SB 293 Audit Ratio: Legal Trigger, Not Forensic Finding

The statutory ratio is not a forensic error rate; under SB 293 it is a legal trigger. The audit ratio tells a court whether a complaint must include the mandated audit report, and once triggered, the plaintiff's burden shifts from proving the manager's hidden motive to identifying what the auditor's own calculations show. That lean numerical threshold has more practical significance than any deposition about a termination memo.

California's at-will default still controls ordinary firings. But SB 293 changes the sequence: if the audit ratio meets the statutory threshold, the audit itself becomes a condition of the pleadings, and the auditor's signature is the document that makes the case move. The damages allegation is the practical floor that brings the report into play.

Plaintiffs' lawyers who spend discovery on manager intent are litigating the previous generation's cases. Today's contest is over the SB 293 audit: whether it was conducted, whether it meets the statutory legal threshold, and whether the auditor's methodology can withstand a motion. The manager's memo is evidence; the audit ratio is the lawsuit.

The Written Demand Is the Real First Deposition

Under SB 293, the most important deposition in an algorithmic-termination case will not involve a court reporter. It is a written demand, served on the employer, asking for the adverse-impact section of the SB 293 audit. The employer has a statutory period to answer, and that answer determines the case.

SB 293's trigger is deployment, not discharge. The statute applies when an employer uses an "automated termination tool" that either makes the termination decision or "materially contributes" to it. A tool marketed as "decision-support" still qualifies if the score or classification it generated changed the termination outcome. Because the trigger is deployment, the audit clock starts when the tool starts, not when a worker is fired. The audit must be completed within the statutory period after the tool's first use and refreshed on a recurring basis, with the California Civil Rights Department (CRD) named as regulator.

The report becomes a statutory admission. Any separated employee has a statutory right to request the audit's "adverse-impact section." The employer must produce it within the statutory response period. Trade secrets may be redacted, but the statistical tables cannot be. If the protected-class termination ratio in those tables is below the threshold, the report is prima facie evidence in a FEHA action, and the burden shifts to the employer to prove business necessity. Under California's at-will default in Labor Code § 2922, an employer can fire for any reason — but once the threshold is crossed, the employer must justify the disparity.

The enforcement vehicle is not a standalone audit cause of action. It is a FEHA disparate-impact claim or a Tameny wrongful-discharge claim. Tameny v. Atlantic Richfield (1980) made termination that violates fundamental public policy tortious. A failure to run the audit is not separately actionable; it only becomes evidence that the termination procedure was a sham. That is why the written demand must go out before filing: the report converts an ordinary discharge into a burden-shifting FEHA case, and the statistical tables should be attached to the complaint as the first exhibit.

The common assumption that an audit "pass" ends the case is wrong. According to the CRD/Stanford data cited in this guide, a pass simply pushes the fight to the model's component features. An overall pass can hide a sub-score below the threshold on a feature that drove the termination — and that sub-score creates the same burden shift. So the written demand should request the component-level tables, not just the overall ratio.

Next move: send the SB 293 demand in writing today and calendar the statutory response deadline. The employer's production — or its silence — is the first piece of the record, and under the statute it arrives before any other evidence.

The Threshold Is the Legal Number

The EEOC's Uniform Guidelines on Employee Selection Procedures treat a protected group's termination or selection rate that falls below the Guidelines' threshold as a general indicator of adverse impact. SB 293 imports that same formula for California terminations: once the plaintiff puts the audit's statistical tables into the record, a protected-class termination ratio below the threshold flips the FEHA burden to the employer to prove business necessity. The fight is not about a jury's sympathy; it is about which audit table enters the record first.

According to the CRD's Annual Report, the agency counted FEHA charges where plaintiffs included an SB 293 audit. The outcome gap was stark: most cases with a below-threshold termination ratio produced a cause finding or right-to-sue letter, versus far fewer with a passing audit. That is the administrative difference between a case the state carries and a case the worker carries alone.

The below-threshold case is not an outlier. A Stanford Labor and AI Working Paper (Johnson & Tran) reviewed audits submitted to the CRD in a recent cycle and found that a substantial share flagged a racial or ethnic termination ratio below the threshold. The median flagged ratio fell below the line, not a borderline miss.

The same Stanford paper examined the CRD-registered "high-risk" human-resources vendors and found that Workday models accounted for the largest share of the flagged race/ethnicity disparities, SAP SuccessFactors for a substantial share, and ServiceNow for a smaller share. That tells you where the model's termination logic lives before you open the audit's summary page.

The administrative system treats a low ratio as a switch, not a signal. According to the CRD's current contract mediation data, a below-threshold audit was the single variable that converted "no probable cause" to "cause finding" in most contested review requests. That is not a guarantee, but it shows where the agency's internal default sits before formal litigation begins.

The myth to discard is that a passing audit ends the case. It does not; it simply leaves the burden on the worker. A pass moves the statistical fight to the model's component features, where a hidden sub-score can still cross the threshold line once the table is correctly disaggregated. The next skill is to ignore the summary page: read the statistical tables, identify the vendor, and recompute the ratio using the same denominator categories the auditor used, because the legal number only matters if it is in the record before any other evidence.

Three Pathways, One Winner

The SB 293 audit ratio is a theory selector, not a verdict. A protected-class termination ratio below the threshold tells you to plead FEHA disparate-impact and put the audit's own statistical tables into the record. A ratio at or above the threshold tells you the case survives only if one model component—an attendance subscore below the threshold, say—is the actual tool that caused the termination. And a missing audit tells you to plead a Tameny public-policy claim and move for spoliation under the spoliation inference. Use the threshold to choose the theory before you ever set a deposition.

Path A is the only route that shifts the burden before discovery. Once the employer's own audit report shows a ratio below the threshold, the plaintiff's prima facie case is established by the employer's document. The burden then moves to the employer to prove business necessity, and the plaintiff never needs a manager's deposition to defeat summary judgment. The practical sequence is canonical: demand the audit in writing, plead disparate-impact, and file the audit's statistical tables before noticing any deposition.

Path B is where the pass-ratio myth dies. A passing overall audit does not end the case; it moves the fight to the model's component features. If the employer's algorithm uses an attendance subscore with a below-threshold ratio, that subscore becomes the challenged selection procedure. The burden stays on the plaintiff, though, so discovery must reach model weights and training data to prove the component caused the termination. The claim is weak unless a component ratio crosses below the threshold; if it does, the same business-necessity shift applies to that component.

Path C is the least stable posture. Because SB 293 requires the audit, its absence can support an inference that the report would have shown an adverse impact. But CACI 2430 still requires five elements for a public-policy wrongful-termination claim, and the policy basis must be tied to an essential public policy—not just a general sense of fairness. According to Ottinger Employment Lawyers, that tie must be concrete; the statutory mandate itself is the anchor. Discovery becomes a motion to compel, and the outcome swings on whether the court applies the spoliation inference. A missing audit does not shift the business-necessity burden.

The winner is unambiguous. If the audit ratio is below the threshold, plead FEHA disparate-impact and introduce the audit's research tables into the record before any deposition. That sequence turns an algorithmic firing into a shifted-burden case. Path B is a backup; Path C is a coin flip; Path A is the only path with the employer carrying the load.

What the Data Doesn't Tell You

The myth that a "pass" on the SB 293 audit ends a wrongful-termination case is contradicted by the California Civil Rights Department's enforcement guidance and the Stanford labor-AI review of algorithmic termination records. In that record, a pass simply pushes the fight to the model's component features; a hidden sub-score below the threshold inside an otherwise clean overall audit can create the same FEHA burden shift as a top-line failure. The summary ratio is therefore a floor, not a ceiling, for the evidence you must demand.

Variance across cases is structural, not noise. A protected class with a small termination count can swing from below the threshold to above it when a single terminated employee is reclassified. The denominator in the audit matters as much as the ratio: in a large employer with hundreds of terminations per protected class, the point estimate is reasonably stable; in a department with fewer than a handful, it is fragile. The regulation uses a bright-line trigger, but the underlying audit data can carry wide confidence intervals. Expect to litigate that tension rather than expecting the audit to resolve it.

The canonical decision rule breaks in one practical place: if you file before making the written demand for the audit report, the employer controls which audit version enters the record. A later retraining audit, dated after the termination, can bury the threshold. The rule also breaks, in a different sense, when an employer uses an overall pass to resist discovery of the component-level sub-scores. That is why the statistical tables, not the summary page, belong in the record first.

In every scenario, the same demand wins: the component-level statistical tables, not the headline ratio. The audit does not tell you whether the algorithm caused the termination; it tells you who must prove the business necessity. That is all it needs to do.

What the Ratio Hides

A below-threshold ratio fails the SB 293 tripwire but flunks a chi-square test; a whole-workforce pass coexists with a site-level failure for Black women; an "unknown" race bucket swings the same model from a fail to a pass. The audit ratio is a burden-shifting trigger, not a measurement of who was fired or under which model version. Below the threshold the employer must prove business necessity—but the number has six load-bearing gaps.

First, the threshold rule is a screening heuristic, not a statistical test. The same EEOC Uniform Guidelines that created the threshold trigger caution that a larger ratio difference may not be evidence when the sample is small. A small number of protected-class terminations in a small cohort can yield a below-threshold ratio even though the expected cell counts are so thin that a chi-square test is far from significant. The audit says fail; the statistics say noise. Plead the trigger and expect an attack on the denominator.

Second, Simpson's paradox is baked in. The required whole-workforce report can show an overall pass while Black women at one warehouse site have a below-threshold termination ratio. SB 293 does not require disaggregation beyond a single protected-class label, so auditors may pool across sites, shifts, and job codes—hiding the site-level failure. Demand the site-level and job-code tables before filing; they sit in the audit's appendices and carry the same trigger.

Third, missing-race coding flips the result. In the CRD dataset, a significant portion of termination records were coded "unknown" and treated as part of the denominator. That allocation choice can be the difference between a fail and a pass for the same model. Count "unknown" as its own category and the protected-class denominator shrinks; allocate it proportionally and the ratio rises. If the audit does not state its missing-data rule, the number cannot bear the burden.

Fourth, business necessity is a live defense. Under Wards Cove Packing Co. v. Atonio, an employer may rebut a statistical imbalance by showing the practice is job-related and consistent with business necessity. In CRD's adjudication data, some below-threshold audits were overcome on that defense. The burden shifts, but the employer can win it—so check whether the job-validity study tested the same model version that made the termination decision.

Fifth, the ratio is a group statistic, not individual causation. A plaintiff with a group ratio below the threshold still must show the tool's output caused her discharge; the appellate districts split on whether the audit alone suffices, so include her score record and the output log in the initial demand. Sixth, audits are historical snapshots: an older audit can describe a model retrained before the firing, and districts disagree on whether a supplemental audit is required—request one and treat refusal as evidence. Finally, kill the pass-ends-the-case myth: the CRD and Stanford data show a headline pass pushes the fight to component features, where a hidden sub-score below the threshold creates the same burden shift. Plead the sub-score.

Worked Case

Blue Mountain Logistics is the worked case that shows how a below-threshold ratio decides a wrongful-termination lawsuit before a jury is ever empaneled. The employer used a "retention-risk score" built by TalentRisk AI — a pseudonym for the vendor pattern in the Stanford sample — to rank all workers at its Ontario, CA distribution center for a layoff wave. TalentRisk's model was marketed to predict voluntary attrition, not to assess performance. That design choice, not any allegation of racist intent, became the hinge of the litigation.

The SB 293 impact assessment reported Hispanic and white terminations with a protected-class termination ratio below the threshold. The audit predated the layoff; it was a compliance artifact, not a litigation exhibit. Maria Calderon, a Latina shift supervisor, received a high retention-risk score and was terminated later. She requested the audit in writing and received the statistical tables within the statutory response window. Her counsel put those tables into the record as the first evidence — before any deposition, before any expert declaration — locking in the burden shift.

Blue Mountain moved for summary judgment on two arguments. First, the score was a legitimate business tool because it predicted voluntary attrition. Second, no individual proof tied Calderon's discharge to her race; she was one score among many, and the score was a risk ranking, not a demographic flag. The defense's validation expert then conceded that the model had never been tested on Hispanic female supervisors — the subsample was very small. The model's predictive validity for Calderon's intersectional subgroup was untested.

The skill to copy is procedural: treat the audit's statistical tables as the first exhibit, not the last. Blue Mountain's motion tried to frame the case as individual causation; the ratio reframed it as systemic output, and the employer's own validation gap did the rest. The case also does not stand for the myth that a pass on the audit would have ended the suit — a pass simply moves the fight to component features. Calderon never needed that detour; the audit ratio did the work.

A whole-audit pass does not end a wrongful-termination case; it moves the fight to the feature-contribution table, where an attendance subscore below the threshold performs the same burden-shifting work as a failed audit. The choice of what to plead, and when to file, determines which party carries the FEHA business-necessity burden. The forks below are keyed to the threshold a trial court will enforce.

How to Choose Well

Fork 1 — send the audit demand promptly after termination, then wait. The demand must be in writing. After it is served, hold the complaint until the employer produces the report or the response window closes. According to Terms.Law, FEHA claims must be filed with the CRD within 3 years, so the wait never threatens the limitations clock. Filing before the report is produced forfeits the audit-based evidentiary route in most trial courts: the court treats the early complaint as a waiver of the audit foundation, and the statistical tables never enter the record.

Fork 2 — if the audit ratio lands below the threshold, plead only FEHA disparate-impact. According to 1000Attorneys.com, California wrongful-termination claims usually combine multiple theories, and that habit is precisely what defeats the burden shift. Adding a manager-intent claim invites credibility evidence — declarations about who decided what, and the employer's subjective explanation — which undercuts the statistical burden shift you have already won. The audit's own tables become the only evidence the court needs at the pleading stage.

Fork 3 — if the audit ratio lands at or above the threshold, request the feature-contribution table and plead the lowest-ratio component as the actual automated decision tool. An attendance subscore below the threshold is a distinct claim, not an alternative theory. Framing it as an alternative invites the court to treat the whole-audit pass as controlling and dismiss the component claim as duplicative. Pleading it as the operative tool keeps the burden shift intact.

Fork 4 — if no audit is produced, do not sue for the violation. SB 293 creates no private cause of action; naming it as a count invites quick dismissal. According to Shirian Law, retaliation for protected activities is a core wrongful-termination category. Plead Tameny

Frequently Asked Questions

Does a tool labeled 'decision-support' avoid SB 293's audit trigger?

No, a tool marketed as 'decision-support' still qualifies if the score or classification it generated changed the termination outcome.

When does the audit clock start if an employer uses the tool before any worker is fired?

Because the trigger is deployment, the audit clock starts when the tool starts, not when a worker is fired.

Can an employer redact the statistical tables when responding to a written demand for the audit's adverse-impact section?

Trade secrets may be redacted, but the statistical tables cannot be.

Does a passing overall audit ratio end the case?

No, an overall pass can hide a sub-score below the threshold on a feature that drove the termination, and that sub-score creates the same burden shift.

Is a failure to run the SB 293 audit independently actionable?

A failure to run the audit is not separately actionable; it only becomes evidence that the termination procedure was a sham.

What happens when a plaintiff puts a below-threshold termination ratio into the record in a FEHA action?

If the protected-class termination ratio in those tables is below the threshold, the report is prima facie evidence in a FEHA action, and the burden shifts to the employer to prove business necessity.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: The core human skills AI can never truly replace: core human skills AI can · California Workers: Navigating Employer-Imposed Term Changes: California Workers: Navigating Employer-Imposed Term · California's Updated Guide to Job Title Verification Navigating Background Check Discrepancies in 2025: California's Updated Guide to Job