California's robo-boss ban works by inverting the burden of proof

On 30 September 2026, in the closing days of the legislative session and on what local reporting describes as his last day to sign or veto before leaving office in January, Governor Gavin Newsom signed SB 947, chaptered as Chapter 859, adding Part 5.5.5 to Division 2 of the California Labor Code.

The law is new Labor Code sections 1520 through 1526.7. It governs the use of "automated decision systems" (ADS) — defined in § 1520(c) as "any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues simplified output, including a score, classification, or recommendation, that is used to assist or replace human discretionary decisionmaking and materially impacts natural persons."

Three duties and three prohibitions do the real work:

• § 1522(b)(1) — an employer "shall not rely solely on an ADS when making a disciplinary or termination decision."
• § 1522(b)(2) — where an employer "primarily relies upon an ADS output" for such a decision, it "shall direct a human to corroborate the decision using data that was collected or used to produce the ADS output or other relevant corroborating or supporting information."
• § 1522(c) — if the employer cannot corroborate, or the human reviewer concludes the output is "inaccurate, incomplete, or misleading," the employer "shall not use the ADS output to make a disciplinary or termination decision."
• § 1522(d) — the employee may request, and the employer must provide, "a meaningful, objective description of the employee's own data used by the ADS."
• § 1524 — the employer "primarily relied upon" an ADS must give the affected employee a written post-use notice at the moment it tells the employee of the decision, containing four mandated elements.
• § 1522(a) — three categorical prohibitions with no human-review escape hatch: no ADS use to prevent compliance with or violate labour, occupational-safety, employment or civil-rights law; no ADS inference of an employee's protected status under Government Code § 12940 (FEHA); no ADS use to predict and take adverse action against a worker for exercising legal rights.

Separately, in the same session Newsom signed SB 951 (Cal-WARN notice when AI causes a mass layoff, relocation or termination), AB 1331 (no workplace surveillance tools in workplace bathrooms), AB 1883 (restrictions on AI workplace surveillance, including emotion inference and neural data), and AB 1979 (AI advisory only in nursing, no replacement of clinical judgment). He vetoed AB 2656 and AB 2575 the same day.

The sequel, not the debut

SB 947 is a narrowed re-run of SB 7 (2025), which Newsom vetoed on 13 October 2025. His veto message called the earlier bill's "unfocused notification requirements on any business using even the most innocuous tools" and its "overly broad restrictions." He also pointed to duplicative existing regulation — the California Privacy Protection Agency's newly finalised ADMT regulations. McNerney then re-introduced the bill in February 2026 and removed the pre-use notification requirement, dropped gig-worker coverage that had drawn heavy criticism from Uber and Lyft, and tightened the ADS definition. The 2026 version is what the Governor signed.

The headline change is that "a robot fired me" stopped being a grievance and became a statutory cause of action — with a defined burden-shifting rule attached.

Five specific deltas matter:

1. A flat prohibition on sole reliance. § 1522(b)(1) is unconditional: no employer may rely solely on an ADS output for a disciplinary or termination decision. This is not conditioned on any finding about how central the ADS was.
2. A separate, affirmative corroboration duty at the "primary basis" threshold. § 1522(b)(2) is triggered by primary reliance. This is a materially different and lower standard than sole reliance, and the statute keeps the two tiers apart. Any secondary reporting that collapses them into one "human-in-the-loop" phrase loses the operative distinction.
3. A mandatory, statutorily prescribed notice with four fixed elements. § 1524(b) requires the notice to affirmatively state that the employer primarily relied on an ADS, that "a human reviewed the decision and corroborated the ADS output," a human contact, and the anti-retaliation statement. This is a prescribed certification, not a free-form memo.
4. A new employee data right. § 1522(d) creates a request-and-response right to "a meaningful, objective description of the employee's own data used by the ADS."
5. An evidentiary reversal in litigation. § 1526.1(c): in any civil action or administrative proceeding for a violation of § 1522(b) or (c), "once it has been demonstrated that an ADS was used to make a disciplinary or termination decision, the employer must demonstrate that the employer did not primarily rely upon an ADS or that the employer complied with Sections 1522 and 1524." The burden of proof shifts to the employer once ADS use is shown. Sanctions are $500 per violation, with preliminary injunctive relief, punitive damages and attorney's fees available (§ 1526.1(d)–(e)).

The four precision points that matter most

(a) "Sole basis" vs "primary basis" are not interchangeable. framing ("Where an ADS is the primary basis for such a decision, the employer must have human corroboration, provide written notice, and give the employee access to the data used") is directionally right but incomplete. Per the chaptered text, corroboration, notice and data access attach at primary reliance; a flat bar attaches at sole reliance; and both duties attach only to disciplinary and termination decisions. The corollary matters commercially: a workflow in which an ADS is one of several inputs but is not primarily relied upon is still lawful — but only if the employer can prove that under a burden-shifting rule.

(b) The bill does not cover other adverse actions. The corroboration, notice and data-access duties reach discipline and termination only. They do not reach hiring, promotion, compensation, scheduling, or deactivation of gig workers. Those remain governed by other instruments (CCPA ADMT regulations, FEHA, the Civil Rights Council's ADS rules, and the new Cal-WARN expansion in SB 951). The "AI-driven firing and discipline" framing in the story title is therefore accurate, and importantly narrower than a general "AI hiring law" reading.

(c) The employer definition is far wider than private tech. Under § 1520(d), "employer" includes "all branches of state government, including all cities, counties, charter counties, municipalities, charter municipalities, cities and counties, special districts, transit districts, the University of California upon agreement by the regents, the California State University, community college districts, school districts, or any other governmental entity" — plus labour contractors, farm labour contractors (§ 1682) and foreign labour contractors (§ 9998.1 of the Business and Professions Code). This reaches public-sector HR and the staffing supply chain, not just Silicon Valley vendors.

(d) The nickname is not a statutory short title. The chaptered text carries no short title: it reads "An act to add Part 5.5.5 (commencing with Section 1520) to Division 2 of the Labor Code." "No Robo Bosses Act of 2026" is the name used by the sponsor's press release and by press coverage. Treat it as a nickname that has become the de facto handle.

--

Definitions that do the gating (Lab. Code § 1520)

The decision gate, in the order the statute orders it

Adverse action contemplated (disciplinary or termination only)
        │
        ├── §1522(a) screen ────────────────────────────────  FLAT PROHIBITIONS
        │   • would it prevent compliance with / violate labour,
        │     occupational-safety, employment or civil-rights law?
        │   • does it infer protected status under Gov. Code §12940?
        │   • does it predict & take adverse action against a worker
        │     for exercising legal rights?
        │        → any YES = prohibited outright. No review cures it.
        │
        ├── §1522(b)(1) ─────────────────────────────────────  SOLE-RELIANCE BAR
        │   Is the ADS the ONLY basis for the decision?
        │        → YES = prohibited. Flat. No materiality threshold.
        │
        ├── §1522(b)(2) ─────────────────────────────────────  CORROBORATION
        │   Does the employer PRIMARILY rely on the ADS output?
        │        → YES: direct a human to corroborate, using data
        │          collected or used to produce the ADS output, or
        │          other relevant corroborating/supporting info:
        │          (A) supervisory or managerial evaluations
        │          (B) personnel files
        │          (C) work product of employees
        │          (D) peer reviews
        │          (E) witness interviews (may include relevant
        │              online customer reviews)
        │
        ├── §1522(c) ────────────────────────────────────────  STOP RULE
        │   Could not corroborate? OR human reviewer found the
        │   output inaccurate / incomplete / misleading?
        │        → YES: the ADS output may NOT be used for the
        │          decision at all.
        │
        ├── §1524 ───────────────────────────────────────────  POST-USE NOTICE
        │   At the moment the employer informs the employee of the
        │   decision, issue a separate, stand-alone written notice:
        │     (1) plain language
        │     (2) in the language used for routine employee comms
        │     (3) simple easy-to-use delivery (email, hyperlink,
        │         other written format)
        │   Content, all four required:
        │     (1) that the employer PRIMARILY relied on an ADS
        │     (2) that a human reviewed the decision and CORROBORATED
        │         the ADS output
        │     (3) a human contact for more information, including
        │         the employee's data-description right
        │     (4) that retaliation is prohibited
        │
        └── §1522(d) ─────────────────────────────────────────  DATA RIGHT (on request)
            Employee requests a "meaningful, objective description"
            of their own data used by the ADS → employer must provide.
            §1522(e): the description must anonymise other individuals'
            personal information.

Enforcement architecture (§§ 1526–1526.7)

The exemptions, precisely

CBA waiver (§ 1526.5). The part does not apply to parties covered by a collective bargaining agreement only if the agreement all three: (1) explicitly waives the part "in clear and unambiguous terms"; (2) expressly provides for the wages or earning, working conditions, and other terms and conditions of work; and (3) provides protection from algorithmic management. This is a three-part conjunctive test, not a general union opt-out — a materially underexamined detail in most coverage.

Federal necessity (§ 1526.6). No application to the extent ADS use is required by, or reasonably necessary to comply with, a federal statute, federal regulation or binding federal contract relating to (1) development of aircraft for use in the national airspace, or (2) development of products or services for national security, military, space or defense purposes. Subsection (b) narrows this further: the carve-out applies only to the operations covered by the listed federal instruments.

CPPA carve-back (§ 1526.4). A business subject to the CCPA remains subject to any privacy-related ADMT regulation duly adopted by the California Privacy Protection Agency under Civil Code § 1798.185 and § 1798.199.40(b). Confirmed as live regulation: CPPA lists a completed September 2025 package, "CCPA Updates, Insurance, Cybersecurity Audits, Risk Assessments, and Automated Decisionmaking Technology (ADMT) Regulations."

Interstate notice safe harbour (§ 1526.3). Compliance with this part's notice requirements displaces "substantially similar notice provisions related to automated decision systems used for employment-related decisions" under other state law — but expressly not for "quotas, as defined in Section 2100," or any other automated standards applied to working conditions.

INTERPRETATION, grounded in confirmed statutory text.

1. It regulates the process, not the outcome. FEHA and anti-discrimination law already made discriminatory algorithmic outcomes unlawful. SB 947 makes it unlawful for the outcome to be produced by a specific process — one where the machine was the sole or primary basis. That is a categorically different compliance object: it cannot be discharged by an outcome-testing regime (bias audits, disparate-impact analysis) alone.
2. The burden shift makes it enforceable in practice. A worker cannot easily prove what the employer "primarily relied on" — that fact lives in the employer's systems. § 1526.1(c) resolves the information asymmetry by requiring the employer to demonstrate the negative once ADS use is shown. This is the single most consequential procedural provision in the bill.
3. It reaches a compliance surface nobody had instrumented. The corroboration list in § 1522(b)(2) — supervisory evaluations, personnel files, work product, peer reviews, witness interviews — is an inventory of exactly the artefacts HR systems rarely retain per-decision in a reviewable form. Employers will have to prove, per decision, that corroborating material existed and was considered.
4. It converts "human in the loop" from a marketing claim to an auditable control. A vendor that advertises "human-in-the-loop review" is now asserting something with statutory consequences. The law does not require independence, authority or expertise — but it does require the employer to say in writing that a human reviewed and corroborated, under penalty of per-violation penalties and punitive damages.
5. It is deliberately the first. Combined with the same-day package (SB 951 mass-layoff notice, AB 1883 emotion/neural-data surveillance restrictions, AB 1331 bathroom surveillance, AB 1979 clinical judgment), California has assembled a complete employment-AI legislative stack roughly one year before its federal counterparts exist. Per CNBC, federal legislation carrying the same name (Sen. Ed Markey and Sen. Brian Schatz, June 2026) was "effectively dead on arrival," and New York, Louisiana and New Jersey had proposals but had not acted.
6. The political economy matters. The Governor signed this on his way out of office, in a session where he signed an EO requiring state agencies to keep saying "artificial intelligence" rather than "Super Intelligence." The bill became law in an environment of high public distrust of workplace AI (CNBC cites a July Gallup poll: 39% of Americans say AI does more harm than good, up 8 points from 31%; and a Pew study released that August: 71% think AI will take people's jobs). PREDICTION, low confidence: this signing is more likely to be read as a positioning statement about 2028 than as a technical compliance milestone.

What genuinely became possible on 2026-09-30:

• A statutory cause of action with a workable evidentiary path. A terminated worker who can show any ADS was used in the decision gets the burden shifted. Combined with $500/violation penalties and punitive damages, this makes class-adjacent exposure real.
• A discoverable compliance surface for adversarial use. Antitrust, employment and consumer plaintiffs now have a statutory hook — "the employer primarily relied on an ADS" — that does not require proving discrimination.
• A compliance product category. Per-decision corroboration records, ADS-output lineage, "primarily relied on" classification, prescribed notice generation, and data-description response workflows become sellable enterprise software. This is the clearest new business surface the story creates.
• A bargaining chip. § 1526.5's CBA carve-out (waiver plus explicit working-terms coverage plus algorithmic-management protection) gives unions a concrete template: "we will waive the statute only if our CBA bans algorithmic management." Unions can now price the statute's protections.
• A template other states can copy. A narrowed, veto-proof, union-sponsored, 29–10 / 53–14 bill with an explicit interstate-notice safe harbour (§ 1526.3) is a genuinely portable artifact. California solved the multistate-conflict problem that stalled other states.
• A defensible "we don't rely on robots" claim. Employer marketing in California now has a statutory backing.

What did NOT become possible:

• Employers cannot lawfully automate discipline or termination. The gate is a stop, not a speed-up.
• There is no right to pre-decision notice, no pre-decision data access, no appeal or independent-review mechanism, no right to an explanation of the reasoning, and no audit requirement. All of those remain absent.
• Nothing changed for hiring, promotion, compensation or scheduling under this statute.
• Nothing changed for gig contractors classified as independent contractors under Proposition 22. KQED's reporting is explicit that rideshare drivers were "notably left out of the SB 947 party," and CNBC confirms McNerney "stripped language that would have extended the Act's protections to gig workers."

1. Will the Labor Commissioner issue regulations, model notices or interpretive guidance before 1 July 2027? Nothing in the Act requires it.
2. What does "primarily relies upon" mean as a matter of evidentiary burden when the record will be produced by the employer? This is the whole case.
3. Is a corroborating human reviewer who is the author of the original performance documentation — the manager whose write-up seeded the data — a meaningful corroborator under § 1522(b)(2)?
4. Do § 1522(a)'s categorical prohibitions trigger on feature presence (an unused but deployed protected-status inference model) or on use?
5. Does the § 1522(d) "meaningful, objective description" reach model weights, prompts and feature importances, or only data points? The statute says "data," and § 1520(d)(4) says "any information that identifies, relates to, or describes an employee," which points toward the broader reading.
6. Does § 1522(e) anonymisation permit partial redaction that preserves the explanation, and who decides?
7. Does § 1526.1(c)'s burden shift apply in arbitration, or only in "civil action[s] or administrative proceeding[s]"? It names only those two forums — a gap for California employers with mandatory arbitration clauses.
8. Is "disciplinary" narrower than "discipline" in § 1524(a)? The statute uses "disciplinary" without definition, and a manager's write-up might be either a "discipline" (covered) or performance management (not covered).
9. Do school districts and UC (§ 1520(d)(2)(B)) face special constraints from other law — e.g. constitutional due process or statutory tenure protections — that conflict with § 1526.1(a)'s interim relief?
10. Will § 1526.3's "substantially similar" preemption be asserted against New York, New Jersey or Massachusetts laws, and who arbitrates that?
11. Are CBA waivers under § 1526.5 already being negotiated into 2027 contracts, and what does "protection from algorithmic management" have to say to be sufficient?
12. Will § 1526.6's federal-necessity carve-out be read expansively by defence contractors using it as a general shield?
13. Will anyone test whether a federal No Robo Bosses bill revives, and would a federal statute preempt § 1526.3's interstate scheme?

CONFIRMED / scheduled

• 1 July 2027 — § 1526.7: the Part becomes operative. All duties and prohibitions apply.
• The Labor Commissioner gains authority under § 1526.1(a) to investigate, order interim relief and issue citations "through the procedures set forth in Section 98.3, 98.7, 98.74, or 1197.1." Nothing in the Act schedules guidance, model notices or interpretive rules.

PREDICTION — near-term watch items (labelled as prediction, not fact)

1. No Labor Commissioner guidance before 2027 is plausible, and its absence would be the single biggest enforcement variable. Expect employer-side law firms and consultancies to produce the first interpretive artefacts instead.
2. First test of "primarily relies" will arrive as a summary-judgment motion on a § 1526.1(c) burden shift, likely in Los Angeles or San Francisco, within 12–24 months of the operative date. PREDICTION, moderate confidence on timing; low confidence on outcome.
3. Interim status-quo relief is the sharpest near-term exposure. A Labor Commissioner ordering a termination or suspension to be reversed pending investigation is a far larger cost than $500 per violation.
4. Vendor product cycles will be visibly reorganised around California by mid-2027: reliance classification, corroboration records, notice generation, data-description response. PREDICTION, moderate-high confidence.
5. At least one state will introduce a copy during the 2027 sessions, and the interesting contest will be over whether to adopt § 1526.3's interstate notice preemption. PREDICTION, moderate confidence.
6. Proposition 22 gig-worker exclusion will be litigated or legislated, because it is the most visible gap in an otherwise labour-movement-sponsored bill. PREDICTION, moderate confidence.
7. A challenge on § 1522(a)(2)'s breadth is likely, on the theory that it bans a capability rather than regulating conduct. The statute's own words — "shall not use an ADS to infer an employee's protected status" — read as capability-neutral but are capable of capability-level reading. PREDICTION, low-moderate confidence.
8. Enforcement will run through the Labor Commissioner and public prosecutors long before private employment litigation. § 1526.1(c)'s burden shift is most useful to a plaintiff's counsel with records access; the Commissioner route needs no private bar. PREDICTION, moderate confidence.

The headline was right and the operative text is narrower and stranger than the headline.

Three things are worth taking away in order.

First, the actual new thing is a burden shift, not a "human in the loop." Every headline in this story's coverage — mine included — describes a human-in-the-loop requirement. The statute's most consequential provision is § 1526.1(c): once it is demonstrated that an ADS was used to make a disciplinary or termination decision, the employer must prove it did not primarily rely on the ADS, or that it complied. That reverses the ordinary logic of employment litigation, where the worker must prove the employer did something wrong and the employer holds the records. It is why a $500-per-violation penalty is not a small number: it is a multiplier applied to a burden the employer cannot easily discharge.

Second, "first US law" is defensible but only if you say what you mean. California is the first state statute to bar sole reliance on an automated decision system and require human corroboration for firing and discipline. Illinois has required notice since January 2026; New York City has required bias audits for automated hiring tools since 2021; Colorado's AI Act does not reach employment discipline; the EU AI Act reaches this ground ex ante rather than through a corroboration duty. The press release saying "first such law in the nation" is a stakeholder claim; the version that survives scrutiny is "first state statute of its kind." Note also that KQED, reporting from California, chose the softer "some of the strongest in the country." That editorial restraint is informative.

Third, the most consequential line in the bill is not about robo bosses. It is § 1522(a)(2): an employer "shall not use an ADS to infer an employee's protected status under Section 12940." That is a capability prohibition aimed at feature engineering, not at decision review. It is the provision most likely to generate surprise, because almost nobody outside the bargaining room described it as the point of the bill. A statute announced as being about robot firing contains its sharpest rule about what data you may model at all.

And the quiet cost. Proposition 22 contractors were removed from the bill after Uber and Lyft objected (CNBC). So the workers KQED quotes as most exposed to algorithmic management — rideshare drivers whose work, pay and deactivation are algorithmically determined — receive none of this protection. The statute reaches employees, public-sector staff and labour contractors, and stops at the platform boundary. Whether the line falls there for a principled reason or a lobbying reason is the question worth carrying into every other state's copy of this law.