Garrison Lovely wants people to stop using "AI" for two very different things. One is a tool like AlphaFold, which predicts the shapes of proteins. The other is a deliberate effort to build machines that can do nearly any job a person can do. Lovely, a journalist and author, calls the second one "the obsoleting project." He argues that it can be stopped, and that the medical breakthroughs its backers promise can be pursued more directly without it.
Lovely made the case on The Cognitive Revolution in an episode published on September 29, 2026. The occasion was his book Obsolete: The AI Industry's Trillion-Dollar Race to Replace Us—and How to Stop It, which OR Books published together with The Nation. The conversation mixed sympathetic questions with the strongest counterarguments to his position.
Two kinds of AI
Lovely begins with the companies' own words. The discussion noted that OpenAI defines AGI (artificial general intelligence) as something better than humans at most economically valuable work. The OpenAI Charter does describe AGI as highly autonomous systems that outperform humans at most economically valuable work. The same charter commits the company to making sure AGI benefits all of humanity. Lovely also cited Dario Amodei, the Anthropic CEO, whom he quoted as saying that AI is not like other technologies because it is "a general substitute for human labor."
To Lovely, that goal is different in kind from ordinary AI. He said he sees it as inevitable that people will keep using deep learning, the technique behind today's AI models, to build systems that do new things. The mistake, he said, is treating AGI as if it were all of AI, just because the companies chasing AGI have been best at building capable, autonomous systems. "We could just be like, we have all these diseases and medical problems," he said, and ask whether "alpha-fold type systems" could be built for specific problems.
AlphaFold shows what that narrower kind of AI can do. In a blinded test in 2020, DeepMind's AlphaFold2 predicted protein structures with a median backbone error of 0.96 ångströms, compared with 2.8 for the next-best method. Those figures are the median backbone accuracy (Cα root-mean-square deviation at 95% residue coverage) measured on CASP14 protein domains. Its authors describe uses such as supplying candidate structures for lab analysis and helping interpret electron-microscopy maps. The paper also notes where it falls short: accuracy dropped for proteins with few known relatives, and for proteins whose shape depends heavily on other protein chains.
Lovely was skeptical of the bigger vision: build a superintelligence, use it to "solve intelligence," then use that to solve everything else. He noted that this was DeepMind's mission statement for a while and called it "the ultimate techno solutionist fantasy." He accepts that, technically, such a machine could invent transformative technologies. But he said you cannot solve ethics, ideology or politics in the same way. He called the idea a kind of "revenge of the STEM people on the humanities": skip history and political theory, build the machine that is smarter than everybody, then ask it what to do.
The book's introductory excerpt frames the economics this way: earlier machines replaced particular tasks, while the new goal is to manufacture labor itself. Workers have to be raised, educated and persuaded to work, which limits how much any owner can produce. Loosening that limit would shift bargaining power from workers to the people who own the machines.
Why the leaders keep racing
The conversation turned to what the people running frontier AI companies believe they are doing. One view put to Lovely was that they are not well described as simply trying to get as rich as possible. Lovely agreed. In his account, the founders of the AGI companies were "chasing mission" and worked in obscurity until their progress drew profit-seeking investors, who then poured in huge sums. The leaders, he said, are still driven by some mix of idealism, a sense of inevitability and "megalomania, you know, messiah complex."
Investors add pressure of their own. Lovely pointed to Sam Altman's firing and return as OpenAI's CEO. He said he is skeptical of accounts that Altman did not really want to come back, and he named Thrive Capital as a big player in restoring him. Once you have invested billions, he said, safety-minded people replacing the CEO who presided over a "meteoric rise" is "just not going to fly."
He named Altman, Amodei, Demis Hassabis and Elon Musk as people he thinks are motivated more by power and by being "a great man of history" than by money, while noting that each is unique and hard to generalize about. He acknowledged a counterexample: OpenAI co-founder Greg Brockman's diary entry asking what would take him to $1 billion, which Lovely called "a pretty crazy thing to write" in the diary of a nonprofit he co-founded. Overall, Lovely said, it is "about being in the room where it happens," and he said Eliezer Yudkowsky told him something similar when Lovely interviewed him in 2023.
Lovely also mocked the plan of simply making the superintelligence and asking it what to do, calling it "the ultimate like question mark, question mark, question mark profit thing."
He thinks this combination is dangerous. He argued that introducing a universal labor-replacing machine safely and democratically would be hard under any circumstances. It is worse, he said, when it happens as fast as possible, with minimal regulation, in a country cutting its social safety net or adding work requirements to Medicaid, while other countries have no chance to tax the companies putting their people out of work.
Why he thinks it can be stopped
Lovely's optimism rests on how narrow the project is. He said it is "really just a handful of companies in two countries," and only one of those countries has truly been at the frontier since ChatGPT launched. It costs enormous sums and depends on the most advanced AI chips, with single companies controlling different parts of the chip supply chain. Whether it can be stopped forever, he said, he does not know.
He was careful about what he is and is not proposing. He said he is not against ever building AGI, but it should happen only "with strong public buy-in and a scientific consensus so it can be done safely." He sees the current backlash against AI as intense and "undiscerning," trending toward banning AI across the board. He thinks that would be very hard to do, especially with open-weight models (models whose files anyone can download and run). His aim instead is to stigmatize the obsoleting project because it is risky and undemocratic, and to "save the baby from the bathwater" by keeping specialized systems.
The case for automation
The discussion also put the strongest counterargument for automation. The argument was that AI saves a great deal of tedious work and has, at least counterfactually, replaced people who would otherwise have had to be hired, which was considered good so far. It pointed to farming: not long ago nearly everyone tilled fields, and now a small share of the population feeds everyone, which almost everybody considers overwhelmingly good. Could AI make each of us an executive of our own AI corporations? Or finally deliver the short working week that the economist John Maynard Keynes predicted about a century ago?
Lovely joked that "the faster AI goes, the more I work." Then he agreed with the premise. Automating labor, he said, is what let humanity go from nearly everyone being very poor to rising living standards after the Industrial Revolution, and he does not want progress to stop. "But trying to automate all labor is pretty different," he said. He admitted that where to draw the line is "not super obvious." His answer is to freeze frontier development now and not resume it without public buy-in and safety.
Cures for all
Lovely's alternative starts with medicine. He argued that AI drug discovery, left to the current market, will follow the incentives of monopoly patents. Companies will look for drugs that make money: treatments for chronic conditions or male-pattern baldness rather than cures. He said deep learning is also good at jobs like matching patients to clinical trials. But getting the most out of it, he argued, means reforming how society decides which drugs get made.
His first proposal is a prize system. The public picks the drugs it wants to exist, pays a prize to whoever develops them, and then produces them at generic cost. He called his overall strategy "a judo move." The "boosters' world" of transformative cures for everybody, he said, will not come from "letting it rip." It is more likely to come from taking a strong stand against being replaced, then using industrial policy (government action to steer investment toward chosen goals) and Operation Warp Speed-type approaches to build the technology that delivers the most public benefit.
The conversation also included an account of a medical emergency in which running a son's test results through three AI systems in parallel proved very valuable, along with praise for OpenAI for making its health product free and unlimited. Lovely was then asked for his positive vision three to five years out. He described a political program: a "third New Deal" with Medicare for all and possibly a locally run jobs guarantee. Alongside it would come heavy investment in institutional and state capacity to develop the treatments society wants.
The idea he said he is "playing around with" is "cures for all": the government would treat diseases as if they were all COVID and it were all Operation Warp Speed. Diseases would be prioritized by tractability and disease burden, among other factors. Then the government would use the whole-government approach, including advance market commitments (promises to buy a product before it exists), human challenge trials in which volunteers are deliberately exposed to a disease, and AI where it helps. "In a lot of cases, it's really about the institutions," he said. The cures could be made freely available around the world or provided at cost.
The model he is borrowing has a mixed record. The Government Accountability Office's review of Operation Warp Speed found that the program sped up COVID-19 vaccine development in two main ways. Companies overlapped some development steps and expanded manufacturing before the vaccines were authorized, while initial human safety and immune-response testing still came before large trials. Federal help covered facilities, equipment, supplies and specialized staff. By January 31, 2021, however, 63.7 million doses had been released: about 32 percent of the 200 million contracted from companies with emergency authorizations for delivery by March 31. The original goal of 300 million doses had not yet been met.
How much delay is acceptable
The hardest question put to Lovely was about time. It cited Amodei, who has recently told the story of his father dying of a disease that became curable just a couple of years later, and called that argument compelling. So how much delay in life-saving advances would Lovely accept to pursue them through narrow tools rather than a general labor replacer?
Lovely began by questioning the premise. We are not getting to cures as fast as we could, he said, partly because lab funding is being cut in many cases, and a Warp Speed approach across many diseases could produce a great deal of medical technology on its own. He also doubted that a superintelligence can be made safe, because safety is "a property of not the model, but the whole system." People in power would have it, he said, and would use it for purposes he strongly disagrees with.
On extinction risk, he argued that very basic economic models with conservative assumptions justify spending hundreds of billions or trillions of dollars a year to reduce existential risk even slightly. Those assumptions include not counting future generations or non-Americans. Increasing that risk in exchange for faster cures, he said, "doesn't pencil." He called the deaths of loved ones "incredibly tragic." He also noted that many people already die of preventable diseases because they are poor, and called fixing that "a deep moral obligation," to be pursued as fast as possible but without taking huge risks against everyone's will.
His final answer was democracy. Suppose citizens' assemblies around the world had to support moving forward with any AGI project. Then the developers would carry the burden of showing that it is not too dangerous and will benefit everybody, and they would have to answer the hard questions before going ahead. Right now, Lovely said, the approach is "ask for forgiveness, not permission while you're gambling with the lives of everybody on the planet."