PRIVILEGED AND CONFIDENTIAL — ATTORNEY-CLIENT COMMUNICATION / ATTORNEY WORK PRODUCT
MEMORANDUM
TO: Carolyn Whitford, General Counsel & EVP, Quorum Insurance Holdings Corp. Gregory Stahl, VP, Legal Affairs, Quorum Insurance Holdings Corp.
FROM: Wendy Schultz, Associate, intellectual Property, Calderwood & Harkness LLP Omar Siddiqui, Associate, intellectual Property, Calderwood & Harkness LLP
DATE: January 25, 2023
RE: Ownership of Machine-Learning Model Outputs — Copyright and Trade Secret Analysis; Quorum Insurance Holdings Corp. / Meridian Credit intelligence, Inc. Data License Negotiation
C&H Matter No.: 1025-00008
CC: Leonard Cross, Partner (Responsible Partner) Patricia Heng, Senior Counsel, Technology Transactions
Right-click to update Table of Contents
I. Executive Summary
You have asked us to analyze whether the outputs of Quorum Insurance Holdings Corp.'s ("Quorum") machine-learning models — trained in part on data licensed from Meridian Credit intelligence, Inc. ("Meridian") — constitute copyrightable "derivative works" under the Copyright Act or protectable trade secrets, and whether Meridian can assert intellectual property rights over such outputs. The Meridian data products at issue are the MCI Consumer Credit File™, MCI CommercialView™, and MCI AltSignal™ (collectively, the "Meridian Data Products"). This memorandum is intended to inform Quorum's negotiating position in the ongoing Master Data License Agreement ("MDLA") negotiation with Meridian.
The central question arises from a negotiation impasse over Article VII of the proposed MDLA. Meridian, represented by Kessler Pruitt & Ashe LLP (lead partner Graham Whitley), has taken the position that any "Derived Data" — defined in Meridian's draft as any data or output "incorporating, based on, or traceable to" Meridian's raw data — constitutes a "derivative work" under federal copyright law and therefore remains subject to Meridian's underlying intellectual property rights. On the basis of this characterization, Meridian has advanced three specific demands: (a) a perpetual, royalty-free license-back on all Derived Data for Meridian's own use; (b) a prohibition on third-party sublicensing of Derived Data unless it has been "sufficiently transformed" as determined in Meridian's sole discretion; and (c) a fifteen percent (15%) revenue share on all Derived Data that Quorum commercializes to external parties.
Quorum's position, which we believe is well-supported by prevailing law, is that Quorum should own all intellectual property in Derived Data — including data outputs, trained models, risk scores, indices, and analytics outputs — created using Meridian data as one of several inputs, and that Quorum should have unrestricted commercialization rights without payment of additional fees to, or requirement of consent from, Meridian.
The commercial stakes are substantial. Quorum's proprietary QuorumScore™ predictive underwriting model (USPTO application pending, Serial No. 97/482,116) blends Meridian credit data with Quorum's own first-party claims, policy, and telematics data to generate proprietary risk scores and underwriting recommendations. QuorumScore is central to Quorum's underwriting competitive advantage and is the foundation for Quorum's planned business-to-business ("B2B") licensing program, which management projects will generate $40–50 million in annual revenue by 2027. Acceptance of Meridian's demands would expose Quorum to annual payments of $6–7.5 million on top of the $12.3 million annual license fee, would give Meridian an effective veto over Quorum's commercialization strategy, and — most troubling — would give Meridian a license to exploit Quorum's own proprietary model outputs, potentially sharing them with Meridian's other data licensees or using them to enhance Meridian's own product portfolio in advance of a potential 2025–2026 IPO.
Our conclusions are as follows:
First, under prevailing copyright law, machine-learning model outputs — particularly live inference predictions generated from new, non-Meridian inputs — do not constitute "derivative works" of the training data. Meridian's underlying credit data consists primarily of uncopyrightable factual information under Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), and the statistical abstraction inherent in the machine-learning training process means that trained model weights do not "incorporate" copyrightable expression from the training data in any recognizable form.
Second, the legal landscape in this area is unsettled. No court has definitively ruled on whether ML model weights or inference outputs constitute derivative works of training data. This uncertainty makes contractual allocation paramount.
Third, trade secret protection offers a more robust and practical framework for protecting Quorum's model outputs than copyright. QuorumScore and its outputs qualify as Quorum's trade secrets. Meridian has no trade secret claim to Quorum's model outputs.
Fourth, the contractual provisions in Article VII of the MDLA must be carefully crafted to prevent Meridian from asserting rights through contract that it could not assert under copyright law alone. The MDLA will effectively define the parties' rights regardless of what background intellectual property law provides, making the drafting of Article VII dispositive.
We set forth our detailed analysis below.
II. Scope and Limitations
This memorandum addresses the ownership and protectability of Quorum's machine-learning model outputs under two bodies of law: (1) United States copyright law, specifically the Copyright Act of 1976, as amended, 17 U.S.C. §§ 101 et seq., and (2) federal and state trade secret law, including the Defend Trade Secrets Act of 2016 ("DTSA"), 18 U.S.C. §§ 1836 et seq., and applicable state trade secret statutes. Because the MDLA negotiation specifies New York law as the governing law, we address New York common-law trade secret principles where relevant, noting that New York has not adopted the Uniform Trade Secrets Act but applies analogous common-law principles.
This memorandum does not address the following topics, each of which is the subject of separate analysis:
• Patent issues relating to QuorumScore, including patent prosecution strategy, freedom-to-operate considerations, and the interplay between patent disclosure requirements and trade secret protection. These matters are being analyzed separately in the QuorumScore IP Protection Memorandum, in coordination with Angela Firth at Caldwell & Strauss LLP, Quorum's longstanding intellectual property counsel in Hartford, Connecticut.
• Privacy and regulatory compliance, including issues arising under the Fair Credit Reporting Act ("FCRA"), the Gramm-Leach-Bliley Act ("GLBA"), and the California Consumer Privacy Act ("CCPA"), as applicable to Quorum's use of Meridian consumer credit data. These dimensions are being addressed in separate memoranda prepared by David Cartwright and Lisa Johannsen of Calderwood & Harkness's Privacy & Data Security group.
This analysis is intended to inform Quorum's negotiating position on Article VII (Derived Data and IP Ownership) of the proposed MDLA with Meridian. It should be read in conjunction with the internal negotiation strategy memorandum dated December 15, 2022, and the forthcoming memoranda on market practice for license-back provisions and on the "sufficient transformation" standard, discussed in Part VIII below.
As a coordination note, we have consulted with Angela Firth at Caldwell & Strauss LLP regarding the implications of this analysis for Quorum's existing intellectual property portfolio. Ms. Firth's background guidance on Quorum's patent and trademark portfolio has been incorporated into relevant portions of this memorandum where appropriate, and we have ensured that our recommended contractual positions are consistent with her advice on QuorumScore patent prosecution strategy.
III. Factual Background
A. The Parties and the Transaction
Quorum Insurance Holdings Corp. ("Quorum") is a Delaware corporation, publicly traded on the Nasdaq Stock Market under the ticker symbol QIHC. Quorum is headquartered at 1200 Atlantic Street, Suite 3400, Stamford, Connecticut 06901. Quorum operates as a multi-line insurance carrier and insurtech platform, offering products and services across property and casualty ("P&C"), life insurance, and specialty lines. For the fiscal year ending December 31, 2022, Quorum reported total revenues of approximately $4.18 billion and total assets of approximately $27.6 billion. Quorum's strategic priority over the past three years has been the development and deployment of advanced analytics and machine-learning capabilities to enhance underwriting accuracy, reduce claims fraud, and develop new data-driven products and services for the insurance industry.
Meridian Credit intelligence, Inc. ("Meridian" or "MCI") is a Georgia corporation, privately held, with its principal offices at 3300 Peachtree Road NE, Suite 2100, Atlanta, Georgia 30326. Meridian is a consumer and commercial credit bureau that maintains and licenses comprehensive credit data on consumers and businesses throughout the United States. Meridian is majority-owned (approximately 62.5% equity) by Breckenridge Capital Partners Fund IV, L.P., a private equity fund. We understand from publicly available filings and Quorum's business intelligence that Whitmore Zanthor & Co. has been retained as Meridian's investment bank and is advising on a potential initial public offering targeted for 2025 or 2026. This context is relevant to understanding Meridian's commercial motivations in the MDLA negotiation, as discussed further below.
The Transaction. The proposed MDLA contemplates a five-year enterprise data-licensing arrangement under which Meridian would provide Quorum with access to the Meridian Data Products described in Part III.B below. The total contract value is estimated at approximately $60 million, structured as follows: $10.8 million in Year 1, $12.3 million in each of Years 2 through 5, plus a one-time implementation fee of $1.5 million. The contract was initially scoped by Quorum's procurement team led by Vanessa Chen (VP, Strategic Sourcing) in consultation with Rajesh Anand (Chief Data Officer) and Diane Lowell (SVP, Underwriting Analytics).
Quorum's intended use of the Meridian Data Products is to ingest Meridian's consumer credit-file data (encompassing approximately 310 million consumer records), commercial credit data (encompassing approximately 42 million business records), and alternative-data feeds (encompassing approximately 85 million consumer alt-data records) to enhance Quorum's automated underwriting models, claims-fraud detection algorithms, and customer propensity scoring across its P&C, life, and specialty insurance lines. The Meridian Data Products are intended to serve as one of several critical data inputs to Quorum's machine-learning pipeline, alongside Quorum's own first-party data assets.
B. Meridian's Licensed Data Products
The three Meridian Data Products that are the subject of the proposed MDLA are as follows:
MCI Consumer Credit File™. This product consists of full-file consumer tradeline data encompassing approximately 310 million consumer records. The data includes consumer credit accounts, payment histories, outstanding balances, credit inquiries, public records, and related information compiled from creditor reporting. Meridian delivers the MCI Consumer Credit File via daily batch feeds and a real-time application programming interface ("API") for on-demand individual consumer inquiries.
MCI CommercialView™. This product provides commercial credit data on approximately 42 million businesses in the United States, including business credit scores, payment performance data, financial statements, corporate linkage information, and related business credit intelligence. MCI CommercialView is delivered via weekly batch feeds.
MCI AltSignal™. This product contains alternative data derived from non-traditional credit sources, including utility payment records, telecommunications payment records, and rental payment histories, covering approximately 85 million consumer records. MCI AltSignal is designed to supplement traditional credit-file data with additional payment-behavior signals, particularly for consumers with limited traditional credit histories ("thin-file" or "credit-invisible" consumers). MCI AltSignal is delivered via daily batch feeds.
The total estimated monthly data volume under the proposed MDLA is approximately 437 million records, representing approximately 87.4% of the base tier of 500 million records per month specified in Meridian's volume-based pricing structure.
C. Quorum's Machine-Learning Pipeline and QuorumScore™
Quorum's data-science team, led by Chief Data Officer Rajesh Anand and SVP of Underwriting Analytics Diane Lowell, has developed the QuorumScore™ predictive underwriting model over a period of approximately two and a half years. QuorumScore is designed to generate proprietary risk scores, propensity scores, and fraud indicators that enhance Quorum's underwriting accuracy and efficiency across all lines of business.
QuorumScore blends Meridian credit data with Quorum's own proprietary first-party data assets, which include: (i) historical claims data spanning more than fifteen years of Quorum's underwriting experience; (ii) policy data, including policyholder demographics, coverage selections, and renewal patterns; (iii) telematics data collected from Quorum's connected-device programs for auto and property insurance; and (iv) customer behavioral data derived from Quorum's digital platforms, including web and mobile application interactions.
Quorum's machine-learning pipeline involves five principal stages, each of which is relevant to the legal analysis set forth in Part IV:
1. Ingestion and Preprocessing of Training Data. Raw data from multiple sources — including Meridian Data Products, Quorum's first-party data, and data from other third-party providers — is ingested into Quorum's data lake, cleaned, normalized, and prepared for use in model training.
2. Feature Engineering. Quorum's data scientists design and construct "features" — derived variables and statistical measures computed from the raw data that serve as inputs to the machine-learning model. Feature engineering is a highly proprietary process that reflects Quorum's domain expertise and data-science capabilities. Quorum's feature library for QuorumScore encompasses more than 1,400 engineered features.
3. Model Training. The machine-learning model is trained using the prepared training dataset. During training, the model iteratively processes the training data and adjusts its internal numerical parameters (commonly referred to as "weights") to minimize prediction error. The output of the training process is a trained model — a mathematical function defined by millions of numerical weight values that encode the statistical patterns learned from the training data. Critically, the trained model does not "contain" the training data in any recognizable or recoverable form; rather, it contains mathematical abstractions of patterns derived from the aggregate dataset.
4. Model Validation and Testing. The trained model is evaluated against held-out validation and test datasets to assess its predictive accuracy, calibration, and fairness. This stage may involve iterative refinement of the model architecture, hyperparameters, and feature set.
5. Deployment for Live Inference. Once validated, the trained model is deployed into Quorum's production environment, where it processes new data inputs in real time (or near-real time) and generates predictions — including risk scores, propensity scores, and fraud probability indicators. These live inference predictions are the primary outputs that Quorum intends to use operationally and, through the QuorumScore B2B licensing program, to commercialize to third parties.
It is essential to understand that during live inference, the trained model processes new data inputs that may or may not include Meridian data. In many use cases, the live inputs consist entirely of Quorum's first-party data (e.g., a new insurance application supplemented by telematics data), with no Meridian data present at the inference stage. In other use cases, a real-time Meridian credit pull may be one of several inputs processed during inference. In all cases, the output is a numerical prediction — a risk score, a propensity index, or a fraud probability — not a reproduction or reorganization of any input data.
Quorum has filed a utility patent application with the United States Patent and Trademark Office for the QuorumScore system and methodology, Serial No. 97/482,116. The application is currently pending. Management projects that B2B licensing of QuorumScore will generate $40–50 million in annual revenue by 2027, making QuorumScore a cornerstone of Quorum's growth strategy.
D. The Derived Data Dispute
The ownership and treatment of "Derived Data" is the central unresolved issue in the MDLA negotiation. It was flagged at the January 18, 2023, first in-person negotiation session in New York — attended by Quorum's team (Gregory Stahl, Rajesh Anand, Vanessa Chen, and outside counsel Leonard Cross and Wendy Schultz of Calderwood & Harkness) and Meridian's team (Amanda Cruz (VP, Enterprise Licensing), Graham Whitley (KP&A), and Jordan Kiefer (Senior Director, Data Products)) — as one of two principal open issues (the other being data security and audit provisions in Article IX).
Meridian's position, as set forth in the November 4, 2022, draft of the MDLA prepared by Kessler Pruitt & Ashe LLP, defines "Derived Data" expansively to include any output "incorporating, based on, or traceable to" Meridian raw data. Meridian characterizes all such Derived Data as a "derivative work" under the Copyright Act and asserts that Meridian retains underlying intellectual property rights in all Derived Data. On this basis, Meridian has advanced three specific demands:
(a) Perpetual Royalty-Free License-Back. Quorum would be required to grant Meridian a perpetual, royalty-free, non-exclusive license to all Derived Data for Meridian's own internal and commercial use.
(b) Sublicensing Restriction — "Sufficient Transformation" Standard. Quorum would be prohibited from sublicensing or distributing Derived Data to third parties unless the data has been "sufficiently transformed" such that Meridian raw data cannot be "reverse-engineered, reconstructed, or identified," with Meridian having sole discretion to determine whether the "sufficient transformation" standard has been met.
(c) Revenue Share. Quorum would pay Meridian fifteen percent (15%) of net revenues attributable to any Derived Data that Quorum commercializes to external parties.
Quorum's position is that Quorum should own all intellectual property in Derived Data — including data outputs, trained model weights and parameters, algorithms, risk scores, indices, and analytics outputs — created using Meridian data as one input among many, and that Quorum should have unrestricted rights to use, sublicense, and commercialize such Derived Data without payment of additional fees to Meridian and without requiring Meridian's consent.
The following Parts IV through VIII analyze the legal basis for Quorum's position and provide our recommendations for the MDLA negotiation.
IV. Copyright Analysis: Are ML Model Outputs "Derivative Works"?
A. Statutory Framework — Derivative Works Under 17 U.S.C. § 101
The Copyright Act defines a "derivative work" as:
a work based upon one or more preexisting works, such as a translation, musical arrangement, dramatization, fictionalization, motion picture version, sound recording, art reproduction, abridgment, condensation, or any other form in which a work may be recast, transformed, or adapted. A work consisting of editorial revisions, annotations, elaborations, or other modifications which, as a whole, represent an original work of authorship, is a "derivative work."
17 U.S.C. § 101. Under Section 106(2), the copyright owner has the exclusive right to "prepare derivative works based upon the copyrighted work." 17 U.S.C. § 106(2).
Several requirements must be satisfied for a work to constitute a derivative work subject to the original copyright owner's exclusive rights:
First, the derivative work must be "based upon" one or more preexisting works — that is, it must incorporate material from the preexisting work. The mere fact that a new work was inspired by, or created with reference to, an existing work does not make it a "derivative work" within the meaning of the statute. The new work must actually incorporate copyrightable expression from the preexisting work.
Second, the derivative must incorporate a copyrightable portion of the preexisting work. If the elements borrowed from the preexisting work are not themselves protected by copyright — for example, if they are uncopyrightable facts, ideas, or methods — then the new work does not implicate the original copyright owner's derivative-work right, regardless of how extensively the new work draws on the preexisting material.
Third, courts generally require "substantial similarity" — the derivative must contain recognizable elements of the underlying copyrighted work. See, e.g., Litchfield v. Spielberg, 736 F.2d 1352, 1357 (9th Cir. 1984); Warner Bros. Inc. v. Am. Broad. Cos., 720 F.2d 231, 240 (2d Cir. 1983). A work that has been so thoroughly transformed that no recognizable copyrightable expression from the original remains is not a derivative work.
It is important to note that copyright in a derivative work extends only to the new material contributed by the author of the derivative work, and does not imply any exclusive right in the preexisting material. 17 U.S.C. § 103(b). The copyright owner of the preexisting work retains copyright in the preexisting material but gains no rights in the new contributions of the derivative author.
These principles are foundational to our analysis of whether Quorum's ML model outputs can properly be characterized as "derivative works" of Meridian's underlying data.
B. Threshold Question: Is Meridian's Raw Data Copyrightable?
Before analyzing whether Quorum's model outputs are "derivative works" of Meridian's data, we must first address whether Meridian's underlying data is copyrightable at all. If it is not, then the derivative-works analysis is largely moot, because one cannot create a "derivative work" of uncopyrightable material.
The Supreme Court's decision in Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), is the controlling authority on the copyrightability of factual compilations. In Feist, the Court held that factual compilations receive only limited copyright protection: the copyright extends to the compiler's original selection, coordination, and arrangement of the facts, but not to the underlying facts themselves. The Court emphasized that "facts are not copyrightable" because "copyright asserts no control over facts." Id. at 344–45. The Court further held that a compilation of facts is copyrightable only to the extent it features an "original selection or arrangement of facts," and that the "sweat of the brow" doctrine — which had previously awarded copyright protection based on the labor invested in gathering facts — was inconsistent with the Constitution's originality requirement. Id. at 352–54.
Meridian's credit data consists of factual information: consumer tradeline data (account numbers, balances, payment histories, credit limits, account statuses), business credit data (business credit scores, payment performance records, financial statements), and alternative data (utility payment records, telecommunications payment records, rental payment histories). Each individual data record — a consumer's credit score, a payment history entry, an account balance — is a fact. Under Feist, these individual facts are not copyrightable. No amount of labor or expense invested by Meridian in compiling these facts changes this fundamental principle.
Meridian's databases may be protectable as compilations to the extent that Meridian's particular selection, coordination, and arrangement of data elements reflect original creative choices. For example, if Meridian exercises original judgment in selecting which data fields to include, how to organize them, or how to present them in relation to one another, those choices could receive thin compilation copyright protection. However, even if Meridian's databases are copyrightable compilations, the copyright extends only to the original selection, coordination, and arrangement — not to the underlying data elements themselves. A party that extracts individual facts from a copyrighted compilation does not infringe the compilation copyright, so long as it does not replicate the original selection, coordination, and arrangement. See Feist, 499 U.S. at 350–51.
The relevance of this analysis to the Derived Data dispute is significant. If the underlying data elements ingested by Quorum's ML pipeline are uncopyrightable facts, then outputs derived from processing those facts — whether model weights, inference predictions, or blended scores — are even less likely to constitute "derivative works" of Meridian's compilation. The model does not replicate Meridian's selection, coordination, or arrangement; it processes individual data elements (facts) through a statistical algorithm to produce entirely new outputs.
We anticipate that Meridian's counsel may respond to this analysis by shifting the basis for Meridian's claims from copyright to trade secret and contractual protections. Indeed, it is our assessment that trade-secret and contractual arguments — not copyright — are the true foundation for Meridian's demands, even though Meridian has strategically deployed copyright terminology ("derivative work") in the MDLA draft to lend an air of statutory authority to what are, in reality, purely contractual claims. We address trade secret issues in Part V below and contractual issues in Parts VII and VIII.
C. Application to ML Training Data
The application of derivative-works doctrine to machine-learning systems requires careful analysis of the three distinct stages of the ML pipeline, because the relationship between Meridian's data and Quorum's outputs changes fundamentally at each stage.
Stage 1: The Training Data Itself
At this initial stage, Meridian's raw data is ingested as one of several data sources into Quorum's training pipeline. During this phase, Meridian's data exists in its original form (or a preprocessed form derived from it) within Quorum's data environment. If Quorum were to reproduce, distribute, or publicly display Meridian's data at this stage, it would raise potential copyright issues (to the extent compilation protection applies) as well as contractual and trade secret concerns. However, the use of data for training purposes — as input to a computational process that will produce a fundamentally different output — is qualitatively different from reproduction of the data.
This stage is not the focus of the Derived Data dispute, because Quorum does not intend to redistribute Meridian's raw data. The dispute concerns the outputs generated after the data has been processed through Quorum's ML pipeline.
Stage 2: Trained Model Weights and Parameters
Machine-learning training is a mathematical optimization process. During training, the ML algorithm iteratively processes training examples (data points) and adjusts the model's numerical parameters (weights) to minimize a loss function — that is, to reduce the difference between the model's predictions and the actual outcomes in the training data. The output of this process is a trained model: a mathematical function defined by millions (or, in deep learning architectures, potentially billions) of numerical weight values.
These weight values encode the statistical patterns, correlations, and relationships that the algorithm has identified in the training data. Crucially, however, the weights do not "contain" the training data in any meaningful sense. One cannot recover individual training examples from the model weights. The weights are mathematical abstractions — numerical values such as "0.00347" or "-1.29482" — that represent learned statistical relationships. They bear no resemblance to, and do not incorporate any recognizable expression from, individual data records in the training set.
An apt analogy is the relationship between a textbook and the knowledge gained by a student who reads it. A student who reads a medical textbook and absorbs its teachings gains knowledge and understanding derived from the textbook. But the knowledge in the student's mind is not a "copy" of the textbook, and a medical diagnosis made by that student — even one informed by knowledge gained from the textbook — is not a "derivative work" of the textbook. Similarly, a machine-learning model that is trained on data absorbs statistical patterns from that data, but the resulting model weights are not a reproduction of the data, and outputs generated by the model are not derivative works of the data.
This analogy, while imperfect (a trained ML model operates through mathematical computation rather than human cognition), captures the fundamental point: the training process is a transformative statistical abstraction, not a reproductive process. The model does not recast, transform, or adapt the training data into a new form of expression; rather, it extracts abstract statistical patterns from the data and encodes those patterns as numerical values.
We note that neither the U.S. Copyright Office nor any court has, as of the date of this memorandum, definitively addressed whether ML model weights constitute derivative works of training data. This is an area of active scholarly debate and ongoing policy discussion. The Copyright Office's 2022 Notice of Inquiry on Artificial intelligence and Copyright (87 Fed. Reg. 59,942 (Oct. 3, 2022)) solicited public comment on, among other things, the question of whether AI-generated outputs implicate the derivative-work right. Multiple pending litigation matters — including Andersen v. Stability AI Ltd. (N.D. Cal., filed Jan. 13, 2023), Kestrel Industries (US), Inc. v. Stability AI, Inc. (D. Del., filed Feb. 3, 2023), and the New York Times litigation against Ashgrove Industries (anticipated but not yet filed as of this memorandum's date) — may produce judicial guidance on these questions, though likely not for several years.
Notwithstanding this uncertainty, we believe the stronger argument is that trained model weights are not derivative works of the training data. The weights do not incorporate copyrightable expression from any particular training example in recognizable form. They do not recast, transform, or adapt any particular copyrighted work. They are, at most, indirect statistical reflections of aggregate patterns across millions of data points from multiple sources. This is insufficient to satisfy the derivative-works standard under Section 101.
This conclusion is further reinforced in the present case by the fact that Meridian's underlying data consists of uncopyrightable facts (as discussed in Part IV.B above). Even if trained model weights could in theory constitute derivative works of copyrightable training data, they cannot constitute derivative works of uncopyrightable factual inputs.
Stage 3: Inference Outputs
The final and most commercially significant stage is live inference — the generation of predictions, scores, or classifications when the trained model is applied to new input data in Quorum's production environment. When a new insurance application is submitted, for example, QuorumScore processes the applicant's data (which may include a real-time Meridian credit pull, Quorum's own first-party data, and data from other sources) through the trained model and generates a numerical risk score (e.g., "742") or a fraud probability (e.g., "0.87").
These inference outputs represent a further degree of removal from Meridian's training data:
• If the new inputs processed during inference do not include Meridian data — for example, if the inference is based solely on Quorum's first-party telematics data — then the output is multiple steps removed from Meridian's data. The chain is: Meridian data → training process → model weights → inference on non-Meridian inputs → output. At each stage, the connection to Meridian's original data becomes more attenuated and abstract.
• Even if the new inputs processed during inference do include some Meridian data (e.g., a real-time credit score pulled from Meridian's API), the output is a numerical prediction — a single number or set of numbers reflecting the model's assessment of risk, propensity, or fraud probability. This numerical output does not contain any recognizable copyrightable expression from Meridian's underlying data. A risk score of "742" does not incorporate, recast, or adapt any element of the Meridian credit file from which one of its inputs may have been drawn. It is a fundamentally new and different type of information — a predictive judgment, not a factual record.
For these reasons, inference outputs are, in our assessment, clearly not "derivative works" of Meridian's training data or of any Meridian data used as a live input. They do not incorporate copyrightable expression from Meridian's data. They do not bear any recognizable similarity to Meridian's data. They are the product of Quorum's proprietary algorithms and models applied to a mix of data inputs, producing outputs that are qualitatively different in kind from any of the inputs.
D. Case Law and Analogies
No court has issued a directly controlling ruling on the question of whether ML model outputs constitute derivative works of training data. As of January 2023, this remains a question of first impression. However, several lines of authority provide persuasive analogies and doctrinal foundations:
*Feist Publications, Inc. v. Rural Telephone Service Co.*, 499 U.S. 340 (1991). As discussed in Part IV.B, Feist establishes that facts extracted from a copyrightable compilation are free to use. This principle applies directly here: to the extent Quorum's models extract factual information (statistical patterns, correlations) from Meridian's data, those facts are free for Quorum to use. The Copyright Act "does not protect facts." Id. at 344.
*Authors Guild v. Severin Industries, Inc.*, 804 F.3d 202 (2d Cir. 2015). In this case, the Second Circuit held that Severin Industries's mass digitization of copyrighted books for the purpose of creating a searchable index was a transformative fair use. The court emphasized that the purpose of Severin Industries's use was "highly transformative" — Severin Industries's search engine provided "a new and different function" from the original books. Id. at 216–17. The analogy to ML training is persuasive: Quorum ingests Meridian's data not to reproduce or display it, but to extract statistical patterns for the purpose of generating predictions — a fundamentally different function from the storage and reporting of credit data. The Second Circuit's emphasis on the transformative nature of a use that "communicates something new and different from the original" strongly supports the conclusion that ML model outputs are not derivative works of training data.
*Drayton & Co. v. Calderon Corporation*, 593 U.S. 1 (2021). The Supreme Court's most recent major fair use decision addressed Severin Industries's copying of certain Java API declarations for use in the Android operating system. The Court held that Severin Industries's use was a transformative fair use, emphasizing that Severin Industries's purpose was to create "a new and different task-related system for a new and different computing environment" and that the value of Severin Industries's use lay in its own creative contributions, not in the expression it had copied. Id. at 36–37. While the fair use context differs from the derivative-works question presented here, Ridgeline Industries v. Severin Industries reinforces the general principle that the more transformative a use is — the more it produces something "new and different" from the original — the less likely it is to infringe the original copyright owner's rights or to create derivative-work obligations.
General principle. Across these authorities, a consistent theme emerges: where a use transforms the original work to produce something functionally new and different, the original copyright owner's rights are limited. The use of raw data to train a machine-learning model that generates numerical predictions is, by any measure, a highly transformative use. The original data (consumer credit records) serves the purpose of providing factual information about consumer creditworthiness; the model output (a predictive risk score) serves the fundamentally different purpose of informing an underwriting decision. No recognizable expression from the original data survives in the output.
Scholarly commentary further supports this analysis. Leading commentators have observed that ML model training is fundamentally a statistical process, not a reproductive one, and that the derivative-works framework was designed for cases in which copyrightable expression is recognizably incorporated into a new work — not for cases in which factual data is processed into abstract mathematical parameters. See, e.g., Mark A. Lemley & Bryan Casey, Fair Learning, 99 Tex. L. Rev. 743 (2021) (arguing that ML training on copyrighted works is transformative and should be treated as fair use); Matthew Sag, Copyright and Copy-Reliant Technology, 103 Nw. U. L. Rev. 1607 (2009) (analyzing "non-expressive" or "copy-reliant" uses of copyrighted works for computational purposes); Andres Guadamuz, Artificial intelligence and Copyright, WIPO Magazine (Oct. 2017) (discussing the inapplicability of traditional derivative-works analysis to AI outputs).
We reiterate that the legal landscape in this area is rapidly evolving. Pending litigation matters (including those referenced in Part IV.C above), ongoing Copyright Office proceedings, and potential congressional action could alter the analysis presented here. This reinforces our recommendation that the MDLA must be drafted to allocate rights clearly by contract, rather than relying on uncertain background copyright law.
E. Analysis of Meridian's "Traceability" Standard
Meridian's proposed MDLA defines "Derived Data" as any data, output, analysis, model, or derivative thereof that "incorporates, is based on, or is traceable to" Meridian raw data. The "traceability" standard warrants particular scrutiny because it dramatically expands the scope of Meridian's claimed rights beyond anything that copyright law would support.
Under the copyright derivative-works framework, as discussed above, a work is a "derivative work" only if it incorporates copyrightable expression from the preexisting work. The work must contain recognizable elements of the original. This is a substantive standard that requires actual incorporation of protected expression.
The "traceability" standard, by contrast, has no such requirement. Under this standard, any output of any model that was trained, even in part, on Meridian data would be deemed "Derived Data" subject to Meridian's asserted IP rights — regardless of whether the output contains, incorporates, or resembles any element of Meridian's data. The standard would sweep in:
• Model weights trained on datasets in which Meridian data constituted a minority of the input (as is the case with QuorumScore, which also incorporates Quorum's substantial first-party data and data from other third-party providers);
• Inference outputs generated from inputs that contain no Meridian data whatsoever, simply because the model was trained in part on Meridian data at some prior point;
• Blended scores and indices in which Meridian data contributed only marginally to the output alongside Quorum's own proprietary data and algorithms;
• Downstream analytical products built upon any of the above.
The "traceability" standard would effectively give Meridian an intellectual property claim over outputs of models that use hundreds of other data sources and that reflect years of Quorum's own proprietary algorithmic development, simply because Meridian data was one input to the training process. Carried to its logical conclusion, this standard would give Meridian IP claims over QuorumScore itself, despite the fact that QuorumScore reflects Quorum's independent investment, innovation, and proprietary data assets.
This standard has no basis in copyright law. It is a purely contractual expansion of Meridian's rights that would grant Meridian control over assets that Meridian did not create, did not invest in, and to which Meridian has no legal entitlement under any background IP regime.
Risk assessment. If Quorum agrees to the "traceability" standard in the MDLA, Quorum would be contractually granting Meridian rights that far exceed what copyright law (or trade secret law) would provide. Those contractual rights would be enforceable in accordance with their terms, regardless of the weakness of Meridian's underlying IP claims. This is precisely the outcome that the MDLA negotiation must avoid.
Recommendation. We recommend that Quorum categorically reject the "traceability" standard. If any contractual test is adopted to define the scope of Meridian's protectable interest in Quorum's outputs, it should be based on an objective, narrow criterion: specifically, whether the output contains or can be used to reconstruct identifiable, non-aggregated Meridian raw data records. This standard protects Meridian's legitimate interest in preventing the redistribution of its raw data while preserving Quorum's ownership of the model outputs that are the product of Quorum's independent investment and innovation.
F. Conclusions on Copyright
Based on the foregoing analysis, we reach the following conclusions:
1. ML model weights and parameters are very unlikely to be "derivative works" of training data under current copyright doctrine. The training process is a statistical abstraction, not a reproductive process. Trained model weights do not incorporate copyrightable expression from any particular training example in recognizable form.
2. Live inference outputs (predictions, scores, and classifications) are even less likely to be derivative works, particularly when generated from non-Meridian inputs. These outputs are multiple layers of transformation removed from the training data and bear no resemblance to any element of that data.
3. Meridian's factual credit data has limited (if any) copyright protection. Under Feist, individual data records are uncopyrightable facts. Meridian's databases may receive thin compilation protection for original selection, coordination, and arrangement, but the underlying data elements are free for others to use.
4. Meridian's "derivative work" characterization in its proposed MDLA overreaches copyright law. There is no basis in the Copyright Act for the proposition that ML model outputs constitute derivative works of factual training data, and the "traceability" standard proposed by Meridian is wholly foreign to the derivative-works framework.
5. However, the unsettled nature of the law means contractual allocation is critical. Because no court has definitively ruled on these questions, the parties cannot rely on background copyright law to resolve their dispute with certainty. The MDLA will effectively define the parties' rights regardless of what copyright law provides, making the drafting of Article VII dispositive. Quorum must ensure that Article VII reflects Quorum's ownership of its model outputs, rather than deferring to Meridian's self-serving copyright characterization.
V. Trade Secret Analysis: Protection of Quorum's ML Model Outputs
A. Trade Secret Framework
Trade secret protection is available under both federal and state law. The Defend Trade Secrets Act of 2016 ("DTSA"), 18 U.S.C. §§ 1836 et seq., provides a federal civil cause of action for the misappropriation of trade secrets. At the state level, while New York has not adopted the Uniform Trade Secrets Act ("UTSA"), New York courts apply common-law trade secret principles that are substantially consistent with the UTSA framework.
Under both the DTSA and New York common law, trade secret protection requires three elements:
1. Independent Economic Value. The information must derive independent economic value, actual or potential, from not being generally known to, and not being readily ascertainable through proper means by, other persons who can obtain economic value from the information's disclosure or use. 18 U.S.C. § 1839(3)(B). New York courts have described this element as requiring that the information provides a competitive advantage. See, e.g., Ashland Mgmt. Inc. v. Janien, 82 N.Y.2d 395, 407 (1993).
2. Reasonable Measures to Maintain Secrecy. The owner of the information must have taken reasonable measures to keep the information secret. 18 U.S.C. § 1839(3)(A). This includes physical and electronic security measures, access controls, confidentiality agreements, and other safeguards appropriate to the nature and value of the information. Under New York law, the reasonableness inquiry is fact-specific and considers the totality of the precautions taken. See North Atl. Instruments, Inc. v. Haber, 188 F.3d 38, 48 (2d Cir. 1999) (applying New York law).
3. Not Generally Known or Readily Ascertainable. The information must not be generally known to, or readily ascertainable by proper means by, the relevant public or industry participants. Information that is publicly available, commonly known in the industry, or easily reverse-engineered does not qualify for trade secret protection.
Trade secret protection is particularly well-suited to machine-learning models, trained weights, proprietary algorithms, feature engineering methodologies, and scoring outputs. Unlike copyright, which requires originality of expression, or patent, which requires novelty and nonobviousness, trade secret protection applies to any information that meets the three elements above — regardless of whether it constitutes "expression" or "invention" in the traditional sense. This makes trade secret law the most natural and robust IP framework for protecting ML assets.
B. QuorumScore™ as Trade Secret
QuorumScore's predictive model, including its architecture, training methodology, feature selection and engineering, hyperparameter settings, trained weights, and inference logic, clearly qualifies for trade secret protection under both the DTSA and New York common law.
Independent Economic Value. QuorumScore is central to Quorum's underwriting competitive advantage. It enables Quorum to price risk more accurately, detect fraud more effectively, and identify profitable customer segments more efficiently than competitors relying on traditional underwriting methods. Management projects that QuorumScore's B2B licensing program will generate $40–50 million in annual revenue by 2027. The independent economic value of QuorumScore derives directly from the fact that its algorithmic methodology, trained model weights, and feature engineering are not known to Quorum's competitors. If a competitor obtained access to QuorumScore's trained model or its methodology, it could replicate Quorum's underwriting advantage, causing substantial competitive harm. This element is clearly satisfied.
Reasonable Measures to Maintain Secrecy. Quorum maintains comprehensive data security and confidentiality protocols. These include:
• Role-based access controls limiting access to QuorumScore's model architecture, weights, and source code to authorized data-science and engineering personnel;
• All employees and contractors with access to QuorumScore are subject to confidentiality and invention assignment agreements;
• Quorum's data environment is SOC 2 Type II certified (audit performed by Felton Thorpe & Co. LLP, covering the period ending December 31, 2023);
• Technical controls including encryption at rest and in transit, audit logging of all access to model assets, and network segmentation isolating the ML development environment from general corporate systems;
• External parties receiving QuorumScore outputs (e.g., B2B licensing customers) will be subject to contractual use restrictions and confidentiality obligations.
These measures are more than sufficient to satisfy the "reasonable measures" element under both the DTSA and New York common law.
Not Generally Known or Readily Ascertainable. QuorumScore's model architecture, feature library (comprising more than 1,400 engineered features), hyperparameter settings, and trained weights are proprietary to Quorum and are not publicly available, published, or otherwise accessible to competitors or the public. While the general concept of using machine learning for insurance underwriting is widely known in the industry, Quorum's specific implementation — the particular combination of data sources, features, algorithmic choices, and training methodology — is not. A competitor could not readily ascertain these details through proper means.
Model outputs as trade secrets. In addition to the model itself, the outputs of QuorumScore — the specific risk scores, propensity indices, and fraud probability indicators generated for individual insurance applicants or policyholders — may also constitute trade secrets to the extent they reflect proprietary algorithmic judgments not generally known to competitors. A particular QuorumScore risk score for a given applicant reflects the combined effect of Quorum's proprietary model architecture, training, feature engineering, and data inputs — information that a competitor could not replicate without access to Quorum's model. While individual scores, once disclosed to the customer, may lose their secret character, the scoring methodology and the aggregate corpus of scoring outputs remain protectable.
Note on patent disclosure. We note that Quorum's pending USPTO application for QuorumScore (Serial No. 97/482,116) may, if the patent issues, require public disclosure of certain aspects of the QuorumScore methodology. This could potentially limit trade secret protection for disclosed elements. We recommend continued coordination with Angela Firth at Caldwell & Strauss LLP to ensure that the patent prosecution strategy is aligned with Quorum's trade secret protection objectives — for example, by carefully considering which aspects of QuorumScore are disclosed in the patent specification and which are maintained as trade secrets.
C. Does Meridian Have Trade Secret Rights in Quorum's Outputs?
Meridian can properly assert trade secret protection over its own raw data — the consumer credit files, commercial credit data, and alternative data feeds that comprise the Meridian Data Products. This data is not publicly available (individual consumers and businesses have access to their own records, but the comprehensive compiled datasets are proprietary to Meridian), derives independent economic value from being licensed to data consumers like Quorum, and is maintained under security protocols (as required by the MDLA itself and by applicable FCRA and GLBA regulations).
However, Meridian's trade secret protection over its raw data does not extend to the outputs of Quorum's independent processing of that data. The transformation of raw data through Quorum's ML pipeline — ingestion, preprocessing, feature engineering, model training, and inference — fundamentally changes the nature and character of the information. What emerges from the pipeline (trained model weights, risk scores, fraud probabilities) is qualitatively different from what went in (consumer tradeline records, payment histories, account balances). The outputs reflect Quorum's proprietary algorithmic choices, feature engineering, and independent data contributions as much as — or more than — any individual Meridian data element.
An analogy illustrates the point. A steel manufacturer that purchases iron ore from a mining company does not cede trade secret rights in its proprietary alloy formulations, manufacturing processes, or finished steel products to the ore supplier. The ore supplier has trade secret rights in its own mining data, geological surveys, and extraction techniques — but not in the finished products manufactured by the buyer using the ore as one raw material among many, combined with the buyer's own proprietary processes. Similarly, Meridian has trade secret rights in its own raw data but not in the finished analytical products that Quorum manufactures using that data as one input among many, combined with Quorum's own proprietary data, algorithms, and analytical methodologies.
The key distinction is that Quorum is not redistributing Meridian's data. Quorum is using Meridian's data as one of many inputs to generate fundamentally different outputs. The transformation involved is not merely cosmetic or superficial — it is a statistical and computational transformation that produces outputs bearing no resemblance to the inputs. A risk score of "742" does not disclose, contain, or reveal any element of the Meridian consumer credit record that contributed to its generation.
We note, however, that contractual restrictions can create enforceable obligations beyond what trade secret law would independently provide. This is precisely what Meridian is attempting through the license-back and revenue-share provisions in its proposed MDLA. Meridian may not have trade secret rights in Quorum's model outputs under background law, but if Quorum contractually grants Meridian a license-back or revenue-share right, that contractual commitment will be enforceable in accordance with its terms. This underscores the critical importance of the MDLA negotiation, as discussed in Parts VII and VIII below.
D. Interaction Between Trade Secret and Contractual Provisions
The interplay between trade secret law and contract is a critical consideration for the MDLA negotiation. Two principles are particularly relevant:
First, contractual provisions can expand a party's rights beyond what background trade secret law would provide. A data licensor can, by contract, obtain rights in a licensee's outputs that the licensor would have no claim to under trade secret law alone. Courts routinely enforce contractual restrictions on the use of licensed data, including restrictions on the creation and exploitation of derivative products, even where such restrictions exceed the scope of the licensor's background IP rights. See, e.g., ProCD, Inc. v. Zeidenberg, 86 F.3d 1447 (7th Cir. 1996) (enforcing contractual restrictions on the use of uncopyrightable data).
Second, a contract provision granting Meridian ownership, license-back, or revenue-share rights in Quorum's model outputs would be enforceable against Quorum, subject only to narrow defenses such as unconscionability, duress, or public policy. In a negotiated commercial transaction between sophisticated parties represented by counsel, these defenses are unlikely to be available.
This dynamic reinforces the paramount importance of the MDLA Article VII negotiation. The background law — both copyright and trade secret — is favorable to Quorum. But the contractual provisions will supersede the background law. If Quorum agrees to contractual provisions that grant Meridian rights in Quorum's model outputs, those provisions will be enforceable regardless of the weakness of Meridian's underlying IP claims. Conversely, if Quorum successfully negotiates contractual provisions that clearly vest ownership of model outputs in Quorum and disclaim any Meridian rights, those provisions will provide definitive protection for Quorum's interests.
Recommendation. Quorum should refuse contractual provisions that would grant Meridian any ownership, license-back, or revenue-share rights in Quorum's model outputs. Such provisions would create enforceable rights where none exist under background law and would undermine Quorum's ability to exploit its own proprietary assets. The MDLA should clearly allocate ownership of model outputs to Quorum, with no contractual encumbrances in favor of Meridian.
VI. The Critical Distinction: Model Weights vs. Inference Outputs
A. Technical Explanation
To properly frame the legal analysis and inform the MDLA drafting, it is essential to understand the technical distinction between trained model weights and inference outputs.
Model Weights/Parameters. When a machine-learning model is trained, the training algorithm processes the training dataset and iteratively adjusts the model's internal numerical parameters — commonly called "weights" — to minimize prediction error. In a neural network architecture of the type used by QuorumScore, the model may contain millions of individual weight values, each of which is a floating-point number (e.g., 0.00347, -1.29482, 2.71828). Collectively, these weights define the mathematical function that the model computes. The weights are the product of the training process: they encode the statistical patterns, correlations, and regularities that the training algorithm identified in the aggregate training dataset.
Critically, the training data itself is consumed during the training process — it is not stored within, or recoverable from, the model weights. The relationship between training data and model weights is analogous to the relationship between raw ingredients and a cooked dish: the ingredients are combined and transformed through a process that produces something qualitatively different, and one cannot recover the original ingredients from the final product. A researcher could not examine QuorumScore's millions of weight values and reconstruct any individual Meridian consumer credit record. The weights are mathematical abstractions, not data records.
Inference Outputs. Once the model is trained and deployed, it is used for "inference" — the process of generating predictions by applying the trained model to new input data. When QuorumScore is deployed in production, it receives a vector of input features (derived from a new insurance applicant's data) and processes that vector through the trained model, producing an output — typically a numerical score (e.g., a risk score of "742") or a probability (e.g., a fraud probability of "0.87").
The inference output is a function of two things: (1) the trained model weights, and (2) the new input data. It is not a function of the original training data, except insofar as the training data influenced the model weights during the training process. The output is a new piece of information — a predictive judgment generated by the model — not a reproduction, transformation, or adaptation of any element of the training data or the inference input data.
B. Legal Significance of the Distinction
The distinction between model weights and inference outputs carries significant legal implications for the Derived Data dispute.
Even if one were to argue — contrary to our analysis in Part IV — that trained model weights are somehow "derived from" or "based on" the training data, this would not automatically make inference outputs derivative of the training data as well. Inference outputs represent an additional layer of transformation beyond the training process. The chain is:
Training data → training process → model weights → inference on new inputs → output
At each stage, the connection to Meridian's original data becomes more attenuated and abstract. By the time an inference output