Insurance Coverage and Bad Faith

Listen to this Blog Post with Additional Commentary from the Author Henry M. Sneath

Podcast #6 in the Series: Defending the Algorithm™: A Bayesian Analysis of AI Litigation and Law

This blog post and audio podcast file is another in the series "Defending the Algorithm™" written, edited and narrated by Pittsburgh, Pennsylvania Business, IP and AI related litigation Trial Lawyer Henry M. Sneath, Esq. and was authored with research assistance by Claude® from Anthropic Opus Edition 4.5 Max and research confirmation and assistance from Google Gemini 3.0 Pro and Westlaw Precision AI. This series focuses on AI, the legal practice, and the intersections of AI and substantive law. Bayes Theorem of probability is a key component of Artificial Intelligence. AI platforms can make mistakes but the Author has taken care to check them all for accuracy.

Introduction – Executive Summary

The federal court discovery battle in Estate of Lokken v. UnitedHealth Group (UHC) (see complaint here) at Docket # 2023-cv-03514 (D. Minn. 2025) reveals what every insurer using AI and LLM’s will face: aggressive discovery demands targeting not just whether AI was used, but how it was developed, trained, validated, and deployed. While the Lokken court has postponed its ruling on plaintiffs' motion to compel, the battle lines are drawn—and they expose the tension between plaintiffs' right to test allegations and the insurers' legitimate interests in protecting proprietary systems and avoiding burdensome discovery.

For insurers, our Bayesian Analysis (see explanation here) predicts significantly increased litigation risk. Insurers using AI in claims processing face substantially higher probabilities of both bad faith litigation and contentious discovery disputes over AI systems once litigation is filed. Understanding the Lokken discovery dispute and how UnitedHealth is defending against it provides a roadmap for insurers preparing for similar battles ahead.

Key Takeaway: The discovery fight in Lokken is about more than documents. It’s about whether plaintiffs can force insurers to expose the "black box" of AI algorithmic decision-making. While courts will ultimately balance these interests, insurers should prepare now for comprehensive discovery demands that go far beyond traditional insurance claims litigation.

Chapter 1. The Lokken Case: A Brief Refresher

As we've covered in prior posts in the Defending the Algorithm™ series, Estate of Lokken v. UnitedHealth Group (UHC) is a putative class action challenging UnitedHealth's use of the nH Predict AI system in processing Medicare Advantage claims for post-acute healthcare needs. UHC purchased the corporation NaviHealth which invented nH Predict. The allegations are straightforward: UnitedHealth allegedly used the AI to predict patient recovery timelines, denied coverage when patients exceeded those predictions, and pressured medically trained employees to follow the AI’s recommendations over their own clinical judgment. Plaintiffs claim the system had an error rate of roughly 90% based on reversals on Medicare appeals.

The Core Allegations:

• UnitedHealth used nH Predict to predict patient recovery timelines;
• When patients exceeded AI-predicted timelines, coverage was denied;
• These denials occurred despite treating physicians recommending continued care;
• The AI system had an alleged 90% error rate (based on Medicare reversal rates on appeal);
• Employees were allegedly pressured to follow nH Predict recommendations over their own clinical judgment

The Legal Claims:

1. Breach of Contract - Evidence of Coverage documents promise that "Clinical Services Staff and Physicians" make claims decisions --not algorithms.
2. Breach of Implied Covenant of Good Faith and Fair Dealing - Using AI to make coverage determinations without meaningful human review violates the duty of good faith and fair dealing in contracts.

The court denied UnitedHealth's Motion to Dismiss the claims for breach of contract and breach of the implied covenant of good faith and fair dealing, finding that these claims are not preempted by the federal Medicare Act because the Court will only need to evaluate compliance with the insurance contracts themselves; in other words, plaintiffs stated viable theories that AI was used "in lieu of physicians" to make coverage determinations. The claims for Breach of Contract and for Breach of the Implied Covenant of Good Faith and Fair Dealing are moving forward and discovery has now begun. The court did dismiss other state-law claims on the basis of preemption including the claim for insurance bad faith, unjust enrichment and other tort claims. The Court reasoned that to evaluate these tort claims would require an evaluation of motive and that this would trigger pre-emption by the Medicare Act.

Chapter 2. The Discovery Battle: What the Class Plaintiffs Want

On August 21, 2025, plaintiffs filed a Motion to Compel Discovery, seeking to force UnitedHealth to produce documents and information that the insurer had refused to provide based on relevance, burden, and privilege objections. The discovery requests reveal plaintiffs' aggressive strategy: they don't just want to know whether nH Predict was used—they want to tear open the entire AI system for examination.

A. Request for Production No. 7: The Algorithm Itself

Perhaps most notably, plaintiffs served Request for Production No. 7, demanding “All Documents and Communications concerning the development, design, creation, approval, implementation, and use of the nH Predict algorithm by You, including the data, rules, source code, and medical guidelines nH Predict is based on."

This is extraordinary discovery. Plaintiffs aren't just asking for routine policy documents, claims manuals or employee training materials—they're demanding:

1. Training data - The patient records used to train nH Predict to generate predictive analytics;
2. Source code - The actual algorithm;
3. Rules - The decision logic and thresholds;
4. Medical guidelines - The clinical criteria programmed into the system.

nH Predict, is not a “generative” Large Language Model (LLM), but rather is a Predictive analytics algorithm – designed for making predictions in the healthcare domain. LLM’s generate text output based on probabilities, while nH Predict is a predictive model designed to forecast patient outcomes under certain scenarios. Think of it like a sophisticated actuarial table rather than a conversational chatbot like Chat GPT. It makes the predictions by recognizing patterns in massive amounts of data - by comparing a patient’s medical and personal characteristics with those of similar patients and by then applying the proprietary algorithm to make a prediction on outcome. 

B. The Broader Discovery Demands

Beyond RFP No. 7, plaintiffs seek a comprehensive picture of UnitedHealth's AI deployment:

RFP No. 3 - Organizational structure of employees involved in AI claims processing;

RFP No. 4 - All policies, procedures, and training materials for assessing and adjudicating post-acute claims;

RFP No. 5 - Internal documents analyzing nH Predict, including compliance with medical standards;

RFP No. 6 - Complaints and grievances about NaviHealth and nH Predict;

RFP Nos. 8-11 - Statistical data on:

• Volume of claims processed using nH Predict;
• Denial rates before and after nH Predict implementation;
• Expected vs. actual denial rates;
• NOMNCs (Notices of Medicare Non-Coverage) issued.

RFP Nos. 12-13 - Financial information:

• Acquisition of NaviHealth and projected cost savings
• Revenue and earnings from Medicare Advantage business

RFP Nos. 14-15 - Government investigations and regulatory productions;

RFP Nos. 16-17 - Employee performance evaluations and disciplinary actions;

Interrogatories seeking similar information about:

• Individuals involved in nH Predict development;
• Relationships between UnitedHealth entities;
• All versions of Evidence of Coverage documents (for putative class);Average premiums paid by Medicare Advantage insureds.

C. Bifurcated Discovery

On August 21, 2025, UnitedHealth filed a Motion to Amend the Pretrial Scheduling Order to bifurcate discovery into two phases. The proposed first stage would limit discovery to the "threshold factual question" of whether the Defendants used nH Predict instead of physician medical directors to make adverse coverage decisions for the named plaintiffs' care. Following this initial phase, UnitedHealth believes the case would be ripe for summary judgment regarding the named plaintiffs. Only if genuine issues of material fact remain would a second stage, including expert and class-based discovery, be initiated.

The court denied this motion on September 8, 2025, finding that bifurcation would not promote efficiency or judicial economy. The court reasoned that separating discovery requests would lead to additional, avoidable disputes over which category a discovery request falls in and would potentially delay the case if UnitedHealth’s summary judgment motion fails.

Chapter 3. UnitedHealth's Defense: The "Binary Question" Strategy

In its August 28, 2025 Brief in Opposition to the Motion to Compel Discovery, UnitedHealth mounted a sophisticated defense grounded in three core arguments:

A. "Either It Was Used or It Wasn't"

UnitedHealth's most compelling argument frames the dispute as a binary question: "Either nH Predict was used as Plaintiffs allege, or it was not." This is a clever legal strategy. By characterizing the case as simply asking “was it used?", UnitedHealth argues that plaintiffs don't need to know how nH Predict was developed, what data it was trained on, whether it complies with medical standards, or how accurate it is. If, as plaintiffs allege, nH Predict was used to make final coverage determinations, that alone is the breach. They claim that the "how" and "why" are irrelevant.

UnitedHealth's 'binary question' strategy is like a defendant in a car accident case arguing: “Either I hit the plaintiff's car or I didn't—you don't need to know whether I was drunk, texting, or going 90 mph. The fact of impact is all that matters.” Plaintiffs respond: “The how shows whether it was truly negligent or accidental.”

B. "We've Already Produced Enough"

UnitedHealth argues it has already provided substantial discovery in the form of:

1. Training materials and policy documents showing how medical directors apply Medicare criteria and how nH Predict is used;
2. Administrative records for the named plaintiffs showing how (or if) nH Predict was involved in their specific claims in addition to the medical records and other documents reviewed by medical directors; and
3. Repeated offers to meet and confer about additional discovery needs.

The argument is that the plaintiffs already have what they need to pursue their claims as to the named plaintiffs. Any additional requests are a fishing expedition.

C. "Faulty Premise" Objections

• Throughout its responses, UnitedHealth objected that many discovery requests were "based on the incorrect speculation that every time NaviHealth provides services that it uses nH Predict, or that nH Predict was used to make adverse coverage determinations." This creates an interesting Catch-22. Plaintiffs allege that nH Predict is used systematically and without proper oversight, but UnitedHealth counters that such claims are speculative, asserting that the AI is not always used and does not make final coverage decisions. Plaintiffs, in turn, need discovery to test their allegations. However, UnitedHealth maintains that because the allegations are incorrect, the requested discovery is irrelevant. Plaintiffs allege systematic and unregulated use of nH Predict;
• UnitedHealth says "that's speculation—nH Predict isn't always used and doesn't make final decisions";
• Plaintiffs need discovery to test their allegations;
• UnitedHealth says "your allegations are wrong, so the discovery is irrelevant".

The Court postponed a ruling on the Plaintiffs’ Motion to Compel Discovery as “Pre-Mature” because the parties were still trying to agree on Protective Order and ESI Protocol terms and due to other case factors. The parties were ordered to meet and confer and to try to reach agreement on the scope of discovery. Now that the ESI Protocol and related issues are decided and included in Court Orders, there has been no additional docket activity relating to discovery issues. We will continue to follow this case.

Chapter 4. What This Means for Insurers: A Bayesian Analysis

A. The Probability Shift

Bayesian probability analysis (See explanation here) of AI insurance litigation reveals a significant upward shift in risk for carriers and a significant increase in the cost of litigation discovery, expert testimony and overall proceedings to trial.

Litigation Risk: Insurers using AI in claims processing face substantially elevated litigation risk compared to traditional claims processing. Cases like Lokken, combined with the parallel Cigna and Humana cases, demonstrate that AI-driven claims denials are likely to become and remain a recognized cause of action.

Discovery Dispute Risk: Once AI-related litigation is filed, the probability of contentious discovery disputes is markedly higher than in traditional insurance litigation. Plaintiffs' aggressive pursuit of AI training data, source codes, and other meta data creates new battlegrounds that didn't exist in conventional bad faith cases. Discovery in these cases will be more akin to discovery in patent or trade secret litigation – highly technical and heavily dependent on competent expert witnesses and tech savvy lawyers. The parties in Lokken have entered into a Protective Order and Electronically Stored Information (ESI) Protocol (View here) that is perhaps the most sophisticated that I have ever seen. UHC will undoubtedly claim highest level confidentiality for all electronic data and may even raise trade secret protection claims to prevent discovery of its underlying predictive source codes. The risk of multiple discovery disputes in this case is very high as UHC tried hard and lost the battle to bifurcate discovery – which may have shielded them from production of the algorithmic data – but they now may well raise other objections under the Protective Order and ESI Protocol.

B. Risk Factors That Increase Litigation and Discovery Exposure

Insurers face higher discovery risk in the following situations:

1. Systemic allegations - Claims are framed as company-wide AI deployment, rather than focusing solely on individual claim disputes.
2. Breach of contract theory - Surviving contract-based claims makes "how AI was used" directly relevant to the legal claim.
3. Class certification pending - The need for class-wide data, documentation, and statistical evidence to meet the commonality and typicality requirements for class certification markedly increases the scope of discoverable material.
4. High error rates alleged - Claims asserting extreme flaws, such as a 90% overturn rate on appeal, invite and justify statistical scrutiny.
5. Government investigations – The existence of parallel legislative or regulatory investigations (e.g., the Senate report on UnitedHealth) creates additional discovery targets.

Chapter 5. Other AI Insurance Cases: The Trend Is Clear

While Lokken is the most advanced procedurally, it is far from the only case examining an insurer’s use of AI. A growing number of cases are focused on insurers’ use of predictive algorithms, particularly in making post-acute care decisions.

A. Kisting-Leung, et al. v. Cigna (E.D. Cal.)

Cigna faces similar allegations regarding its use of the PxDx algorithm to deny claims without individualized review. While specific discovery disputes have not been publicly litigated, the case raises the same questions presented in Lokken:

• Was the algorithm used to make final determinations?
• Were physicians pressured to follow algorithmic recommendations?
• What was the algorithm’s error rate?

The most significant event to date is the court’s order on March 31, 2025 which denied, in part, Cigna’s Motion to Dismiss. The court allowed the plaintiffs’ central claim, breach of fiduciary duty under ERISA, to move forward. The judge found that the plaintiffs plausibly alleged Cigna was acting as an ERISA fiduciary. The plan documents stated that a "medical director" would make determinations regarding medical necessity. The court ruled that if Cigna, as alleged, used the PxDx algorithm to auto-deny claims in 1.2 seconds, it was a plausible "abuse of discretion" and a violation of the plan's own terms.

Following that order, the case has moved into discovery. Court records confirm a Stipulated Protective Order was filed on September 4, 2025. This is to manage the discovery of Cigna's highly confidential, proprietary information and the plaintiffs' protected health information.

B. Barrows, et al. v. Humana (W.D. Ky.)

This case is a direct parallel to Lokken, as it also involves the nH Predict algorithm. Consequently, the discovery strategies from Lokken will likely be replicated in this case. The parties are actively engaged in discovery as confirmed by recent filings, including a Joint Stipulation on November 6, 2025.

On August 14, 2025, the court denied, in part, Humana's Motion to Dismiss. Humana argued that the plaintiffs must first go through the full Medicare administrative appeals process, a huge hurdle to overcome. Similar to Lokken, the court agreed to judicially waive the exhaustion requirement, finding that forcing the plaintiffs through this process would be futile. While the court did dismiss several state-law claims as being preempted by the Medicare Act, it allowed the core claims, including breach of contract and breach of the implied covenant of good faith and fair dealing, to proceed just as in Lokken.

In summary, all three of these cases have successfully survived their initial motions to dismiss. The courts have agreed that the core claims, whether for breach of contract or breach of fiduciary duty, are plausible and deserve to be argued in court. Because these cases exhibit substantial similarities, the discovery demands and defense strategies in the suits against Cigna and Humana will likely follow those in Lokken. The upcoming judicial rulings will establish the boundaries of discoverable information, setting precedent that will influence ongoing and future AI-focused insurance litigation.

Chapter 6. Looking Ahead: The Algorithmic Trade Secret Battle Is Coming

The postponement of the key discovery ruling in Lokken means the major questions surrounding algorithmic transparency remain unanswered: Can insurers effectively protect their AI algorithms as trade secrets or otherwise as confidential information as a defense to discovery demands? The ultimate judicial answer will hinge on several key legal and technical determinants. The answer to this question will depend on three factors:

1. How unique is the algorithm? Protection is highly dependent on the proprietary nature of the technology. For instance, off-the-shelf predictive models are unlikely to qualify as trade secrets. In contrast, custom algorithms built with proprietary training data have a much stronger claim to protection.
2. What is really at issue? Courts will attempt to balance transparency with commercial secrecy. They may require the disclosure of the AI's decision logic and thresholds necessary for plaintiffs to prove their claims, but they may still protect the underlying source code if the decision logic suffices.
3. Are there less intrusive alternatives? The necessity of discovery often relies on the lack of other options. Courts may consider if plaintiffs' experts can adequately test the system without seeing the source code or if access to input/output data and statistical summaries alone would suffice to assess the AI's alleged error rate and bias.

Insurers face a fundamental tension: AI systems work best as black boxes (protecting the proprietary methods), but contract obligations require transparent decision-making. Plaintiffs will argue that “it's like hiring a consultant who refuses to explain their advice but insists you follow it—eventually, trust breaks down."

Chapter 7. Conclusion: The Discovery Phase Will Shape AI Insurance Litigation

The Lokken discovery battle appears to be the first major test of whether insurers can shield AI systems from scrutiny in litigation. The outcome will determine:

1. Scope of AI discovery - Will courts require full algorithmic transparency or accept limited disclosure?
2. Trade secret protection - Can insurers successfully invoke trade secret protection for AI systems used in claims processing?
3. Burden and proportionality - How will courts balance plaintiffs' need for information against the burden of massive AI-related document productions?

For insurers, the lesson is clear: the discovery phase is now more critical than ever as to whether AI insurance cases will be won or lost. The substantive legal issues are important. But if plaintiffs can't get discovery on how the algorithms actually work, they will have a hard time trying to prove their case. Plaintiffs will want to examine the credibility of human testimony by looking under the hood. Looking to see how the algorithm was trained and on what material? What is it prompted by its creator coders to do in the black algorithmic box? UnitedHealth’s defense strategy in Lokken, to frame it as a binary question; limit discovery to named plaintiffs only; and protect the algorithm as proprietary, is the template other insurers will follow. Whether it succeeds will shape the future of AI insurance litigation.

To conclude in Bayesian probability terms given the likelihood factors we are learning from the Lokken case: P(discovery in breach of contract and insurance bad faith cases becoming more meta data driven, more complicated and more expensive) | (the need by Plaintiffs to challenge human testimony on what the subject algorithms are doing in the black box) = Very High".

My Thanks to Claude®, Google Gemini and Westlaw Precision for their assistance with the Defending the Algorithm series. Thanks also to Houston Harbaugh Associate Attorney Taylor Hinds for her research and writing for this blog. Houston Harbaugh's intellectual property and AI litigation team continues to monitor legal developments in LLM and predictive software use and implementation, and in AI enterprise litigation across all legal frameworks. For questions about how these evolving legal standards may affect your AI deployment, data acquisition practices, or enterprise AI risk management, please contact the Author Henry M. Sneath, Esq. in our office at email address sneathhm@hh-law.com or 412-288-4013. Thanks for reading or listening. See you next time.

This blog post represents the author's views and does not constitute legal advice. Readers should consult with qualified legal counsel regarding specific legal issues.