Workers Are Already Setting the Terms for AI at Work

Something is happening in workplaces right now that has little to do with technology and everything to do with who gets a say in it.

In late July 2026, Axios reported that with Congress largely inactive on workplace AI, union contracts have become one of the strongest tools workers have against AI-driven disruption. The NewsGuild-CWA now carries AI-related language in something close to 90 collective bargaining agreements. Its president, Jon Schleuss, described the logic plainly: without a contract, a worker has no say in how AI gets introduced to their job. A few months earlier, unionized journalists at Politico won an arbitration ruling forcing the removal of two AI tools their employer had deployed without bargaining over them. SAG-AFTRA's newest television agreement extended protections against unauthorized AI replicas of performers. The International Longshoremen's Association negotiated an outright ban on fully automated terminal technology at the ports it covers.

None of this is hypothetical, and none of it is limited to a handful of high-profile sectors. It is also not new. What is new is the pace, and the fact that AI is now a named subject at the bargaining table rather than an afterthought.

It is also no longer limited to industries with a long history of worker organizing. The tech sector itself, previously among the most union-resistant sectors of the American economy, is shifting. The Guardian reported on July 21 that tech workers are unionizing specifically over the issue of AI, after months of AI-tied layoffs and new monitoring tools left some employees feeling power had shifted too far in favor of their employer. Workers at the University of California system's IT operations recently voted to form the largest tech-worker union in the country, citing AI-driven layoff anxiety among their reasons. A union representing employees at ZeniMax, the Microsoft-owned gaming studio, won contract language requiring that AI support workers rather than replace them, with mandatory notice and bargaining before any new deployment. One Meta employee involved in a UK union drive put the motivation simply: organizing was the way to actually have a seat at the table on how the company operates, not just on pay.

The same pattern is visible abroad, often further along, and not only through formal regulatory systems. In July of this year, in Italy, Amazon signed the first national collective agreement it has reached with any country's workforce, covering 57 sites. The unions say the deal sets specific conditions on how new monitoring technologies can be deployed, including a rule that video surveillance footage cannot be used for disciplinary purposes. That last piece is worth pausing on, because it is a direct, negotiated limit on how a workplace monitoring tool can be used against a worker, arrived at through bargaining rather than through law.

Germany's system goes further still, and is built into statute rather than negotiated site by site. Works councils already hold binding co-determination rights over any technical system capable of monitoring employee behavior or performance, and starting August 2, 2026, the EU AI Act requires them to conduct formal risk assessments on high-risk workplace AI systems. Spain's 2023 national agreement between unions and employer federations requires companies to disclose how AI is used in hiring, evaluation, and dismissal decisions, and to keep a human in the loop. The UK's Trades Union Congress has built a five-point strategy arguing that AI adoption across British workplaces has stayed shallow specifically because workers have been excluded from decisions about how it gets introduced.

There is a further point that runs underneath every one of these examples. A seat at the table only matters if the people in it know what to look for. That means training has to go past how to operate a new tool or how to re-design workflows. Workers exposed to an AI system, its outputs, or the decisions it drives are often the first to see where it creates a new hazard, whether physical, psychological, or procedural. They are also usually the ones best placed to judge which fixes will actually hold up on the ground. Training that stops at usage and workflow, and never gets to hazard identification and mitigation defined by the workers who are exposed, hands the questions that matter most to people much further from the work, and much further from its consequences.

This is not only about protecting workers from AI. Bringing workers in also tends to make the technology work better. Management often reads employee pushback as delay or added cost, when it is closer to the opposite.

A 2026 report from UNI Europa looked at how AI was rolled out in banking, telecommunications, and broadcasting. It found that collective bargaining did not slow adoption down. In its case studies, the opposite held. At Intesa Sanpaolo, Orange, and RTL Group, the employers who brought workers and their representatives in early, and did it in good faith, saw the technology take hold faster and stick better than the ones who did not. The OECD's 2023 Employment Outlook points the same direction. Across a range of countries, workplaces where workers had some form of representation tended to see better working conditions after AI arrived than workplaces where they had none. This is descriptive evidence rather than proof of cause, but it shows up across enough countries and sectors to be worth taking seriously.

Research summarized by the Washington Center for Equitable Growth found that AI models improve when the people who actually use them are involved in developing, checking, and correcting them. It also noted that management research separately found workplaces with higher trust and worker autonomy see better performance outcomes from AI investment specifically. Workers are more than just the people AI is done to, they are frequently the people who know first when it's wrong.

None of this means every workplace with a union has this figured out, or that every non-union workplace is out of options. Most American workers are not in a union and are not likely to be soon. According to the US Bureau of Labor Statistics, about 11 percent of workers were represented by a union in 2025, which leaves roughly 130 million without that specific form of leverage. But the absence of a union doesn't mean those workers are without any negotiation mechanism. The National Labor Relations Act protects workers who raise concerns about workplace conditions collectively, whether or not they belong to a union. A growing patchwork of state laws is filling gaps federal policy hasn't reached, such as Illinois now requiring disclosure when AI is used in hiring and New York City requiring public bias audits of automated hiring tools. And there are early signs that AI is becoming an organizing issue in workplaces that have never had a union at all, as workers push back on scheduling algorithms and monitoring tools they had no say in adopting.

The throughline across all of it, unionized or not, formal or informal, American or not, is the same: workers who have a real channel to name what's happening to their work, ask how a system was built, identify where it introduces hazard, and push back when it gets something wrong are not obstacles to responsible AI adoption. They are close to a precondition for it.

This is the first in a series looking at where that channel exists, where it doesn't, and what fills the gap when it's missing.