Let’s talk about the week California made it illegal for a machine to be the only reason you get fired.
On the last day of September, Governor Gavin Newsom signed SB 947 — the No Robo Bosses Act of 2026. First law of its kind in the United States. A California employer can no longer rely solely on an automated decision system to fire or discipline somebody. If the algorithm is the primary basis for the call, a human has to come in and corroborate it. With actual evidence. A manager’s assessment, peer reviews, the personnel file. Not a click. A look.
And the part most people skimmed past: the worker has to be told, in writing, that AI played a primary role. What personal data the system used. And the name of a real person they can talk to about it.
Read that list again. That isn’t a rule about firing. That’s a rule about proof.
Picture how this used to go. A warehouse floor, a call centre, a delivery route. Somebody gets pulled aside and told their numbers weren’t good enough. And the numbers were the whole conversation — because the numbers were a person once. A supervisor who’d watched you, who knew the kid was sick, who had to look at you while he did it.
Now picture the same moment the other way. The decision drops out of a system that’s been grading you for months — idle minutes, pick rate, keystrokes, timing — and nobody in the room can explain the score. You ask why. The answer is the score. The score is the answer. That’s the thing this law is trying to put a stop to.
Here’s what people are missing, though. The bill doesn’t ban AI from firing decisions. It bans AI alone. And a law that says a human has to sign off is only ever as strong as the human.
If the reviewer gets a live score at nine in the morning with forty other files to clear, no way to see how the score was built, and no upside to disagreeing — that is not oversight. It’s a signature with a pulse. It launders the machine’s decision through a person’s name and hands the company a defence. The same thing I keep saying about doctors and clinical AI: the human in the loop is worthless without the time, the information, and the standing to say no.
And watch the word the whole thing turns on. Primarily relies. Every requirement fires only if the AI was the primary basis for the decision — and that word is undefined. Business groups fought the bill for a year on exactly that. Which means the most important question in California employment law right now is this: who decides whether the machine was primary?
Because if the score said terminate and the manager agreed in four minutes flat, was the human the decision-maker, or the machine’s translator? If the manager writes a reason down after the fact, does that count?
You can guess how that argument goes. The company has the lawyers.
So be honest about what this law is. It’s a floor, not a ceiling, and it got narrower to survive. Newsom vetoed an earlier version last year over its notice rules. It came back with the advance-notice cut and gig workers removed entirely. Business groups still fought it to the end.
What’s left still matters — because of the one thing the lobbying could not strip out. The disclosure. You have to be told the machine was the reason, and which of your data it used. That’s the real teeth. You cannot appeal a decision you cannot see.
Same day, same pen, Newsom signed a whole package around this. Companies now have to say in writing when a mass layoff or termination is caused by an AI system. Surveillance tools are banned in workplace bathrooms. And in an executive order that reads like a middle finger aimed at the White House, California formally declared that Artificial Intelligence shall be called “Artificial Intelligence” — a quiet refusal of the rebrand to “superintelligence.” The state that hosts the industry just spent a day drawing lines around it, in public.
Now the operator lesson, and it’s not aimed at workers. It’s aimed at anybody who runs a company.
If your performance stack, your screening tool, your productivity dashboard is the thing that decides who gets managed out, you just inherited a legal duty in California and a moral one everywhere else. And the lazy version of compliance is to print the score, hand it to a manager, and collect a signature. That version will not survive a court, and it shouldn’t. It won’t survive the first employee who asks the obvious question: what did the human actually add?
So build it properly. Keep the reviewer’s reasoning, not just their yes. Make sure the person signing off can see how the score was assembled and can overrule it on the record. And ask the harder question before a regulator does — if I switched the model off tomorrow, would my managers still know how to judge a person? If the answer is no, the model stopped being a tool a while ago. You just didn’t notice who the boss was.
The machines are going to do more of the watching. Fine. Let them grade the work. But ending someone’s livelihood is an act with a consequence, and consequence is the one thing a model cannot carry. It has no family to feed, no bad night behind it, no reputation to lose when it gets someone wrong. Only a person can answer for a decision. That isn’t a technical limit. That’s the entire reason you put a person there in the first place.
A machine can propose. A person has to decide. And the person has to be able to be wrong — and be held to it.
That’s what California wrote down this week. Not that AI can’t help. That somebody has to sign for the firing, and that the signature has to mean something.
Which is really just the old rule, written into law at last. If a decision is about a human being, a human being has to own it.
Let’s keep it that way. You’ve got this.