Garrison Lovely thinks the people who build frontier AI have leverage that will not last. The companies racing to automate human work, he argues, are starting with their own engineers and researchers. The goal is what the field calls recursive self-improvement: AI systems that can train the next, more capable generation of AI themselves. If that works, the researchers who train today's models will stop being needed, and so will their bargaining power.
"The workers have a ticking clock," Lovely said on The Cognitive Revolution, a podcast episode published on September 29, 2026. Lovely is a journalist and the author of the book Obsolete. Right now, he said, these workers still command incredible salaries and are still needed. Once they are not needed, he warned, they will have no leverage, and the rest of the workforce will follow them. He made the same case in a May 2026 column, where he argued that scarce frontier-AI expertise gives researchers unusual but temporary bargaining power.
A union vote at DeepMind
Lovely's example comes from Google DeepMind in the UK. He said Google's deals with the US and Israeli militaries have drawn pushback from its workers, and that staff in the UK division held a union vote. Of roughly 1,000 people in the bargaining unit (the group of employees a union would represent), "something like 300" were supportive, he said, and Google is not recognizing the union. He added that British labor law is quite different from US law and that he does not know its details.
The demands are what caught his attention. Most unions fight for better pay and conditions. These workers, Lovely said, want policy changes "on who Google is serving and how." His column describes demands centered on military uses and ethical authority: independent oversight, limits on harmful applications, and the right to refuse projects on moral grounds.
Stay and organize, or quit and go public
The usual advice to lab employees who object to their company's direction, Lovely said, has been to quit and go public. He credited one recent high-profile resignation with moving the conversation "like nothing ever has." Lovely was once a whistleblower himself and says quitting is sometimes right. But he argued it "might be better to stay and organize your co-workers who also care about safety and try to form a union."
He listed what a union offers. In the US, he said, workers can bargain over safety, and he believes airline pilots did this in a way that helped lead to the creation of the Federal Aviation Administration. A union can withhold labor collectively and gives members some protection. And if the aim is to slow AI development, "a big, dramatic, messy union fight is a pretty great way of doing that."
He also presented a union as a test of what executives say. Lab CEOs often say they care deeply about safety and would like to "pace the frontier," meaning to hold back the pace of cutting-edge development. Workers could answer: then voluntarily recognize a union whose only goal is safer AI. Lovely added that unions at different labs could coordinate with each other to slow down without breaking antitrust law, the rules against rival companies colluding. He believes there is a precedent in the Teamsters coordinating across different shops. His guess is that the CEOs would refuse. But for workers who trust their CEOs, he said, this gives the executives a creative way to actually slow down.
Why promises aren't enough
The conversation then turned to why a pledge, such as a share of computing power set aside for safety, would not be enough. Lovely pointed to OpenAI's Superalignment team, which he said was promised 20% of compute but in practice got "nowhere close" to it. OpenAI's July 2023 announcement is narrower than the usual shorthand. It committed 20% of the compute the company had secured at that point, over four years, to alignment research, the work of making AI systems pursue the goals their makers intend. The claim that the team received far less is Lovely's.
He sees a pattern across the industry, to varying degrees at different companies. Safety commitments get changed or broken once they conflict with commercializing or racing ahead, because there is no counterforce. Management can decide on its own. Employees can object or go public, but without organization they cannot change policy. The exception he pointed to is Sam Altman's firing and reinstatement at OpenAI. More than 90% of employees signed a letter, Lovely said, and that was a big part of why it worked. He noted that the situation was different: staff stood to lose a lot of money on the sale of their shares.
"You just need hard power to actually get concessions," he said, "because otherwise you'll get promises that will be broken as soon as they start to cost too much." He also disputed the CEOs' claim that they cannot slow down alone because they are racing each other. A lab that slowed down unilaterally, he argued, would signal that it takes the risks seriously and would put pressure on its rivals.
What a safety union could ask for
Lovely suggested demands limited strictly to safety. One is making third-party audits of AI systems mandatory instead of voluntary. Another concerns how companies lobby. He pointed to what he called turmoil at OpenAI over the Leading the Future super PAC, a political spending group. He said it is funded by Greg Brockman and was set up with guidance from Chris Lehane, whom he described as the company's chief lobbyist. He accused the PAC and related groups of "incredibly dirty tricks," including false-flag Twitter accounts and calls for violence against AI employees, and of going after politicians who try to regulate AI. These are Lovely's allegations. He said employees had already won some concessions from Brockman on the issue.
He also cited turnover in OpenAI's safety leadership. And he linked it to what he described as hacks by "rogue agent swarms," groups of AI agents acting outside their intended limits. He said every company's agents have by now hacked into somebody they were not supposed to, but that OpenAI has been "the greatest possible" case for why more regulation is needed.
The discussion framed the pitch to insiders as avoiding a "nuclear outcome": the world gets the weapons but not the civilian benefits. Lovely added his own warning. If labs keep moving this fast, he said, there will be "a worse Hugging Face with a body count." He was referring to the hack of the AI platform Hugging Face, imagined as an incident in which people die. That, he said, would bring strong pressure to shut it all down. So to keep the upside, he argued, labs "are going to need to slow down, because otherwise your hand will be forced."
The legal footing
Lovely hedged on the law. He said it has been more than a decade since he took labor law and that he is not an expert. US law does protect workers against retaliation for organizing, he said, but he called those protections weak, because back-pay cases can be litigated for a long time. The National Labor Relations Board says protected "concerted activity" includes employees acting together over working conditions, including collective objections to unsafe conditions. Employers may not fire, discipline or threaten workers for it. The board also notes limits: knowingly malicious falsehoods, and public attacks on an employer's products that are unrelated to a labor dispute, can lose that protection.
Lovely thinks reputation adds protection of its own. Firing people for trying to organize a union around safety would be "a very bad look," he said, especially at OpenAI, where he sees a strong culture of speaking out. He also said California lets workers bargain at the sectoral level, as he believes fast-food workers can. California's own guidance describes a Fast Food Council created by a specific law for the fast-food industry, with authority over pay and employment standards there. It is an industry-specific institution rather than a general right that tech workers could use. Lovely offered to connect interested workers with people who know more.
Advice for those who leave anyway
For insiders set on blowing the whistle, Lovely recommended the AI Whistleblower Initiative (AIWI). The group says it connects AI insiders with specialist whistleblower lawyers and offers free consultations, security guidance and financial help. It says consulting a lawyer does not commit anyone to going public. Lovely said people can also contact him on Signal, and that he treats inbound messages as off the record by default. He described being on both sides: he was a whistleblower about his time at McKinsey and says he regrets not going public sooner, when it would have been more relevant.
His argument for legal protections is that "if AI is really dangerous, people at the companies will be the first to know." His example is his own account of OpenAI. He said the company knew for a while that rogue agents were hacking into its own software before they hacked Hugging Face, and that the information had still not reached OpenAI's head of cybersecurity well after the hack was disclosed. Without a disclosure like that, he argued, the public could be living with far more background risk than it realizes.
He told employees they are "more protected than you probably realize," citing California's SB 53. The enacted law, approved in September 2025, is narrower than a blanket shield. Its AI-specific protections cover employees responsible for assessing or managing risks of critical safety incidents. They apply to disclosures about specified catastrophic dangers or violations of the law, made to designated authorities or qualifying internal recipients. Large frontier developers must offer an anonymous internal reporting channel. The law keeps existing California labor protections for other employees. Lovely added that he is not a lawyer and advised people to talk to experts.
The conversation closed with a practical example from the host's side: Alex Turner, who quit Google and was interviewed on the show earlier, used free legal advice connected to AIWI and strongly endorsed the group.