job to an inadequate system. That wording helps customers understand that they may receive worse films, medical decisions, software, or other services. Workers and consumers can then become allies rather than opponents.
He argues that individual persuasion is insufficient when management controls tool choice and working conditions. Organized labor provides the leverage to set limits, as demonstrated by screenwriters who won contractual AI protections.
His strongest concern is the loss of process knowledge, the unwritten practical understanding accumulated among workers. It includes workarounds, relationships, safety practices, customer knowledge, and opportunities that no manual or intellectual-property file fully records. If firms dismiss experienced people and later turn off failed AI systems, simply rehiring may not restore that knowledge. Former staff may have retired, retrained, moved, or left the workforce, and rebuilding their collective expertise could take decades.
Doctorow calls AI both a new-market story after ordinary enshittification runs out of room and an especially powerful mechanism for further squeezing workers and customers. Systems can use behavioral data to divide people into fine-grained groups, experiment continuously, estimate the lowest wage each worker will accept, or find the highest price each customer will pay.
Unreliability can itself provide cover. A recommendation system might favor a higher-commission product and, if challenged, characterize the result as an accidental hallucination. Doctorow worries that firms could combine targeting with this plausible deniability, applying more manipulation to users judged less likely to notice.
His conclusion is not that all AI is worthless. It is that a useful technical method has been inflated into an unsustainable financial and managerial project. The sooner that distinction becomes clear, he argues, the less skilled labor and institutional capacity will be destroyed along the way.