The guardrails are arriving through bargaining, not legislation

A trade press story worth reading twice

On September 2, 2026, NW Labor Press, a Pacific Northwest labor outlet, published a piece called "AI, Meet CBA." The number in it is the headline: more than 175 documented union contracts now contain guardrails on how AI can be used against the people covered by them. That is not a category that existed five years ago as something worth counting. It exists now because bargaining committees started writing AI clauses into contracts one local at a time, and someone finally added them up.

Look at the scale for a second. A Kaiser Permanente coalition covering 60,000 members has a provision addressing AI. That is not a pilot agreement at one clinic. More than 50 NewsGuild contracts now restrict how newsrooms can use AI, which means the same fight is playing out in an industry with a completely different labor structure and a completely different set of tools. Contract language does not spread across sectors like that unless it is solving a problem people actually have.

What triggered the language

The article is specific about the flashpoints. Workers at ProPublica staged a walkout involving more than 150 people. At the Skagit Valley Herald, workers raised a demand for a $20 an hour wage-floor tied explicitly to AI. An Oregon AFSCME contract includes a 90-day advance-notice requirement before any new AI system can be deployed. And the piece names Oregon Health and Science University (OHSU) specifically in its coverage of algorithmic productivity surveillance: keystroke counts, screenshot capture, the tools that tell an employer what an employee did at their desk minute by minute.

That last category, keystroke and screenshot monitoring, is most of what the article is actually about. On its face, this is not a story about cameras.

Why the distinction will not hold

Here is the argument this piece exists to make: that distinction, camera versus keystroke logger, will not survive contact with a bargaining committee, and it should not.

A camera system that flags unsafe behavior on a warehouse floor and a keystroke counter that logs how long someone paused between emails are built for different reasons. One is meant to prevent an injury. The other is meant to measure output. The intent is genuinely different. But mechanically, from the point of view of the person being watched, they do the same thing: both observe a person while they work and produce a record of it that someone else can review later, for a purpose the worker may never have agreed to and cannot always see.

A union that has just spent a bargaining cycle writing 90-day notice language for software that counts keystrokes is not going to treat a new camera deployment as a different category of problem. It is going to ask the same three questions it asked about the keystroke tool. That is the correct outcome, not a labor relations inconvenience. DHI sells software that watches people at work. Pretending that is categorically unlike other workplace monitoring because the sensor is a lens instead of a keyboard driver would be dishonest, and it would also be a poor prediction about what happens at the next contract negotiation one of our customers sits through.

The 90-day clause is not a formality

The Oregon AFSCME contract's 90-day advance-notice requirement deserves its own paragraph, because it changes something concrete about how deployments happen. If a customer operates under a contract like that, the deployment date is not something a vendor gets to set by finishing the install. It is set by the calendar in the labor agreement. A sales timeline that assumes go-live the week hardware ships does not survive contact with that clause, and neither does an implementation plan that treats notifying the workforce as a step that happens after the cameras are already live.

That is worth planning around before it becomes a problem on a specific job site, not after.

What a one-page answer has to say

The practical requirement coming out of all of this is not, in most of these contracts, a legal one yet. It is a documentation one. A bargaining committee, or a worker who just heard their workplace is getting new cameras, is going to ask three questions: what does this collect, who can see it, and how long is it kept. If a vendor cannot answer all three on a single page, in plain language, that vendor loses the deal. Not because the detection model is worse than a competitor's. Because the answer took a phone call and a follow-up email instead of one page anyone could read before the meeting.

We think that is a product requirement, not a compliance chore to hand off to whoever writes the terms of service. The page should exist before the first contract that asks for it, not after.

What we can say honestly

Here is the one claim in this piece we can make about our own product with a straight face: DHI's design keeps video on site, and what leaves the building is an event, not a frame. A detection produces a record: what happened, where, when. It does not, by default, produce a video file traveling to a server somewhere else. That is a real architectural choice, and it is relevant to exactly the question a bargaining committee is going to ask about retention and access. It is not a complete answer to every question this article raises, and we are not claiming it is. It is one fact, stated once, that happens to be true.

The takeaway

Statutes move on legislative calendars. Contracts move at the speed of the next negotiation, which is often much faster, and 175 documented examples says this is no longer a fringe activity. The mechanism most likely to constrain workplace camera deployments over the next two years is at least as likely to be a clause a union local wrote into a contract as a bill signed by a governor. Vendors who treat that as someone else's problem, because the current wave of contract language is aimed at keystroke loggers and not cameras, are reading the wrong lesson from the data.