Axios reported this week that union contracts are now among the strongest defenses American workers have against artificial intelligence in the workplace. That is true. It is also an indictment, and not of the unions.
Roughly 16.5 million Americans belonged to a union last year. About 11.2 percent of the workforce is covered by a contract. That leaves something like 130 million people with no negotiated say in whether a system they have never seen rewrites their job description or eliminates it.
The numbers that should end the argument
Here is the part that stopped me. According to the Bureau of Labor Statistics, 4.4 percent of workers in computer and mathematical occupations are union-represented. In finance, it is 1.1 percent.
The occupations most exposed to white-collar automation are the least organized in the country. The protection is real, it works, and it has almost perfectly missed the people who need it.
Where coverage exists, the wins are concrete. The NewsGuild has between eighty-five and ninety contracts with explicit AI provisions. Unionized Politico journalists won an arbitration in November 2025 after management deployed tools that generated subscriber policy reports with no editorial review and, predictably, factual errors in them. SAG-AFTRA’s new television agreement, approved in June, tightens limits on synthetic performers. Workers at ZeniMax won language requiring that AI support rather than replace them, and requiring management to notify and bargain before deploying anything new.
Notice what those have in common. None of them ban the technology. They require a conversation before deployment instead of an announcement after it.
The honest objection
The strongest case against all of this is that bargaining over technology is how you lose the plot. Firms that must negotiate every deployment move slower than firms that do not, and slower firms lose to faster ones, and everyone at the slow firm eventually loses the job the contract was written to protect. That argument is not stupid. It is most of the history of American manufacturing.
But it proves less than it claims. The ZeniMax language does not stop deployment. It stops surprise. A firm that cannot explain to its own workforce why a system is going in probably cannot explain it to a regulator or a customer either.
Where this actually lands
A former FTC commissioner made the point that the rules unions are negotiating could map onto federal policy without much translation. Notice and consultation before deployment. Disclosure when a system touches hiring or termination. Human review of consequential output.
None of that is exotic. It is what any competent engineering organization already does before shipping something that can hurt people.
Congress has left this to private contract negotiation between parties of wildly unequal power, and then only for one worker in nine. That is not a policy. It is an absence with a lobbying budget.
Disclosure: NeuraWeb Global Inc., of which the author is founder and CEO, builds software and competes in the technology market this column discusses.
Add your perspective
Left: the answer is labor law reform that makes organizing possible in tech and finance, not admiration for the few contracts that squeaked through. Right: mandating consultation before deployment hands a veto to the least productive parties in the firm and guarantees the work moves offshore or to a startup with no union at all. Center if the useful floor is disclosure and human review of consequential decisions, obtained by statute rather than by contract, so it covers all 130 million.
