AI Company Employees Should Unionize

Recent discussions have been inconclusive, yet they did connect me with activists on both sides of the issue. Many who plan to stay and fight are exploring the idea of forming labor unions. A group named the Coalition Of Concerned AI Staff already exists, but it functions across multiple companies and includes former staff, making it less ideal for traditional union-building. AI workers, however, may find ways around this obstacle.

Unionization battles in other industries, like Starbucks, have often been intense. AI workers, however, face a different landscape. Employers can replace unionized workers with a seemingly endless supply of faceless replacements, making legal protections less necessary. Given the scarcity of skilled AI researchers, the threat of a dozen top researchers walking out holds significant weight. This reality undermines the intimidation tactics often seen in other fields.

Informal "concerted activity" could suffice for AI researchers to form a union. Simply displaying a simple sign, like "OpenAI Safety Union," and having at least two employees support it would suffice. The term "union" is optional; employees could opt for a "research council" or any other moniker.

Once a union is established, the law would protect employees from being fired for joining or discussing their union. It would also prevent termination for striking over workplace conditions. While the "workplace conditions" requirement may seem to rule out most safety activism, many safety concerns overlap with work-related issues, such as whistleblower protections, auditing requirements, and staffing of safety teams.

Moreover, the fear of recursive self-improvement, which could lead to automation of AI researcher jobs, is inherently linked to workplace concerns. Even if a safety issue has no direct workplace relevance, unions still retain the right to strike against it, though they would lose legal protection against termination. High-caliber AI researchers are nearly irreplaceable, commanding tens of millions of dollars apiece, so AI companies are disinclined to fire multiple researchers simultaneously.

Legal protections against firing striking researchers would likely be minimal, and striking AI researchers would draw significant public attention and face a PR nightmare.

If employees believe their company already prioritizes safety, they should still consider unionizing. This move could provide additional options for negotiations with companies, which are subject to less antitrust scrutiny. Unions could negotiate terms to slow research or prioritize safety even when executives are constrained. For instance, if OpenAI's leadership orders researchers to switch to a less-monitored transformer design, a union might question this decision or mutually refuse to proceed.

If Sam Altman insisted that they must create a looped transformer or risk losing to Anthropic, the organizers could reach out to the Anthropic Employee Union. If Anthropic's union agreed to strike over the same issue, the two unions could join forces, creating a united front that could pressure their respective companies. Although this arrangement offers limited standalone value, it could set a precedent within the industry or serve as a basis for government intervention.

In the worst-case scenario, two unions could negotiate an initial pause deal, enforce it on their parent companies, and the government might take notice, potentially incorporating it into a future international agreement.