What was the nH Predict lawsuit about?

Federal lawsuits alleged that a major health insurer used an AI tool called nH Predict to estimate post-acute care length of stay, then denied coverage that departed from the algorithm's prediction — allegedly overriding treating physicians' judgment at a high rate and without individualized review. The cases argue this violates state insurance bad-faith law and, in some filings, disability discrimination statutes because the tool disproportionately cut off care for elderly and disabled patients.

Does California law now require human review of AI claims denials?

California SB 1120 requires health plans to ensure that medical necessity decisions are made by a licensed physician or appropriately licensed provider, not by an algorithm alone, and that AI or algorithmic tools used in utilization review are based on the individual patient's medical history rather than a generalized dataset. Denials cannot rest solely on an algorithmic output.

Can an AI-driven claims denial trigger disability discrimination liability?

Yes, if the tool systematically produces worse outcomes for patients with certain diagnoses, ages, or disabilities, plaintiffs can argue disparate-impact discrimination under state insurance codes or disability law, separate from any bad-faith claims-handling theory. A pattern of algorithmic denials clustered around specific conditions is the kind of statistical evidence used to support these claims.

Does this liability apply to property and casualty insurers too, or only health plans?

The legal theories extend beyond health coverage. Property and casualty insurers using AI to evaluate claims, estimate repair costs, or flag claims for denial or investigation face similar bad-faith and discrimination exposure if the tool's outputs aren't reviewed by a qualified adjuster and can be shown to disadvantage protected groups or specific geographies in a way that overlaps with protected characteristics.

AI Insurance Claims Denial Lawsuits 2026: Discrimination Risk for Payers

From Underwriting Bias to Claims-Denial Bias

Most AI-insurance compliance coverage focuses on underwriting: can an algorithm price a policy or approve an applicant in a way that discriminates. That's real exposure, but a second front has opened that's arguably more litigated right now — AI tools used after a policy is issued, to decide whether a specific claim gets paid.

The fact pattern behind the highest-profile suits is consistent: an insurer deploys a predictive model to estimate an outcome — length of stay, likely recovery time, expected repair cost — and then denials track the model's output far more tightly than they track individualized clinical or adjuster judgment. Plaintiffs argue the AI wasn't a decision support tool; it was the decision-maker, with a human simply rubber-stamping the output.

That distinction — decision support versus de facto decision-maker — is now the central legal question in this entire category of litigation.

The nH Predict Fact Pattern

The lawsuits that put this issue on the map involved a post-acute care prediction tool used to estimate how long a patient should need skilled nursing or rehab care after a hospital stay. Plaintiffs alleged coverage was cut off in lockstep with the algorithm's predicted discharge date at a rate far exceeding what treating physicians recommended, and that the tool's error rate was known internally but not corrected.

Denials allegedly followed the model's output regardless of individual patient factors like comorbidities or slower recovery
The claims relied on state bad-faith insurance law — failure to conduct a reasonable, individualized investigation before denying a claim
Some filings layered on disability-discrimination theory, arguing elderly and disabled patients bore a disproportionate share of the cutoffs
The reputational and settlement fallout accelerated state legislation requiring human review, discussed below

California SB 1120 and the Human-Review Mandate

California responded directly to this litigation pattern with SB 1120, which requires health plans and insurers to ensure that decisions to deny, delay, or modify coverage based on medical necessity are made by a licensed physician or appropriately licensed health professional — not by an algorithm alone.

Individualized Basis Required

Any AI or algorithmic tool used in utilization review must base its output on the individual patient's medical history and clinical circumstances, not solely on a generalized dataset or population-level statistics.

No Algorithm-Only Denials

A denial, delay, or modification of care cannot rest on an algorithmic recommendation without a licensed reviewer independently evaluating the individual case before the adverse decision is finalized.

Fair, Consistent, Bias-Tested Application

Plans must apply AI tools in a way that doesn't discriminate, and must be able to show the tool was evaluated for bias against the populations it's applied to.

Other states are watching California's approach closely, and multistate payers should expect similar human-review requirements to spread rather than treat this as a California-only compliance item.

Compliance Checklist for Insurers and AI Vendors

1. Document Human Judgment, Not Just Human Sign-Off

☐Require the reviewing physician or adjuster to record the individualized factors they considered, not just an approval checkbox on the AI output
☐Track how often the human reviewer overrides the algorithm in both directions — a near-zero override rate is itself evidence the review is not meaningful
☐Prohibit denial letters that cite the algorithm's output as the primary or sole stated reason

2. Bias-Test the Model Against Outcomes, Not Just Inputs

☐Run periodic disparate-impact analysis on denial rates by age, diagnosis category, and other protected or proxy characteristics
☐Document remediation steps when a bias pattern is found, not just detection
☐Keep model validation records for the retention period required in your state's insurance code

3. Vendor Contract Terms

☐Get contractual representations from AI vendors about training data, known error rates, and bias-testing results
☐Require vendors to disclose material model updates that could shift denial patterns
☐Allocate liability for algorithmic errors explicitly rather than relying on general indemnification language

4. Claimant-Facing Transparency

☐Disclose in plain language when AI was used in evaluating a claim, consistent with emerging state AI-transparency requirements
☐Provide a clear appeals path that guarantees human review distinct from the original algorithmic assessment
☐Preserve the case file, including the algorithm's output and the reviewer's documented reasoning, for the appeals window and beyond

Frequently Asked Questions

We use AI only to flag claims for review, not to deny them directly. Are we exposed?

You have less exposure than an insurer whose denials track the algorithm automatically, but you're not exempt. If your reviewers approve the AI's flag at a very high rate without documenting independent judgment, plaintiffs can argue the human step is nominal rather than substantive — the same theory used against automated denial systems.

Does this only apply to health insurance, or also auto and homeowners claims?

The highest-profile litigation and California's SB 1120 focus on health coverage, but the underlying bad-faith and discrimination legal theories aren't health-specific. Property and casualty insurers using AI to estimate damage, flag claims as suspicious, or recommend denial face comparable exposure under general insurance bad-faith law.

What should we do if an internal audit shows our AI tool has a disparate denial pattern?

Document the finding, the remediation plan, and the timeline for implementing it. Regulators and plaintiffs' counsel treat a documented, acted-upon finding very differently from an undisclosed one discovered later in litigation — remediation in progress is a materially better position than silence.

Is a general AI-use disclosure in our policy documents enough to cover claims-handling AI?

No. Emerging laws like SB 1120 require specific safeguards around medical-necessity decisions — individualized basis, licensed human review, non-discriminatory application — that go well beyond a generic AI-use disclosure. Review your claims AI deployment against the specific requirements, not just general transparency language.

The Override Rate Is the Tell

Regulators and plaintiffs' counsel are converging on the same diagnostic: how often does a human reviewer actually override the algorithm's recommendation. If the answer is "almost never," your review process will be characterized as cosmetic, regardless of what your policy documents say.

Start by pulling that number for your own claims process. If it's near zero, that's the first thing to fix — before a plaintiff's expert pulls the same statistic in discovery.