Employment & Labor Law

Defending the Algorithm® AI and Employment Law Series | Edition 2

Welcome back to Defending the Algorithm® - a LinkedIn newsletter from Pittsburgh law firm, Houston Harbaugh, PC, helping defense attorneys, insurance professionals, employment lawyers, corporate counsel and clients to navigate the intersection between artificial intelligence and the law. This newsletter series is specifically targeted to the defense of employment law claims against employers and their vendors and was written and edited with assistance from Claude Opus 5.0 from Anthropic and Google Gemini 3.0 Pro and with research confirmation by Westlaw Advantage AI with Co-Counsel. It is a companion to our podcast and blog series, available at: Defending the Algorithm® created by Henry M. Sneath, Esq.

In cooperation with the DRI Canter for Law and Public Policy; AI tools used in research: Claude (Anthropic), Google (Gemini) and Westlaw Precision AI. All content was reviewed and verified by the authors.

The Black Box Stays Partially Closed

In this Edition #2, we give an update on the AI-discovery battle in Mobley v. Workday, 3:23-cv-00770-RFL (N.D. Cal.) . As you may recall, this case is brought by inter alia, lead class plaintiff  Derek Mobley, a man who allegedly applied to more than one hundred jobs through Workday’s hiring platform and was rejected from every one. He sued Workday, the vendor whose AI screening system, he alleges, discriminated against him due to his race, age and his claimed disabilities before any of the employers actually saw his resume. The case has since morphed into a nationwide collective action, with the putative class action plaintiffs alleging employment-discrimination. Recently, a tense discovery battle erupted, whereby the plaintiffs moved to compel Workday to produce LLM bias-testing code, results, and underlying data and, separately, documents relating to Workday’s compliance with certain federal employment statutes.

The Discovery Dispute

With respect to algorithmic bias-testing, the Court summarized the relevant factual background of the information sought by the plaintiffs as follows:

Workday Recruiting is a software product that helps customers manage hiring. Customers who purchase Workday Recruiting have access to Candidate Skills Match, an algorithmic feature that determines the extent to which an applicant’s skills match the role to which they applied. In 2024, Workday acquired HiredScore, allowing Workday to offer additional features to customers, including Spotlight []. Spotlight is a candidate review tool that integrates with an Applicant Tracking System like Workday Recruiting and allows customers to assess the level of match between a candidate’s application materials and the job requirements. Workday . . . conducted bias testing of Candidate Skills Match at the direction and under the guidance of its legal counsel for the purpose of rendering legal advice regarding this product. Workday restrict[ed] access to its bias testing, including the underlying data, testing code, and results, to the individuals who need[ed] it for giving legal advice.

In other words, the plaintiffs sought the heart of Workday’s algorithmic method and the guts of its AI program. The parties fully briefed the issues, and in a well-reasoned opinion, United States Magistrate Judge Laurel Beeler denied the plaintiffs’ motion to compel production of Workday’s bias-testing data, code, and results, holding that all three are protected by the attorney-client privilege. The plaintiffs did not walk away empty-handed — the court ordered Workday to produce its own EEO-1 and OFCCP filings[1] — but on the central question, the one every AI vendor and every employer-customer has been watching, the black box stayed closed, or at least partially closed for now.

This is a ruling that defense counsel in AI civil litigation should study.

[1] The Equal Employment Opportunity Commission (EEOC) and Office of Federal Contract Compliance Programs (OFCCP) are federal agencies that protect individuals from workplace discrimination and harassment. The EEOC handles workplace discrimination complaints, and the OFCCP ensures that federal contractors comply with employment laws and regulations.

The Heart of the Ruling: Attorney-Curated Bias Testing is Privileged

The plaintiffs argued that bias-testing artifacts cannot be privileged because the privilege protects communications, not underlying facts. Therefore, they argued, because code is an engineering artifact written by data scientists and the testing parameters were dictated by regulation rather than by counsel, privilege did not attach. Plaintiffs further argued that Workday waived any privilege in three ways: by publicly touting its bias testing in an “AI Fact Sheet,” by failing to retain its 2023 and 2024 audit data, and by disclosing its 2023 HiredScore Spotlight testing data to an external auditor. Plaintiffs summarized their argument in one memorable line: parking an attorney on a datafile does not confer privilege on technical work product.

The court was not persuaded. The decisive factor was the way Workday structured the program. Workday represented — through a detailed declaration from in-house counsel — that its attorneys curated the data that went into the bias testing, that the overall purpose of the testing was to obtain legal advice rather than to serve a business function, and that the results were never submitted to any regulatory body. On that record, the court held that Workday had shown “more than mere direction by its attorneys,” and that the bias-testing data, code, and results were therefore privileged.

The court’s treatment of the plaintiffs’ lead authority is the part worth studying. The plaintiffs leaned on Upjohnv. United States, 449 U.S. 383 (1981), for the rule that privilege protects communications, not facts. The court agreed with the principle but found it beside the point: Upjohn involved employee-to-counsel communications, not attorney-curated data sets. The more applicable authority, per the court, was Suboh v. Bellsouth, No. 03-cv-996, 2004 WL 5550100 (N.D. Ga. Nov. 17, 2004), which held that raw data underlying a statistical analysis is not privileged, but the attorney’s choices about what to compile, how to organize it, and how to analyze it can reveal counsel’s thought process and are, therefore, privileged. Workday asserted privilege over that curation layer, not over raw applicant data. That distinction carried the day.

Why This Is a Defense Roadmap, Not Just a Defense Win

Although it would be easy to read this as just a clean defense victory, it is more useful to read the decision as a roadmap. Three features of Workday’s setup did the work, and each is something defense counsel can advise a client to replicate before litigation is ever on the horizon.

Counsel curated the inputs. The privilege did not attach because a lawyer was copied on an email, but rather because counsel selected the data to be tested, defined the populations and the comparisons, and used the results to advise the company. The court drew a sharp line between attorney direction (not enough) and attorney curation that reveals legal thinking (enough).

The purpose was legal, and the record proved it. Workday supported its privilege claim with a declaration establishing that the testing existed to inform legal advice, not to satisfy a business or regulatory requirement. Had it prepared or submitted the material for regulatory reasons, it likely would have lost the privilege. See U.S. v. Richey, 632 F.3d 559 (9th Cir. 2011) (holding that materials prepared for submission to the IRS are not privileged).

Workday never used the testing as a sword. The plaintiffs’ strongest argument was waiver — that Workday had touted its bias testing in a public “AI Fact Sheet” and could not then hide the data. The court rejected it because Workday had not argued in the litigation that its tools are unbiased. The lesson is unforgiving: the moment a defendant uses favorable bias-testing conclusions to defend the merits, the privilege over the underlying data is at risk.

Two Cautionary Notes Before Any Civil Defendant Celebrates

First, privilege is a structure built in advance, not a label to be applied after-the-fact. Workday won because the program was designed from the outset to be attorney-directed and confidential, and because it could prove that with a declaration and a privilege log. A client that runs bias testing as a routine compliance exercise, then tries to wrap it in privilege after a charge is filed, will not likely be able to lean on privilege. If anything, the regulatory frameworks pushing employers toward routine bias auditing (e.g., the California FEHA regulations, NYC Local Law 144, the federal Uniform Guidelines) make the business-purpose argument easier for a future plaintiff, not harder.

Second, this was a discovery ruling by a magistrate judge, not a merits decision and not binding precedent. It is persuasive and well-reasoned, and it will be oft-cited, but a different record, a sloppier privilege log, or a defendant who leaned on its testing to defend the merits could flip the result. And the plaintiffs were not shut out: the EEO-1 and OFCCP production the court ordered handed the plaintiffs a federally mandated dataset on Workday’s own workforce demographics, because Workday uses the same AI tools on its own hiring.

Four Takeaways for Defense Counsel

1. If bias testing is meant to be privileged, build it that way from day one. Counsel must curate the inputs, define the methodology, and document a legal-advice purpose. Retroactive privilege claims over routine compliance testing will likely fail.

2. Separate privileged legal-advice testing from regulatory-compliance testing. A company may need both, but conflating them risks losing the privilege over the first. Workday kept its privileged testing out of any regulatory submission, and that mattered.

3. Never use the conclusions as a merits defense unless you will produce the data. Touting “we tested for bias and found none” in litigation is the fastest way to waive the privilege.

4. Mind data preservation and the Rule 34(a)(1) “control” question. Plaintiffs argued Workday failed to preserve some testing data. While the court declined to compel the production of such data because plaintiffs had not shown Workday’s control of it, it is best to control and preserve such data to prevent spoliation allegations.

What Comes Next

On June 24, 2026, the District Judge Rita F. Lin issued an order in which she described Magistrate Judge’s Beeler’s decision as “well-reasoned and not clearly erroneous or contrary to law.” For all intents and purposes, that put the issue to bed – at least for the moment. While the privilege question may resurface later in the litigation, Magistrate Judge Beeler’s order is the most detailed federal treatment to date of how attorney-client privilege applies to AI bias-testing artifacts and it should serve as a roadmap for counsel and clients to structure similar programs going forward. More recently, the purported class plaintiffs filed a fourth Amended Complaint (after three prior Motions to Dismiss by Workday) and Workday has now answered the complaint denying all liability and damages. We will follow the case and report further.

Connect With Us

Authors: Eric M. Spada (Senior Attorney, Litigation, 412-288-2220, spadaem@hh-law.com) and Grant A. Allison (Associate, Litigation, 412-288-2211, allisonga@hh-law.com) with creative and editing assistance from Henry M. Sneath (Shareholder, Business Litigation, AI and IP, 412-288-4013 and sneathhm@hh-law.com)

To discuss AI-related employment exposure or to schedule a vendor-contract review, contact the authors at Houston Harbaugh, P.C., 412-281-5060, or click here to visit our website.

Defending the Algorithm® is a federally registered trademark of Houston Harbaugh, P.C. This newsletter is informational only and does not constitute legal advice or create an attorney-client relationship.

Sources Verified

Discovery Order, Mobley v. Workday, Inc., Case No. 23-cv-00770-RFL (LB) (N.D. Cal. May 28, 2026, ECF No. 340, filed May 29, 2026) (Beeler, Mag. J.) (resolving ECF Nos. 310, 314, and 319). Plaintiffs’ Motion to Compel (ECF No. 310, Apr. 20, 2026); Workday’s Opposition (ECF No. 315, May 4, 2026); Plaintiffs’ Reply (ECF No. 321, May 11, 2026); Declaration of Michelle Law (ECF No. 316). Upjohn Co. v. United States, 449 U.S. 383 (1981). United States v. Ruehle, 583 F.3d 600 (9th Cir. 2009). In re Grand Jury, 23 F.4th 1088 (9th Cir. 2021). In re Pacific Pictures Corp., 679 F.3d 1121 (9th Cir. 2012). Suboh v. Bellsouth Business Systems, Inc., 2004 WL 5550100 (N.D. Ga. Nov. 17, 2004). United States v. Nat’l Ass’n of Realtors, 242 F.R.D. 491 (N.D. Ill. 2007).  Ideal Electric Co. v. Flowserve Corp., 230 F.R.D. 603 (D. Nev. 2005). Dothard v. Rawlinson, 433 U.S. 321 (1977). Fed. R. Civ. P. 72(a) and 28 U.S.C. § 636(b)(1)(A) (district-judge review of magistrate-judge orders on non-dispositive matters). Plaintiffs’ Motion for Relief from Non-Dispositive Order of the Magistrate Judge (ECF No. 351, June 12, 2026); Order setting Workday’s response on the Rule 34(a)(1) possession question (ECF No. 353, June 15, 2026) (Lin, J.). Discovery Order re Deposition Location, Mobley v. Workday, Inc., ECF No. 357 (N.D. Cal. June 15, 2026) (Beeler, Mag. J.) (resolving ECF No. 350); Fed. R. Civ. P. 30(b)(1); Fenerjian v. Nong Shim Co., 2016 WL 1019669 (N.D. Cal. Mar. 15, 2016); Mullins v. Premier Nutrition Corp., 2014 WL 4058484 (N.D. Cal. Aug. 15, 2014); Lexington Ins. Co. v. Commonwealth Ins. Co., 1999 WL 33292943 (N.D. Cal. Sep. 17, 1999). Additional authority cited by the authors but not relied on in the order: United States v. Richey, 632 F.3d 559 (9th Cir. 2011). Federal Rules of Civil Procedure 26(b), 33, 34, and 37(e). California FEHA Regulations on Automated-Decision Systems, Cal. Code Regs. tit. 2, §§ 11008.1–11008.4 (eff. Oct. 1, 2025). NYC Local Law 144. EEOC Uniform Guidelines on Employee Selection Procedures, 29 C.F.R. § 1607.4(D). Workday is represented by Orrick, Herrington & Sutcliffe LLP.