{"id":905995,"date":"2026-08-12T15:21:21","date_gmt":"2026-08-12T21:21:21","guid":{"rendered":"https:\/\/www.myconstructionexpert.com\/blog\/?p=905995"},"modified":"2026-08-12T15:21:24","modified_gmt":"2026-08-12T21:21:24","slug":"artificial-intelligence-in-claim-handling","status":"publish","type":"post","link":"https:\/\/www.myconstructionexpert.com\/blog\/artificial-intelligence-in-claim-handling\/","title":{"rendered":"From Colossus to ChatGPT: Artificial Intelligence in Modern Claim Handling \u2014 Efficiency, Explainability, and Exposure"},"content":{"rendered":"\n

Christopher Jacobs | Houston Harbaugh<\/a><\/p>\n\n\n\n

I. Introduction and Historical Context<\/strong><\/p>\n\n\n\n

The integration of artificial intelligence<\/a> into the insurance industry has not occurred in a single leap but through a series of incremental innovations\u2014each testing the boundaries between efficiency and fairness, automation and accountability. While today\u2019s conversation focuses on large language models, image-recognition tools, and predictive analytics, the roots of algorithmic claim handling trace back nearly three decades to the introduction of\u00a0Colossus<\/em>, a software program once hailed as revolutionary in standardizing bodily-injury claim valuations. The trajectory from\u00a0Colossus<\/em>\u00a0to contemporary AI systems offers not merely a technological evolution, but a jurisprudential one: both eras have forced courts, regulators, and practitioners to confront how far insurers may rely on opaque systems to evaluate inherently human losses.<\/p>\n\n\n\n

A. The Promise of Automation in Claims<\/strong><\/p>\n\n\n\n

At its inception in the 1990s, Colossus<\/em> was marketed as a neutral, data-driven means of ensuring consistency across bodily-injury settlements. Adjusters were instructed to input medical details, treatment types, and jurisdictional information, and the program would generate a recommended settlement range based on prior claims data. Insurers adopting the system saw it as an antidote to subjectivity\u2014an algorithmic equalizer that could reduce variance, curb inflated demands, and streamline claim resolution. In a business driven by volume and loss ratios, such efficiency was irresistible.<\/p>\n\n\n\n

Yet, the same characteristics that made Colossus<\/em> efficient also made it inscrutable. The proprietary \u201cvalue drivers\u201d that produced its outputs\u2014reportedly numbering in the thousands\u2014were hidden behind trade-secret protections. Adjusters could see the result, but not the reasoning. Regulators and claim professionals later observed that a system intended to remove subjectivity in valuation often introduced new, systemic biases of its own.[1]<\/a><\/p>\n\n\n\n

B. The Colossus Litigation and Judicial Response<\/strong><\/p>\n\n\n\n

By the early 2000s, Colossus<\/em> had become the focal point of consumer litigation and regulatory scrutiny. Plaintiffs alleged that insurers and the software\u2019s developer, Computer Sciences Corporation (CSC), had used Colossus<\/em> to systematically undervalue bodily-injury claims by enforcing uniform settlement ranges and incentivizing adjusters to conform to them. The central contention was not that Colossus<\/em> malfunctioned in a technical sense, but that its design embodied a \u201cone-size-fits-all\u201d approach to inherently individualized losses. By converting subjective, claimant-specific evaluations into standardized algorithmic outputs, the software replaced human discretion with what might be called artificial standardization<\/em>. That process risked disregarding unique claimant characteristics\u2014such as occupation, pre-existing conditions, or quality-of-life impacts\u2014that traditionally informed the valuation of bodily-injury damages.<\/p>\n\n\n\n

The most notable of these actions culminated in a national class-action settlement in Hensley v. Computer Sciences Corporation<\/em>, filed in Arkansas and approved in 2005. The plaintiffs contended that CSC and multiple insurers had concealed the software\u2019s use and manipulated its calibration to achieve predetermined reductions in claims payouts. While CSC denied wrongdoing, the settlement required it to modify its marketing practices, clarify the system\u2019s intended use, and make aspects of its functionality more transparent to insurers and regulators. Subsequent market-conduct examinations in several states, including California and Michigan, reached similar conclusions: insurers could not require adjusters to adhere rigidly to Colossus<\/em> outputs, nor could they compensate personnel based on compliance with those valuations.<\/p>\n\n\n\n

Although the resulting settlements and administrative orders did not create binding precedent, they altered the industry\u2019s expectations for algorithmic decision-making. The Colossus<\/em> controversy demonstrated that automation does not absolve insurers of the duty of fair claim handling; it merely reframes it. The software\u2019s opacity generated the same evidentiary and ethical challenges now resurfacing with modern AI: explainability, bias, and the tension between internal models and external accountability.<\/p>\n\n\n\n

C. Lessons from the First Generation of Algorithmic Claims<\/strong><\/p>\n\n\n\n

Three lessons from the Colossus<\/em> era remain particularly instructive for today\u2019s practitioners.<\/p>\n\n\n\n

First, transparency matters<\/strong>. Regulators and courts grew skeptical not merely because Colossus<\/em> existed, but because it operated as a hidden arbiter of value. The lack of disclosure\u2014both to claimants and, in some instances, to line adjusters\u2014transformed a management tool into a litigation risk.<\/p>\n\n\n\n

Second, automation amplifies institutional intent<\/strong>. A valuation system trained or tuned to achieve efficiency gains can easily be repurposed to accomplish cost-containment objectives inconsistent with fair-claims standards. Plaintiffs in Hensley<\/em> alleged that Colossus<\/em> was calibrated to produce systematic reductions in claim payouts, reportedly targeting decreases of up to fifteen percent. Whether or not that allegation could be empirically proven, the perception alone was enough to erode confidence in algorithmic fairness\u2014a cautionary lesson for modern machine-learning tools trained on historical data that may embed similar undervaluation biases.<\/p>\n\n\n\n

Third, human oversight is indispensable<\/strong>. Following the Colossus<\/em> settlements, several state insurance departments emphasized that adjusters must retain independent judgment and document reasons for either deviating from or adopting software recommendations. That principle\u2014human-in-the-loop accountability\u2014has since become the ethical baseline for any deployment of AI in claims handling.<\/p>\n\n\n\n

D. Continuity Between Colossus and Contemporary AI<\/strong><\/p>\n\n\n\n

Seen through this historical lens, today\u2019s debates over generative AI and machine-learning systems are less revolutionary than cyclical. Modern claim-analysis models, particularly those that evaluate images of property damage or generate text-based coverage explanations, echo the same dynamics of efficiency versus discretion that defined Colossus<\/em>. The principal difference lies in scale and autonomy. Where Colossus<\/em> applied deterministic rules to structured data, modern AI operates on probabilistic inference from vast, unstructured sources\u2014millions of data points, often processed without direct human review. The opacity has deepened, and with it, the potential for both error and abuse.<\/p>\n\n\n\n

In litigation, this evolution raises familiar questions under new guises. If an insurer relies on an AI-generated damage estimate that proves inaccurate, has it acted unreasonably? If an AI model produces inconsistent results across regions or demographics, does that constitute unfair discrimination under state law? These inquiries trace their lineage directly to the Colossus<\/em> disputes, but they now extend into far more complex evidentiary terrain. Discovery once aimed at Colossus<\/em> calibration settings will soon target training data, neural-network architectures, and algorithmic weighting\u2014issues few courts are yet equipped to handle.<\/p>\n\n\n\n

E. Framing the Discussion Ahead<\/strong><\/p>\n\n\n\n

The following sections will build upon these historical parallels. Section II will examine how insurers currently deploy AI in claim handling, including both legitimate efficiency tools and the emerging risk of AI-generated fraudulent claims submitted by insureds themselves. Section III will turn to the developing legal landscape, analyzing how courts have begun to interpret bad-faith standards in the context of algorithmic decision-making and what ethical obligations arise from reliance on non-transparent systems. Ultimately, the goal is not to condemn automation, but to situate it within a continuum of evolving duties\u2014duties that require insurers to harness innovation responsibly, preserve human judgment, and maintain the public trust that underpins the entire enterprise of insurance.<\/p>\n\n\n\n

II. Modern AI in Claim Handling<\/strong><\/p>\n\n\n\n

A. From Colossus<\/em> to Cognitive Claims \u2014 The Evolution of Automation<\/strong><\/p>\n\n\n\n

The impulse that gave rise to Colossus<\/em>\u2014the desire for faster, more consistent, and ostensibly \u201cobjective\u201d claim evaluations\u2014remains the same animating force behind today\u2019s adoption of artificial intelligence in insurance claim handling. What has changed is the data universe<\/em>. Where Colossus<\/em> relied on structured medical codes, treatment descriptions, and adjuster input, modern AI systems ingest unstructured and multimodal data: photos of property damage, satellite imagery, adjuster notes, invoices, correspondence, and even social media content.<\/p>\n\n\n\n

In essence, the industry has moved from rule-based automation<\/strong> to learning-based cognition<\/strong>. Yet the governing question is still familiar: how much of the inherently human process of claims evaluation can be delegated to machines without undermining the insurer\u2019s duty of good faith?<\/p>\n\n\n\n

Many of today\u2019s AI tools are, in a sense, Colossus\u2019s descendants<\/em>. They pursue the same efficiencies\u2014standardization, cost control, and cycle-time reduction\u2014but do so with greater technical sophistication and, correspondingly, greater opacity. The tension between innovation and accountability, first articulated in the Colossus<\/em> litigation, now resurfaces in the age of generative and predictive AI.<\/p>\n\n\n\n

B. The New Landscape of AI in Claims<\/strong><\/p>\n\n\n\n

Modern claim-handling systems employ a range of artificial intelligence applications, each addressing a different aspect of the loss evaluation process.<\/p>\n\n\n\n

1. Property-Damage Assessment<\/strong><\/p>\n\n\n\n

In property and auto claims, some image-recognition platforms use computer-vision models to analyze photographs of damaged structures or vehicles and generate cost estimates. These systems extract patterns\u2014material types, breakage contours, water staining, roof damage\u2014from uploaded imagery, often producing an initial estimate before an adjuster ever visits the site.<\/p>\n\n\n\n

Other platforms extend this approach to full virtual inspections by using spatial-recognition algorithms to determine room geometry, materials, and fixtures directly from smartphone photos. These systems effectively operationalize the Colossus ideal of standardized evaluation in the property context, replacing onsite measurements with digital modeling. While they can dramatically reduce inspection times, they also risk replicating Colossus\u2019s chief shortcoming: the substitution of algorithmic uniformity for contextual judgment. Lighting conditions, debris, or atypical construction materials can skew an AI\u2019s interpretation, and the line between efficiency and oversimplification again becomes blurred.<\/p>\n\n\n\n

2. Bodily-Injury and Fraud Detection Models<\/strong><\/p>\n\n\n\n

For bodily-injury and personal-lines claims, insurers may rely on predictive and anomaly-detection models reminiscent of Colossus\u2019s original domain but enhanced with modern machine-learning architecture. Some commercially available AI platforms evaluate claim narratives, medical billing, and historical claim data to detect inconsistencies or potential fraud. Certain systems combine structured data with natural-language processing of adjuster notes, producing risk scores that triage claims for additional review.<\/p>\n\n\n\n

Complementing these are advanced text- and pattern-recognition systems that mine claim narratives and historical records for linguistic or statistical anomalies suggestive of misrepresentation or exaggeration. Like their Colossus predecessor, these systems quantify human judgment\u2014here through pattern extraction rather than explicit rule encoding.<\/p>\n\n\n\n

As with Colossus<\/em>, these tools promise consistency and efficiency\u2014but they also pose the same ethical dilemma: if a claim is flagged as \u201csuspicious\u201d by an opaque model, what is the insurer\u2019s evidentiary basis for that conclusion? The program\u2019s internal weighting of factors\u2014location, treatment type, prior claims\u2014may be invisible to the adjuster who relies on it. In this way, the Colossus<\/em> debate over transparency re-emerges in algorithmic form.<\/p>\n\n\n\n

3. Workflow Automation and Communication Tools<\/strong><\/p>\n\n\n\n

Beyond valuation, AI now automates parts of the claim lifecycle once deemed purely administrative. Some carriers use machine-learning triage systems to route claims to appropriate handling paths, while others experiment with generative text tools that draft coverage letters or status updates. These automations can improve customer experience and free adjusters for complex tasks, but they also introduce potential legal risk. A misworded AI-generated communication\u2014particularly one that inaccurately represents coverage\u2014may be discoverable evidence in a later bad faith action. Once again, technology that promises efficiency can quietly erode the precision and care required in insurer communications.<\/p>\n\n\n\n

4. The Parallels Between Colossus<\/em> and Modern AI Tools<\/strong><\/p>\n\n\n\n

Across all these tools, the parallels with Colossus<\/em> are striking. Each new technology advances the same triad of objectives\u2014speed, uniformity, and cost control\u2014while exposing insurers to the same triad of risks\u2014opacity, bias, and overreliance. The lesson of the past two decades remains constant: automation can assist the adjuster, but it cannot become<\/em> the adjuster.<\/p>\n\n\n\n

To synthesize the current marketplace of tools and their potential implications, the following summarizes representative examples of AI applications in claim handling as discussed above:<\/p>\n\n\n\n

Use Case<\/strong><\/strong><\/td>Purpose\/Function<\/strong><\/strong><\/td>Representative Applications<\/strong><\/td>Observations\/Limitations<\/strong><\/strong><\/td><\/tr><\/thead>
Property-damage assessment<\/strong><\/td>Uses computer vision to analyze photos or video and estimate repair costs.<\/td>Computer-vision and virtual-inspection platforms<\/td>Improves efficiency but may misread atypical damage; requires human validation.<\/td><\/tr>
Bodily-injury evaluation & fraud detection<\/strong><\/td>Applies predictive analytics to claim data and medical records to flag anomalies or suspicious patterns.<\/td>Predictive analytics, anomaly-detection, and text-mining platforms<\/td>Enhances consistency; opacity of models raises fairness and \u201cexplainability\u201d concerns.<\/td><\/tr>
Claims triage and routing<\/strong><\/td>Automatically categorizes and prioritizes claims for appropriate handling.<\/td>Proprietary or vendor-developed triage systems<\/td>Risk of misclassification; requires audit trails and override authority.<\/td><\/tr>
Automated communications<\/strong><\/td>Drafts coverage letters and claim-status messages using rule-based or generative AI.<\/td>Internal carrier systems; vendor plug-ins<\/td>Must be reviewed for legal accuracy; errors may create bad-faith exposure.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

C. Fraud in the Age of AI<\/strong><\/p>\n\n\n\n

If Colossus<\/em> represented the insurer\u2019s early flirtation with algorithmic overreach, generative AI marks the insured\u2019s opportunity for technological abuse. Where automation once threatened undervaluation, it now enables fabrication\u2014the deliberate creation of false evidence through AI. The same tools that allow insurers to analyze photographs can also allow claimants to fabricate them.<\/p>\n\n\n\n

The most common manifestation is the AI-generated image of property damage. A claimant might superimpose fire, water, or structural damage onto authentic photos or generate entirely synthetic images of a purported loss. Some falsify metadata to make the photos appear contemporaneous with the claimed event. Others produce fabricated invoices or repair estimates generated by text-based AI systems.<\/p>\n\n\n\n

Large international insurers have reported detecting claims supported by manipulated imagery, often identified through telltale inconsistencies in lighting, shadows, or metadata.[2]<\/a> Specialized forensic-AI systems now scan submitted photos for generative artifacts\u2014digital \u201cfingerprints\u201d left by diffusion or generative models.[3]<\/a><\/p>\n\n\n\n

The asymmetry of technology has created a new cat-and-mouse dynamic: claimants armed with generative tools versus insurers developing discriminative ones. Yet the underlying legal principles are unchanged. The insurer retains both the right and the duty to investigate suspected fraud, and must do so with diligence and documentation sufficient to support its decision if challenged in litigation.<\/p>\n\n\n\n

Ultimately, this emerging threat highlights the mirror image of the Colossus<\/em> problem. Where the first generation of algorithms risked dehumanizing claims through artificial standardization, the new generation threatens to deceive them through artificial fabrication. Both scenarios reveal that the heart of claim handling\u2014truthful, individualized assessment\u2014remains a human responsibility, regardless of technological sophistication.<\/p>\n\n\n\n

D. Balancing Efficiency and Oversight<\/strong><\/p>\n\n\n\n

The path forward lies not in rejecting AI, but in integrating it responsibly. Insurers can preserve the benefits of automation while mitigating its risks through several interlocking practices.<\/p>\n\n\n\n

First, human-in-the-loop review<\/strong> must remain nonnegotiable. No AI output, however sophisticated, should serve as the final basis for a coverage or valuation determination without human confirmation. Adjusters should understand the inputs and confidence thresholds that underpin AI recommendations. Deviations\u2014i.e., whether to accept or reject an AI\u2019s conclusion\u2014should be documented.<\/p>\n\n\n\n

Second, model governance and auditability<\/strong> are essential. Insurers should maintain records of model versions, training data, and updates, as well as track overrides or exceptions to algorithmic recommendations. Such documentation not only strengthens compliance but also provides a defensible record if a dispute arises.<\/p>\n\n\n\n

Third, vendor transparency<\/strong> should be contractually required. Insurers adopting third-party AI tools should insist on access to model documentation and, where possible, audit rights. Without such safeguards, an insurer may find itself unable to explain a decision in discovery\u2014an evidentiary vulnerability reminiscent of the Colossus<\/em> era.<\/p>\n\n\n\n

Finally, training and culture<\/strong> matter. Claim professionals should be educated not only on the functionality of AI tools but on their limitations. The same duty of fairness that applied in the manual-claims era persists in the digital one. Automation should accelerate human reasoning, not replace it.<\/p>\n\n\n\n

The above discussion reveals the dual nature of AI in claim handling: it is both an instrument of efficiency and a potential vector of error or alleged deceit. Like Colossus<\/em> before it, each innovation carries with it a renewed obligation\u2014to preserve transparency, maintain discretion, and ensure that the pursuit of technological progress never eclipses the foundational principle of fair and individualized claim evaluation.<\/p>\n\n\n\n

III. Legal Landscape<\/strong><\/p>\n\n\n\n

If Colossus<\/em> marked the industry\u2019s first encounter with algorithmic accountability, modern artificial intelligence represents its jurisprudential sequel. Courts have yet to articulate a comprehensive framework for AI-assisted claim handling, but familiar doctrines\u2014bad faith, evidentiary reliability, and fairness\u2014provide a starting point. The question is not whether existing law applies, but how it adapts.<\/p>\n\n\n\n

A. Bad Faith and the Standard of Reasonableness<\/strong><\/p>\n\n\n\n

At the core of every bad-faith inquiry lies a simple question: Was the insurer\u2019s conduct reasonable under the circumstances?<\/em> Artificial intelligence complicates that inquiry by adding a layer of mechanical judgment between the insurer and its insured. When an AI system influences claim valuation or denial, the reasonableness of the insurer\u2019s conduct becomes inseparable from the reasonableness of its reliance on that system.<\/p>\n\n\n\n

Courts assessing Colossus<\/em>-related claims applied this logic implicitly. They did not condemn automation per se<\/em> but scrutinized whether insurers used it responsibly\u2014whether adjusters exercised independent judgment rather than deferring to the software\u2019s recommendation. The same reasoning will likely govern modern AI cases in the extracontractual context. If a carrier denies, limits, or delays payment based on an algorithmic recommendation it cannot explain or verify, policyholders may argue that the insurer abdicated its non-delegable duty of good faith.<\/p>\n\n\n\n

In Anderson v. Nationwide Mutual Insurance Co.<\/em>, for example, the court held that \u201cfailure to evaluate the insured\u2019s claim based on individualized evidence, rather than rigid internal metrics, may support a finding of bad faith.\u201d[4]<\/a> While Anderson<\/em> predated the current AI wave, its principle maps directly onto contemporary claim handling: an insurer that substitutes model output for individualized analysis risks crossing the line from efficiency to indifference.<\/p>\n\n\n\n

Modern case law in adjacent contexts reinforces this approach. In State Farm Fire & Casualty Co. v. Slade<\/em>, the Alabama Supreme Court emphasized that \u201can insurer\u2019s claim-handling practices must remain guided by professional judgment, even when informed by standardized procedures.\u201d[5]<\/a> And in Rancosky v. Washington National Insurance Co.<\/em>, the Pennsylvania Supreme Court reaffirmed that bad faith encompasses reckless disregard for the absence of a reasonable basis to deny benefits.[6]<\/a> These holdings collectively suggest that automation offers no safe harbor: an unreasonable belief in the accuracy of an opaque algorithm may be analogous to an unreasonable belief in a flawed human assessment.<\/p>\n\n\n\n

B. Potential Judicial Treatment of Algorithmic Evidence<\/strong><\/p>\n\n\n\n

Although no U.S. court has yet ruled squarely on the admissibility or reliability of AI claim-handling tools, early litigation hints at the contours of the debate. In discovery disputes, plaintiffs are likely to begin to seek model documentation, training data, and internal correspondence regarding algorithmic decision-making. Courts may soon confront questions analogous to those raised in Colossus<\/em> litigation, principally: Must an insurer disclose the internal logic of an AI model if it forms part of the claim-decision process?<\/em><\/p>\n\n\n\n

A few cases already gesture toward answers, or at least provide a snapshot of potential judicial treatment of the issue. In United States v. Loomis<\/em>, a criminal-sentencing case involving a proprietary risk-assessment algorithm, the Wisconsin Supreme Court upheld use of the algorithm but warned that reliance on an undisclosed model required \u201ccaution and transparency.\u201d[7]<\/a> Although Loomis<\/em> was not an insurance case, its reasoning\u2014emphasizing due process and \u201cexplainability\u201d\u2014may influence civil discovery disputes involving insurer AI systems. Similarly, in In re State Farm Lloyds<\/em> (Tex. Sup. Ct. 2020), the court addressed electronic claim-handling databases, holding that insurers could not withhold data essential to understanding how loss values were calculated.[8]<\/a> Both decisions underscore the judiciary\u2019s growing expectation that technological processes affecting substantive rights be explainable<\/em>.<\/p>\n\n\n\n

In practice, the evidentiary challenge for insurers will be twofold: first, to preserve and produce sufficient documentation to show that an AI recommendation was reasonable; and second, to train personnel capable of articulating that reasoning in deposition or trial testimony. A company that cannot explain its own system invites the same skepticism that plagued Colossus<\/em> users two decades ago.<\/p>\n\n\n\n

C. Looking Forward<\/strong><\/p>\n\n\n\n

Taken together, these developing doctrines suggest a clear trajectory. Courts are unlikely to craft new \u201cAI law\u201d for insurers; instead, they will likely apply traditional standards\u2014reasonableness, good faith, and \u201cexplainability\u201d\u2014to new facts. Just as Colossus<\/em> taught that automation does not immunize a carrier from scrutiny, modern AI will teach that sophistication does not substitute for accountability. The insurer\u2019s safest course remains the oldest one: exercise judgment, document reasoning, and treat every claim as an individual undertaking rather than a data point. In the end, fairness and transparency\u2014the same qualities arguably absent from the Colossus<\/em> controversy\u2014remain the best defenses against the next generation of litigation.<\/p>\n\n\n\n

IV. Discovery and Evidentiary Issues<\/strong><\/p>\n\n\n\n

As insurers begin to integrate artificial intelligence into claim handling, disputes will inevitably test how courts treat AI-influenced decisions\u2014both in contractual coverage cases and in extracontractual \u201cbad faith\u201d litigation. Discovery requests will probe the extent to which AI shaped claim outcomes, while evidentiary questions will determine how such systems and their outputs are presented to a jury. Many of these challenges mirror those that arose during the Colossus<\/em> era, but modern AI adds layers of technical and procedural complexity.<\/p>\n\n\n\n

A. Scope of Discovery in Contractual and Extracontractual Claims<\/strong><\/p>\n\n\n\n

In coverage litigation, discovery will typically focus on what<\/strong> an insurer\u2019s AI system produced\u2014namely, the estimates, analyses, or other outputs that contributed to a coverage determination or valuation of the loss. These outputs are part of the claim file, and policyholders may seek to understand how those AI-generated results influenced the insurer\u2019s interpretation of the scope\/availability of coverage or the amount of loss. While discovery at this stage generally stops at the product of the system\u2019s analysis rather than its inner workings, courts may allow limited inquiry if the reliability of that output becomes relevant to the contractual dispute.<\/p>\n\n\n\n

In bad-faith litigation, discovery reaches further\u2014to how<\/strong> the AI system was used, understood, and supervised within the claim-handling process. Plaintiffs are likely to explore whether the insurer\u2019s reliance on an algorithm was reasonable and whether the adjuster understood the system well enough to make an informed decision. Depositions may examine the adjuster\u2019s familiarity with the AI tool, the extent of training received, and the level of discretion retained in accepting or rejecting its recommendations. Discovery may also probe the degree to which supervisors or management personnel were aware of or participated in the use of AI, and whether internal measures existed to ensure that adjusters were not applying AI outputs without oversight or human verification.<\/p>\n\n\n\n

The Colossus<\/em> disputes previewed this dynamic. Policyholders in those cases contended that adjusters had become conduits for opaque software rather than independent evaluators, resulting in mechanical and uniform claim decisions. The same argument will likely reemerge in AI-related bad-faith claims, reframed around machine learning and \u201cexplainability\u201d\u2014the ability of human decision-makers to articulate how and why they relied on algorithmic recommendations.<\/p>\n\n\n\n

B. Preservation of AI-Specific Data<\/strong><\/p>\n\n\n\n

Preservation duties apply in every claim dispute, but AI introduces a new wrinkle: models evolve.<\/em> Machine-learning systems are dynamic, updating their parameters as new data enters the training pipeline. If a claim decision is later challenged, it may be difficult\u2014or impossible\u2014to reproduce the precise model state that produced the disputed output unless it was preserved contemporaneously.<\/p>\n\n\n\n

Accordingly, insurers that employ AI should ensure that their litigation-hold procedures capture:<\/p>\n\n\n\n

    \n
  1. Model version identifiers and configuration files<\/strong>\u00a0\u2014 preserving the specific build or release number used during the relevant claim period;<\/li>\n\n\n\n
  2. Input and output data pairs<\/strong>\u00a0\u2014 the photos, documents, or structured data fed into the system and the resulting AI analysis or valuation;<\/li>\n\n\n\n
  3. System logs and metadata<\/strong>\u00a0\u2014 timestamps, user IDs, and confidence scores showing how and when the model processed each claim; and<\/li>\n\n\n\n
  4. Retraining documentation<\/strong>\u00a0\u2014 if the model was later updated, retaining records of when and how those updates occurred.<\/li>\n<\/ol>\n\n\n\n

    This level of preservation allows an insurer to recreate, if necessary, the analytic environment that produced the AI recommendation. Unlike ordinary claim documentation, these artifacts may exist outside the traditional claim file and require coordination between claims, IT, and data-science personnel. While courts have not yet ruled on preservation obligations specific to AI, best practice dictates erring on the side of completeness\u2014particularly where an insurer anticipates that its use of AI could become a contested issue.<\/p>\n\n\n\n

    C. Protecting Proprietary and Trade-Secret Information<\/strong><\/p>\n\n\n\n

    At present, relatively few insurers use AI tools in live claim environments, but those that do typically rely on vendor-developed, commercially available systems<\/strong> rather than proprietary in-house software. When litigation arises, policyholders may seek discovery into the inner workings of those systems\u2014source code, training data, or algorithmic logic\u2014to challenge the reliability of the insurer\u2019s reliance on them. Because such materials often belong to third-party vendors and contain trade-secret information, courts will likely balance competing interests: the policyholder\u2019s need for relevant discovery versus the vendor\u2019s right to protect its intellectual property.<\/p>\n\n\n\n

    Protective measures should mirror those employed in prior Colossus<\/em> disputes:<\/p>\n\n\n\n