California's AI Firing Law: The Human-Review Record That Meets the Burden of Proof

From 1 July 2027 California employers can't discipline or fire on an AI's output alone — and once a worker shows AI was used, the employer must prove it complied. Here's the record that does it.

What SB 947 Actually Says — and What It Doesn't

California Labor Code §§ 1520–1526.7, added by SB 947 (Stats. 2026, ch. 859) and available in full on the Legislature's site, becomes operative on 1 July 2027 — that date is still ahead. Until then, it has no legal force, so the window to build compliant processes is open but finite. The law covers every employer of California employees. There is no headcount floor. A 12-person startup whose account manager works remotely from Sacramento is covered on the same terms as a 5,000-person enterprise.

The statute defines an "automated decision system" (ADS) broadly: any computational process that uses machine learning, statistics, or other data-processing techniques to generate a score, classification, or recommendation that influences a consequential employment decision. That language covers performance-management platforms, attendance-tracking tools, productivity-monitoring software, and AI-assisted hiring systems if they feed into a discipline or termination decision.

Two threshold concepts govern the whole statute. "Solely" relying on an ADS is prohibited outright under § 1522(b). "Primarily" relying on one triggers the corroboration requirement. The line between those two concepts is where the litigation risk lives, because an employer who thinks a manager's five-minute rubber-stamp clears the "solely" bar may find a plaintiff arguing it still meets the "primarily" standard.

The Corroboration Requirement: What § 1522(b) and (c) Demand

Under § 1522(b), when an employer relies primarily on an ADS output to reach a discipline or termination decision, a human reviewer must independently corroborate that output. Corroboration is not approval — it means reviewing the underlying data or other evidence and forming an independent assessment. The statute names the evidence categories explicitly: supervisory evaluations, personnel files, work product, peer reviews, and witness interviews.

§ 1522(c) is the kill switch. If the human reviewer cannot corroborate the output — or finds it inaccurate, incomplete, or misleading — the output may not be used in the decision. This means the reviewer's assessment must be genuine. A reviewer who flags concerns but is overruled by the ADS output creates exactly the record a plaintiff's attorney wants to see.

The corroboration requirement only works if the reviewer could have reached a different conclusion. If the process is designed so the human always confirms the machine, you don't have corroboration — you have laundering.

Build the review process so the reviewer documents what they would have decided based solely on the non-ADS evidence, before recording whether that matches the system output. That sequencing is the difference between a defensible record and a paper trail that proves the human was decorative.

Notice and Data-Access Rights Under §§ 1522(d) and 1524

SB 947 adds two employee-facing rights that HR teams must operationalize before 1 July 2027. First, § 1524 requires a written, plain-language, stand-alone notice to the employee at the moment they are told of the decision. "Stand-alone" matters: burying the disclosure in a termination letter or an employee handbook update does not satisfy the requirement. Under § 1524(b) the notice must state that the employer primarily relied on an automated decision system for the decision, that a human reviewed the decision and corroborated the system's output, who the employee can contact for more information and about their right to a description of their own data, and that retaliation for exercising these rights is prohibited.

Second, § 1522(d) gives the employee the right to a description of their own data the system used. That means HR and Legal need to know — before a decision is made — exactly which data fields fed the ADS output for that individual. If the vendor can't tell you what data went in for a specific employee, you cannot satisfy this disclosure right, and you should be asking hard questions about whether to use that vendor for consequential employment decisions at all.

Track the notice delivery the same way you'd track receipt of a WARN Act notice: document who delivered it, in what format, and when. For remote employees, email delivery with a read-receipt or portal acknowledgment is the safest approach. You'll want that documentation if a worker later claims they weren't told.

The Burden-Shift: Why § 1526.1(c) Changes Everything

This is the provision that distinguishes SB 947 from other US AI employment laws. Under § 1526.1(c), once an employee shows that an ADS was used in the decision — not that the employer violated the law, just that a system was used — the burden shifts to the employer to prove either that it did not primarily rely on the output or that it complied with §§ 1522 and 1524.

In practice, that means the review record is not optional documentation. It is the employer's defence. If the record doesn't exist, is incomplete, or was clearly created after the fact, the employer loses the burden-shift and faces liability. § 1526.1(e) sets the civil penalty at $500 per violation, and § 1526.1(d) makes injunctive relief, punitive damages, and attorney's fees available in a civil action. For a series of terminations processed through the same ADS workflow without adequate records, the exposure stacks quickly.

Keep your AI regulation compliance deadlines tracker bookmarked — § 1526.7 sets the operative date as 1 July 2027. The Labor Commissioner enforces the law (§ 1526.1(a)); SB 947 does not direct any agency to write implementing rules, so the statute itself is what your review record has to satisfy.

How California Compares to Connecticut and GDPR Enforcement

California's § 1522 is not the only rule your employment decisions may need to satisfy. The table below compares three overlapping frameworks as of today, 6 October 2026.

Dimension	California SB 947 § 1522 Operative 1 July 2027	Connecticut PA 26-15 No-defence rule since 1 October 2026; notice duties from 1 October 2027	GDPR Article 22 Uber enforcement, 21 Aug 2026
Core prohibition	No sole reliance on an ADS for discipline or termination (§ 1522(b)(1)); primary reliance requires human corroboration (§ 1522(b)(2))	Use of an AEDT is no defence to a § 46a-60 discriminatory employment practice claim	No solely automated decisions producing legal or similarly significant effects without human oversight or explicit consent
Who bears the burden	Employer, once ADS use is shown (§ 1526.1(c))	Employer — AEDT use cannot shift discrimination liability to the tool	Controller — must demonstrate lawful basis and human review mechanism
Human review standard	Must corroborate using underlying data or named evidence categories; reviewer can veto the output (§ 1522(c))	Not separately defined — discrimination analysis applies regardless of whether a human approved the AEDT output	Meaningful human assessment required; Dutch DPA found Uber's automated driver deactivation lacked it — EUR 824,990,000 fine
Employee notice	Written, plain-language, stand-alone notice at decision point (§ 1524); right to data description (§ 1522(d))	For AEDT deployed in Connecticut on or after 1 October 2027: written notice before a decision in which the AEDT is a substantial factor (purpose, trade name, data categories and sources). No notice duty before then	Right to human review, right to contest, right to explanation (Art. 22(3))
Penalty	$500/violation (§ 1526.1(e)); injunctive relief, punitive damages, fees (§ 1526.1(d))	Existing CFEPA remedies: back pay, compensatory and punitive damages, attorney's fees	Up to 4% of global annual turnover or EUR 20M, whichever is higher
Headcount threshold	None	None (§ 46a-60 applies to all employers)	None

The Uber enforcement is the clearest signal of where regulators draw the line. The Dutch DPA found that Uber's system deactivated drivers on fraud suspicions and low customer ratings without any human assessment of individual cases. Between 2018 and 2022, accounts were deactivated — temporarily, or permanently for persistent low ratings — with no human assessment at all, and drivers were not adequately told that automated decisions were being made. Uber has appealed. California goes further than the GDPR baseline: a human in the loop is not enough under § 1522(b), because the reviewer must actually corroborate the output, and § 1522(c) bars an output that cannot be corroborated. If you want more context on the regulatory landscape across jurisdictions, see our full AI regulation compliance deadlines tracker.

The Human-Review Record: What It Must Contain

Your review record is your § 1526.1(c) defence. Build it as a structured document, created contemporaneously with the decision, retained with the personnel file. Here is what it needs to cover:

• Identification of the ADS and its output. Name the system, the version or model date, and the specific score, flag, or recommendation it produced for this employee on this date. If your performance platform uses a third-party AI scoring engine, name both.
• Who conducted the human review. Full name, title, and relationship to the employee (direct manager, skip-level, HR business partner). Document that this person had access to the underlying data — not just the ADS summary output.
• What evidence the reviewer examined. List each item specifically: the three performance reviews from the past 18 months, the Q2 work-product samples, the incident report from 14 August, the peer survey results, the attendance records. Vague references to "personnel file reviewed" are not enough.
• The reviewer's independent assessment. This is the critical field. Before recording whether the ADS output was corroborated, the reviewer must record what conclusion they reached based solely on the non-ADS evidence. Write it out: "Based on the evidence above, my independent assessment is that [conclusion]."
• Corroboration finding. Does the independent assessment corroborate the ADS output? If yes, the decision may proceed. If the reviewer found the output inaccurate, incomplete, or misleading, document that finding and record that the ADS output was not used — per § 1522(c), that closes the loop.
• What the reviewer would have decided if the ADS had reached the opposite conclusion. One sentence. This documents that the review was genuine, not a formality.
• Employee notice compliance. Attach or reference the written § 1524 notice delivered to the employee: the date, the delivery method, the recipient's acknowledgment (or the method used when acknowledgment was unavailable), and a copy of the notice text itself.
• Data-description response. If the employee invoked § 1522(d) and requested a description of their data, attach the response and the date it was delivered.

Store this record with the same retention schedule as your termination documentation — California's statute of limitations for employment claims runs up to three years for statutory violations, and longer under some discrimination theories. Your AI acceptable use policy template should reference this record-keeping requirement explicitly so managers know it's not optional.

One practical step you can take before 1 July 2027: audit every performance management, attendance, and productivity platform your HR team uses today. For each one, answer three questions: Does it produce a score, flag, or recommendation that a manager sees before making a discipline or termination decision? Can you identify exactly which data fields it used for a specific employee? Can a manager independently evaluate the underlying data without relying on the platform's summary? If the answer to any of those is no, you have a gap to fix. For more on controlling which AI tools your team uses in the first place, see our guide on shadow AI and how it enters employment workflows.

About Shadow AI Policy: We build AI acceptable use policy tools for HR and operations teams at 50–500 person companies. We publish guides on shadow AI, acceptable use policies, and AI governance, updated as regulations and AI tools change.