id: PIR-2026-0005title: Class action alleges UnitedHealth's nH Predict model cut off Medicare Advantage post-acute care against physician judgmentdate_occurred: 2022-2023 (named plaintiffs' coverage cutoffs; practice alleged to be ongoing)date_detected: unknown (families experienced cutoffs directly; litigation followed investigative reporting)date_disclosed: 2023-11-14 (complaint filed, D. Minn., 0:23-cv-03514); motion to dismiss partially denied 2025-02-13status: disputed (core mechanism is plaintiff allegation; UnitedHealth contests; no judgment)agent_description: nH Predict (NaviHealth/UnitedHealth), a predictive model applied to Medicare Advantage post-acute care claims, generating length-of-stay predictions alleged to function as de facto coverage-termination decisions.operator_type: enterpriseautonomy_level: human-on-the-loop nominally (human sign-off formally required); alleged de facto autonomous-within-policy (predictions alleged to override treating physicians)model_stack: unknown (proprietary predictive model; internals subject to ongoing discovery)harness: NaviHealth utilization-management pipelinefailure_locus: operator-config (alleged: rigid targets imposed on model output; the model did what it was deployed to do)authority_scope: decision authority over insurance coverage (effective control of care funding for elderly patients); data access (claims and clinical records)funds_at_risk_usd: unknown at class scale; named plaintiffs allege tens of thousands of dollars each in out-of-pocket costsblast_radius: customers/third parties (Medicare Advantage members across the plan)root_cause: policy-violation - primary: alleged breach of contract and of good-faith coverage obligations. Contributing: operator-error (deployment of predictions as rigid cutoffs)mechanism: Per the complaint: nH Predict generated length-of-stay targets, and coverage was cut when the prediction was reached regardless of treating physicians' assessments. Plaintiffs allege that over 90 percent of the denials were reversed through internal appeal or ALJ proceedings (a complaint allegation, not an independently measured model error rate), that UnitedHealth knew this, and that it relied on most patients not appealing. Named plaintiffs' families paid out of pocket; two named patients died during or after the coverage cutoffs. The court (Judge Tunheim, 2025-02-13) dismissed several Medicare-preempted counts with prejudice but allowed breach-of-contract and implied-covenant claims to proceed, with broad discovery following; the ruling establishes that those claims proceed, not that the model cut care unlawfully or caused the alleged deaths.adversary_present: noexploitation_status: in-wild-malfunction (real coverage denials in production; no attacker - reclassified from in-wild-exploited, which implies an adversary)severity: loss (named plaintiffs' realized out-of-pocket costs; alleged patient harm up to death) - class-scale loss unknown and unadjudicateddirect_loss_usd: unknown (no judgment; named-plaintiff allegations in the tens of thousands each)indirect_loss_usd: unknowndowntime: n/adata_exposure: nonedetected_by: third-party (affected families and counsel; investigative reporting preceding the suit)time_to_detect: unknowntime_to_recover: ongoing (litigation active as of Aug 2026)remediation: none by operator to date on the public recordstructural_fix: none disclosedcontrols_that_worked: the appeals process - if the alleged >90% reversal rate holds, the human appellate layer caught errors, but only for patients who invoked itliability_holder: contested (UnitedHealth Group/NaviHealth as operator; no adjudication)precedent_set: 2025-02-13 ruling lets contract and good-faith claims over AI-driven coverage denial proceed past dismissal, with broad discovery into the model's usesealed_material: unknown (discovery ongoing)telemetry_grade: operator-logs (claims-system records in the operator's custody, now subject to discovery)sources:independence: court record + academic tracker + independent legal press - strong on procedure.aiid_incident_id: 608 (https://incidentdatabase.ai/cite/608/) - cross-reference; primaries verified independentlyconfidence: medium - procedural facts high (documented rulings); the 90% reversal rate and rigid-cutoff mechanics are plaintiff allegations not yet proven.Filing date, case number, and the 2025-02-13 partial denial verified against the docket and coverage; allegations kept clearly labeled as such throughout. Boundary case (alleged de facto autonomy) recorded under the v0.1 legal block.
aiid_incident_id cross-reference (AIID 608), matched against the AIID weekly database export (2026-08-17). A cross-reference, not a re-verification; no claim changed.See also - this event in the AI Incident Database: incident 608.