They Said the AI Was 'Just a Guide.' A Judge Disagreed.

Issue 09 | The AI Detective: GovernIntel | April 8, 2026
Circulation: 772 investigators

THE SCENE OF THE CRIME

A judge just gave a health insurer 30 days to hand over tens of thousands of documents. Nine years of records. The algorithm at the center of the case has, according to the lawsuit, a 90% error rate. Elderly patients were denied rehabilitation care — care their own doctors had prescribed — because a machine said no. The company called it "just a guide."

The AI system your legal team calls "a guide" is not a guide. It is a decision-maker — and courts are now treating it exactly like one.

CASE FILE: THE INVESTIGATION

Let me tell you about two cases that landed in the same week, in two completely different industries, with the exact same problem at the center.

The first is UnitedHealth. A Minnesota judge just extended a discovery deadline to April 29, ordering the company to produce tens of thousands of documents related to nH Predict — an AI model UnitedHealth uses to determine how long patients should stay in rehabilitation facilities. The lawsuit alleges that elderly patients on Medicare Advantage plans were prematurely discharged from care, not because their doctors said they were ready, but because the algorithm said their time was up. Internal documents, according to the complaint, show UnitedHealth managers were instructed to keep patient stays within 1% of what the algorithm projected.

A 90% error rate. Manager targets tied to an algorithm's output. Elderly patients depleting family savings or going without care they needed. UnitedHealth has called the tool "a guide." Courts are not accepting that framing.

The second case involves Eightfold AI — a hiring platform used by Microsoft, PayPal, Morgan Stanley, Starbucks, Chevron, and Bayer, among others. A January 2026 class action, brought by former EEOC chair Jenny Yang and the nonprofit Towards Justice, alleges that Eightfold scraped data on over one billion workers, scored every applicant on a zero-to-five scale, and filtered out lower-ranked candidates before a single human being ever saw their name.

Here is what makes the Eightfold case genuinely different from every AI discrimination case that came before it. The plaintiffs are not claiming the algorithm was biased. They are claiming it existed in secret.

The Fair Credit Reporting Act has required companies to disclose when they compile consumer reports used in employment decisions since 1970. Eightfold allegedly did not do that. Every rejected candidate had a right to know a report was created about them, to receive a copy, and to dispute errors. None of that happened. And at statutory FCRA damages of $100 to $1,000 per willful violation, multiply that across a database of one billion profiles.

What connects these two cases is something accelerating across every industry. Companies deploy AI tools. They frame those tools as assistants, guides, support systems. When outcomes are catastrophic and courts start asking questions, the AI becomes a neutral tool that humans were always in control of. That argument is running out of runway.

Meanwhile, the regulatory environment is not waiting for courts to settle these questions. Since mid-March, 19 new AI laws have been signed across multiple states. The total for 2026 is now 25 new laws, with another 27 bills awaiting signature. Colorado's AI Act takes effect in June. The enforcement calendar has dates on it. The question is whether your governance calendar does too.

EVIDENCE LOG

🔍 Finding 1/3: A Minnesota judge ordered UnitedHealth to produce tens of thousands of documents by April 29 in a case alleging its AI model denied elderly patients medically necessary care at a 90% error rate.
What it means for you: Every internal document about your AI system's error rate and human override policies is potentially discoverable.

🔍 Finding 2/3: The Eightfold AI class action targeting platforms used by Microsoft, PayPal, Morgan Stanley and others does not allege bias — it alleges the algorithm existed in secret, affecting over one billion worker profiles.
What it means for you: Your AI hiring tools do not need to be biased to create massive legal exposure. They need only to exist without the right disclosures.

🔍 Finding 3/3: 19 new AI laws were signed across U.S. states in the last two weeks alone, bringing 2026's total to 25 new laws with 27 more awaiting signature.
What it means for you: Multi-directional enforcement from state AGs, private plaintiffs, and federal agencies means there is no single compliance checkbox that protects you.

THE VERDICT

1. Audit every AI system your organization uses for any consequential decision. Document what the system does, who reviewed its outputs, and what disclosures were made to affected individuals.

2. Review every vendor agreement for the word "guide." If a contract describes an AI tool as a guide while that tool is functionally filtering or denying outcomes for real people, your legal team needs to review it immediately.

3. Map your FCRA exposure today. If any AI system in your HR stack compiles data on applicants from sources beyond what they submitted, you may be operating a consumer reporting agency without knowing it.

NEXT ON THE CASE

Next week, GovernIntel goes inside the doctrine quietly rewriting liability for every enterprise deploying AI in high-stakes environments. What does accountability look like when the algorithm moves faster than the audit trail?

If your organization is building out an AI governance function and wants a structured framework for 2026, the June GovernIntel Cohort is designed exactly for that. Details at governintel.com.