California didn't ban employers from using AI to fire or discipline people. It banned them from hiding behind it, which is the right distinction and a more useful one than it sounds. SB 947 doesn't ask the obvious but unanswerable question — is the algorithm fair? — it asks the answerable one: did a human actually look at the file before someone lost their job? That's a sensible design choice. Bias in automated decision systems is a documented, real-world risk, and requiring a human to cross-check performance records before an adverse action takes effect is a low-cost, high-plausibility safeguard against the worst failure mode: a model making an irreversible call with nobody accountable for it.
But the law's central phrase — 'primarily relies' on automation — is doing enormous work without much definition behind it. The Chamber of Progress's complaint isn't reflexive business griping; it's a legitimate governance problem. If employers, regulators, and courts can't agree on what counts as primary reliance, the practical effect is unpredictable: some companies will over-comply out of litigation fear and slow adoption of genuinely useful workplace tools, others will quietly structure their systems to stay just outside the definition, and the ambiguity itself becomes the compliance cost. Vague thresholds don't eliminate risk, they just relocate it into legal discovery.
There's also a familiar pattern worth watching: mandated human review can become theater rather than substance. A manager glancing at a file for thirty seconds to satisfy a legal requirement is not meaningfully different from the automated decision it's supposed to check, unless the law or its enforcement regime demands something more rigorous than a signature. The bill doesn't spell out what 'confirming the outcome' actually requires in practice, which means its real-world teeth depend entirely on how aggressively state labor regulators choose to interpret and enforce it — and California's labor enforcement capacity is not infinite.
The fact that a business-aligned group calls the bill underdefined and a conservative gubernatorial candidate calls it too narrow tells you this isn't really a left-right fight. It's a fight about regulatory precision, and precision is exactly what's missing. The instinct behind the law is correct. The execution leaves the hardest questions — what counts as primary reliance, what counts as adequate human review, who checks any of it — to be fought out later, probably in court, at employers' and workers' expense.
How it may affect me
In the near term, California workers facing AI-influenced discipline or termination gain two concrete things: written notice of what data was used against them, and a human contact who can explain the decision. That's a real, immediate improvement in transparency and a potential foothold for disputing an unfair call — something that previously may not have existed at all for workers managed by automated systems.
For employers, the immediate effect is compliance friction: documenting human review, training managers to actually examine personnel files rather than rubber-stamp an algorithm's output, and absorbing legal uncertainty over what 'primarily relies on automation' actually means. Smaller employers with thinner HR infrastructure will likely feel this more than large firms with compliance teams already built for California's regulatory environment.
Longer term, two outcomes are plausible and depend heavily on enforcement and future clarification. If regulators or courts define 'primary reliance' narrowly and police human review seriously, this could become a genuine check on algorithmic bias in hiring and firing decisions. If the definition stays vague, the practical effect may be thinner: some companies restructure their AI use just enough to avoid the law's trigger, some workers get a formal notice but no substantive change in outcome, and the law's real value shrinks to the paperwork it generates rather than the protections it was meant to deliver.
