California Bans Firing Workers on an Algorithm's Word Alone

California has become the first US state to stop employers from terminating or disciplining workers solely on the say-so of an automated system. Governor Newsom signed SB 947, and the law takes effect in July 2027.

The bill's core requirement is simple to state and harder to comply with. When an algorithm is the primary basis for a decision to fire, demote, or discipline someone, a human has to independently review it. The employee gets written notice. The employee also gets a right to see the data the system used.

That last piece is the one that will keep compliance teams busy. A right of access to the data behind a decision forces companies to know where that data came from, how it was weighted, and whether it can be produced in a form a person can understand.

The rule targets the quiet part of automation

Most employment law already assumes a human decided. SB 947 addresses the case where no one really did.

Workforce management software has been assigning shifts, scoring performance, and flagging attendance for years. The systems are marketed as decision support. In practice, a score that drops below a threshold can trigger an automatic termination, and the manager who clicks the final button may have no more information than the score itself.

The bill draws its line at that arrangement. If the algorithm is the primary basis, a person has to look at it. If a person does look, the notice and data-access duties kick in. The design assumes the human review is real rather than nominal, which is where the enforcement question will live.

This is a state patch on a patchwork

California is not the first jurisdiction to regulate automated employment decisions, but it is the first to write a termination ban into state law. New York City's Local Law 144 required bias audits for automated employment decision tools and notices to candidates, and it has drawn criticism for how many employers found ways to stay outside its reach. Colorado passed a broader AI act that classifies employment decisions as high risk. The EU AI Act treats employment and worker management systems as high risk too, which brings documentation, human oversight, and logging duties for anyone placing them on the European market.

The result is a stack of overlapping rules that do not line up neatly. A company operating in California, New York, and the EU has to satisfy three different definitions of a covered system, three different notice regimes, and three different audit expectations. For large employers with in-house counsel this is manageable, if annoying. For the staffing agencies and small logistics firms that lean hardest on automated scheduling, it is a genuine operational burden.

The wrinkle nobody has solved

Human review is easy to mandate and hard to define. A manager who spends four minutes confirming a score has technically reviewed it. A reviewer who has twenty such cases before lunch will approve most of them. The law's protection depends on the review being substantive, and substance is not something a statute can specify in advance.

There is a second wrinkle. The right to inspect the data presumes the employer can explain what the system looked at. Many commercial HR analytics products are built on models whose weights are opaque even to the vendor's own support team, and whose inputs include proxies that nobody consciously chose. Producing a meaningful answer may require rebuilding a decision the employer believed was already made.

The bill's backers are betting that the disclosure duty, more than the human review, is what changes behaviour. When a company has to hand over the reasoning, systems that cannot produce reasoning become hard to defend, and that pressure reaches back into procurement. The vendor that can explain itself wins the renewal.

What employers should do before July

The practical steps are not exotic. Inventory which systems touch hiring, evaluation, scheduling, promotion, and termination. For each, determine whether it is the primary basis for a decision or a supporting input, because the answer changes what the law requires. Keep records of human involvement at each stage, including timestamps and who signed off. Then test whether the system can produce an explanation a worker could actually follow.

The most useful exercise is a dry run. Pick one recent disciplinary action, and reconstruct it as if an employee had asked for the data. Where the trail runs out is where the gap is.

The vendor side of the equation

The bill's reach extends further than the employer's own procedures, because procurement is where the constraint actually bites. A company that has to produce an explanation on demand cannot buy a system that will not provide one, and that requirement travels up the supply chain into how HR analytics products are built and sold.

Expect three adjustments from vendors. Documentation that describes inputs and scoring logic in a form a non-specialist can read. Logging that records which version of a model produced which decision, so a company can reconstruct a call months later. And a human-in-the-loop feature that is a real workflow step rather than a checkbox that a manager clears in one click.

The vendors that already invested in explainability will frame this as a feature. The ones that did not will have to decide whether to rebuild or to sell into jurisdictions with looser rules, and the second option shrinks every year as more states move.

The timing is the point

California chose a July 2027 effective date, which is roughly a year after signing. That gap is deliberate. It gives covered employers time to inventory their systems and renegotiate contracts, and it gives the state time to write enforcement guidance that the statute does not supply.

It also creates a window in which the practical rules are being set by the first few complaints rather than by the text. How much human review counts as independent, what a data-access request has to contain, and whether a score that is one of several inputs is a primary basis will all be resolved case by case in the first year. Companies that document their reasoning now will be arguing from a stronger position than ones that improvise after the first inquiry arrives.

What this actually settles

For all the procedural detail, SB 947 answers one question clearly. When a machine makes a call that costs someone their job, a person has to own that call, and the worker gets to see why. That is a narrow principle, and it is a hard one to argue against in public.

Everything else is unfinished. The definition of a covered system, the standard for independent review, and the reach of the data-access right will be litigated and revised. The important part is that the burden of proof has shifted. Before, a worker had to show that an automated decision was unlawful. Going forward, an employer has to show that a person genuinely made the decision, and that it can explain how.

SB 947 will not stop automation in HR, and it is not meant to. It sets a floor: a person has to own the decision, and the worker has a right to see the basis. How much that floor holds will depend on whether the review stays real once the paperwork is routine. Companies that treat the disclosure duty as a design constraint rather than a form to file will find the compliance easier, and the systems they buy more defensible.