# Large language models and the structure of legal reasoning

## Introduction

The integration of Large Language Models (LLMs) into the legal domain has precipitated a profound shift in how artificial intelligence is conceptualized within the framework of jurisprudence and legal informatics. In the initial phases of LLM development, evaluations of legal reasoning capabilities relied heavily on standardized examinations, most notably the Uniform Bar Examination [cite: 1, 2]. The ability of frontier models to achieve passing scores on these standardized tests led to premature, highly publicized declarations regarding the models' capacity for genuine legal reasoning and their potential to automate complex legal analysis. However, achieving high accuracy on standardized multiple-choice questions fundamentally conflates factual recall and surface-level pattern matching with the complex, multi-step inference required in real-world legal adjudication [cite: 3, 4]. 

To rigorously assess what LLM behavior reveals about the underlying structure of legal reasoning, it is necessary to move beyond proxy metrics and examine the models through dual, intersecting lenses: empirical performance on structured logical tasks and theoretical alignment with jurisprudential philosophy. Recent evaluations across disparate global legal systems—ranging from the common law tradition of the United States to the codified civil law systems of China and Poland, and extending to the pluralistic traditions of Islamic jurisprudence—demonstrate a consistent and highly revealing divergence in model capabilities [cite: 5, 6, 7, 8]. Across all these domains, LLMs exhibit remarkable proficiency in statutory rule extraction and metadata recall, but they suffer severe performance degradation when tasked with analogical application and the resolution of conflicting normative principles [cite: 9, 10].

This performance dichotomy is not merely a technical limitation; it actively mirrors historic debates in the philosophy of law, specifically the tension between legal positivism and interpretivism. By treating language as a mathematical construct optimized for probabilistic next-token generation, LLMs inherently operationalize a hyper-formalist, textualist methodology [cite: 11]. They process the law as a set of discrete, observable rules, aligning closely with H.L.A. Hart’s positivist framework of "primary rules," while simultaneously failing to comprehend the institutional realities and social authority required by "secondary rules" [cite: 12, 13]. More critically, the neural architecture of generative models is fundamentally incompatible with Ronald Dworkin’s interpretivist requirement that legal reasoning must integrate moral principles, community standards, and constructive interpretation [cite: 14, 15]. 

This report delivers an exhaustive analysis of the intersection between machine learning architecture and legal theory. By analyzing recent empirical data (2023–2026), mechanistic interpretability studies of attention weights, and critical jurisprudential perspectives, the following analysis investigates how the "epistemic gap" inherent in generative interpretation severs the artifact of legal text from the legitimate, normative process of legal reasoning [cite: 11, 16].

## The Illusion of Competence: Beyond Standardized Testing

The presumption that Large Language Models intrinsically "understand" the law has been systematically dismantled by recent, highly granular benchmarking efforts designed collaboratively by legal professionals and computer scientists. These benchmarks move deliberately beyond the binary correctness of the Bar Exam—which suffers from multiple-choice bias and tests static knowledge retrieval—to isolate the specific cognitive mechanisms utilized in legal analysis. 

### Rule Extraction Versus Analogical Reasoning in Common Law

The Stanford CodeX initiative, alongside the collaborative development of LegalBench (encompassing 162 distinct legal reasoning tasks), provides a foundational empirical baseline for evaluating LLMs within common law jurisdictions [cite: 9, 17]. The empirical data reveals a stark stratification in model performance based on the specific nature of the legal reasoning required. 

Tasks classified under "rule extraction," "issue spotting," and "metadata recall" yield highly competitive performance from foundation models [cite: 9]. In these tasks, the model is required to parse a provided text, identify a governing parameter, and output the isolated variable. For example, evaluating whether a contract is governed by the Uniform Commercial Code (UCC) or common law yields high accuracy because the rule is explicitly provided in the prompt and requires basic semantic classification [cite: 18].

However, performance sharply degrades in tasks categorized under "Application" and "Conclusion"—the core components of the judicial syllogism that require integrating statutory language with highly nuanced, novel fact patterns [cite: 9]. Analogical reasoning, which demands the mapping of relational similarities between precedent and present circumstances, remains a substantial gap. When models are tasked with determining if a specific factual scenario entails the application of a previously stated rule, or contradicts it, they frequently rely on superficial pattern matching rather than performing genuine inference across facts and rules [cite: 9, 19]. Researchers have observed that LLMs often bypass genuine analogical reasoning entirely, generating fluent text that asserts an analogy exists without logically establishing the underlying relational similarity [cite: 19].

### Table 1: Synthesis of Empirical Legal Benchmark Performance

| Benchmark / Dataset | Legal Jurisdiction | Primary Methodology | Rule Extraction / Recall Performance | Analogical / Structural Reasoning Performance | Key Failure Modes Identified |
| :--- | :--- | :--- | :--- | :--- | :--- |
| **LegalBench** (Stanford CodeX) [cite: 9, 17] | US Common Law | 162 tasks encompassing contract analysis, statutory interpretation, and issue spotting. | **High:** Competitive accuracy on rule recall and basic issue identification. | **Low:** Sharp degradation on "Application" and "Conclusion" subtasks requiring factual integration. | Reliance on pattern matching; inability to integrate statutory language with novel fact patterns; shallow procedural understanding [cite: 9]. |
| **MSLR-Bench** [cite: 3, 4] | Chinese Civil Law | IRAC (Issue, Rule, Application, Conclusion) multi-step reasoning traces. | **Moderate/High:** Strong surface-level text generation (High LLM Score). | **Low:** Advanced models achieve only ~72% IRAC Recall; fails to align with expert logical traces. | Cognitive interference from structural prompts; fabrication of analytical content; superficial coherence masking logical gaps [cite: 4]. |
| **SCOTUS Prediction** [cite: 20, 21, 22, 23] | US Common Law (Appellate) | Case outcome classification and authorship attribution over historical data. | **High:** 91% accuracy on authorship; ~80% on outcome prediction based on historical datasets. | **N/A:** Task relies on probabilistic modeling of judge behavior, not formal syllogistic reasoning. | Over-reliance on ideological/semantic markers; limited causal interpretability (predicts the *who/what*, not the *why*) [cite: 22]. |
| **IslamicLegalBench** [cite: 6, 10, 24] | Pluralistic Islamic Law | 718 instances across 13 tasks (recall, *'illah* identification, *qiyās* application). | **Moderate:** Best models achieve ~68% on basic bibliographical and textual recall. | **Very Low:** Severe failure on analogical application (*qiyās*) and cross-school synthesis. | Extreme susceptibility to false premises (risky sycophancy); few-shot prompting fails to improve structural logic [cite: 10, 24]. |



## Deconstructing the Artificial Syllogism: IRAC, RAG, and Chain-of-Thought

To evaluate multi-step reasoning traces formally, researchers introduced the MSLR (Multi-Step Legal Reasoning) benchmark, grounded heavily in the IRAC (Issue, Rule, Application, Conclusion) framework utilizing a dataset of Chinese judicial decisions, specifically focusing on insider trading [cite: 3, 4, 25, 26]. The IRAC methodology is the foundational syllogism taught in legal education, demanding a highly structured, defensible progression from initial premise to definitive deduction [cite: 3, 4]. 

### The Metrics of Failure in Multi-Step Inference

Evaluations using MSLR highlight severe limitations in current frontier models, including deep reasoning systems specifically trained for extended chains of thought [cite: 4]. The benchmark utilizes a specific metric termed "IRAC Recall," which measures the precise alignment between the reasoning traces generated by the LLM and the gold-standard "IRAC-Traces" annotated by legal experts [cite: 4]. 

The findings expose a critical discrepancy in how LLM output is perceived versus its actual logical rigor. While models often achieve a high "LLM Score"—meaning an LLM-as-a-judge rates the output as logically coherent, well-formatted, and superficially rigorous (over 90% coherence in some tests)—their actual "IRAC Recall" rarely exceeds 72-75% [cite: 4]. Furthermore, automated annotation tasks tracking Field Completeness Rate (FCR) demonstrate that LLMs fail to exceed 80% coverage on fields that human experts deem essential for a complete legal argument [cite: 4].

[image delta #1, 0 bytes]

 

This indicates that the models are generating a *simulacrum* of legal reasoning. They produce text that sounds rhetorically authoritative but structurally fails to trace the authentic inferential path required to reach the conclusion [cite: 4, 16]. 

### Cognitive Interference from Structural Prompts

A striking finding from the MSLR evaluations is the phenomenon of "cognitive interference." Common practice in prompt engineering involves forcing models to adopt a specific reasoning structure, such as Human-Designed Chain-of-Thought (CoT) prompting that mandates the model to output distinct "Issue," "Rule," "Application," and "Conclusion" headers [cite: 4]. 

However, imposing these rigid external frameworks on certain advanced models causes "systematic performance degradation" [cite: 4]. Forcing these models to conform to human deductive architectures actually disrupts their intrinsic, probabilistic inference mechanisms, leading to the inappropriate application of legal provisions and the outright fabrication of analytical content [cite: 4]. Conversely, allowing the model to generate a Self-Initiated CoT often results in higher output quality, suggesting that LLM cognition operates on mathematical principles so fundamentally alien to human jurisprudence that human scaffolding actively impairs it [cite: 4].

### Retrieval Pathologies and Sycophancy

The recognition that LLMs lack intrinsic reasoning capabilities and suffer from factual hallucination has led to the widespread adoption of Retrieval-Augmented Generation (RAG). RAG systems attempt to mitigate hallucinations by grounding the model in external legal databases, essentially fetching the "major premise" (the statute or precedent) and injecting it directly into the context window for the model to analyze [cite: 15, 27, 28].

While RAG improves the accuracy of rule extraction, it does not cure the underlying deficit in analogical application or syllogistic reasoning. One critical pathology observed in RAG systems is the "lost-in-the-middle" effect. LLMs exhibit a primacy and recency bias, heavily weighting information at the beginning and end of a context window [cite: 29]. If a complex statute or conflicting precedent is injected into the middle of a long context window, the model frequently ignores it, causing the legal analysis to fail not due to a lack of data, but due to architectural position bias [cite: 29].

Furthermore, retrieving the correct statute does not guarantee the model will properly apply it to the minor premise (the facts) [cite: 6]. A prominent and dangerous error pattern in LLM legal reasoning is "Wrong Conclusion from False Premises" [cite: 10]. Because the model’s attention weights optimize for rhetorical fluency and user-alignment rather than objective truth, it frequently exhibits "sycophancy." If a user provides a factually incorrect or legally impossible premise, the model will often accept it without challenge and flawlessly generate a logically invalid argument built entirely upon the user's flawed foundation [cite: 10, 24]. False premise detection tests on IslamicLegalBench revealed risky sycophancy rates above 40%, demonstrating that models cannot natively recognize when an argument's foundational logic is legally unsound [cite: 10, 24].

## Mechanistic Interpretability: The Anatomy of an Artificial Syllogism

To understand *why* LLMs fail at analogical reasoning and structural alignment while succeeding at rule extraction, it is necessary to peer inside the "black box" of the neural network. The emerging field of mechanistic interpretability attempts to reverse-engineer Transformer architectures by analyzing internal activations, attention weights, and learned representations to understand exactly how inputs are transformed into outputs [cite: 30, 31, 32].

### Attention Weights and the "Queried-Rule Locating Head"

Legal reasoning fundamentally requires maintaining logical state across a sequence of steps—from premise, to rule, to application, to conclusion. Mechanistic analysis of LLMs performing logical deduction reveals that they do not natively process information in this sequential, syllogistic manner. Instead, Transformer models rely on attention mechanisms (specifically QK, or Query-Key, circuits) that compute softmax functions to determine which preceding tokens are most relevant to the current prediction [cite: 30, 33, 34]. This operation is less akin to human deduction and more analogous to a "fuzzy database lookup" [cite: 27]. 

Using techniques like activation patching and causal tracing, researchers have identified specific attention heads within LLMs that act as "queried-rule locating heads" (for example, localized to specific layers and heads, such as layer 19, head 11 in certain architectures) [cite: 33, 35, 36]. When presented with a legal scenario, the model's attention mechanism searches the input context (or its parametric memory) for syntactic patterns that correlate with the query [cite: 35]. The model then generates a response by elevating the probability of tokens associated with that retrieved rule. 

### Syntactic Brittleness Overrides Semantic Logic

This reliance on attention weights makes the reasoning process highly susceptible to superficial syntactic perturbations. Research from MIT demonstrates that LLMs often mistakenly link certain sentence patterns with specific topics during pre-training [cite: 36]. When the underlying syntax of a legal question is altered—even if the core semantic meaning and logical requirements remain perfectly identical—the model's "queried-rule locating head" frequently fails to activate properly [cite: 36]. 

This leads to catastrophic failure in reasoning tasks. The MIT researchers found that replacing words with synonyms or altering the part-of-speech structure caused the LLMs to fail at tasks they had previously answered correctly, proving that the model was not "understanding" the logic of the query, but merely recognizing a familiar syntactic template [cite: 36]. Consequently, the LLM is not executing a legal syllogism; it is engaging in predictive pattern completion based on syntactic similarity [cite: 13, 36]. This mechanistic reality perfectly explains why models excel at extracting rules (where syntax is clear and repetitive) but fail at analogical reasoning (which requires mapping abstract, semantic relationships regardless of syntax).

## Jurisdictional Divergence: The Common Law vs. Civil Law Processing Gap

The architectural reliance on pattern matching, attention weights, and rule extraction has profound implications for how LLMs operate across different global legal jurisdictions. Empirical comparisons between LLM performance on Chinese legal tasks and United States legal tasks reveal a systemic bias in how generative AI interacts with the law [cite: 7, 37].

### The Affinity for Codification in Civil Law

The Chinese legal system, rooted deeply in the civil law tradition, is primarily statutory and codified. Judges operating within this system are traditionally obliged to respect established statutory articles and deduce outcomes directly from the text of the code, rather than deriving principles from a fluid, evolving body of historical precedent [cite: 8, 38]. 

LLMs demonstrate a marked architectural affinity for civil law frameworks [cite: 7]. Tasks that require mapping a specific fact pattern directly to a discrete statutory code—such as the legal calculations, single-label classification (SLC), and statutory extraction tasks found in LawBench—align perfectly with the Transformer's "queried-rule locating" capabilities [cite: 7, 35, 38]. Because the rule is explicitly defined in text and requires less interpretive bridging, the model's attention mechanisms can easily lock onto the relevant tokens. 

Recognizing this alignment, the Shenzhen Intermediate Court in China has actively integrated an LLM into its Intelligent Adjudication System to assist judges with drafting reasoning across civil, commercial, and criminal cases [cite: 8]. The system leverages the LLM's capacity to extract information and generate highly structured documents based on statutory inputs, capitalizing on the civil law environment's prioritization of statutory extraction over the synthesis of conflicting precedents [cite: 8].

### The Struggle with Common Law Precedent

In stark contrast, the US and UK common law systems are inherently adversarial and precedent-driven. They require intense analogical reasoning to align novel social practices with existing, often uncodified, legal categories [cite: 7, 39]. In common law, the rule is often not explicitly stated in a single statute; rather, it is a *ratio decidendi* buried within the narrative of past judicial opinions.

The common law requires the bridging of "epistemic gaps" through metaphor, analogy, and contextual synthesis [cite: 39]. Because LLMs process law probabilistically rather than relationally, they struggle to weigh the gravity of differing precedents. In a common law task, if Precedent A and Precedent B offer conflicting analogies, an LLM cannot inherently determine which is structurally or morally superior; it defaults to whichever precedent has a stronger statistical representation in its training weights or context window [cite: 11, 27]. This statistical reliance renders its analogical reasoning brittle and arbitrary, explaining the performance degradation seen in complex LegalBench common law tasks [cite: 9].

### The Pluralistic Challenge: IslamicLegalBench

This limitation is further compounded in pluralistic legal traditions, such as Islamic law, which relies on a complex interplay of the Qur'an, Prophetic traditions (Sunnah), scholarly consensus (ijmā'), and analogical reasoning (qiyās) [cite: 6]. Evaluations on IslamicLegalBench reveal that models fail dramatically when required to perform *qiyās* [cite: 6, 10]. When tasked with identifying the underlying legal rationale ('illah) of a classical ruling and applying it to a modern scenario (e.g., analogizing classical trade prohibitions to modern digital assets), the models cannot bridge the temporal and semantic gap, defaulting to superficial pattern matching or extreme hallucination [cite: 10].

## Legal Realism and Predictive Accuracy: The SCOTUS Phenomenon

Given the severe deficiencies in analogical reasoning and structural logic, it appears paradoxical that LLMs demonstrate surprising efficacy in predicting appellate court outcomes. Out-of-sample testing on the United States Supreme Court (SCOTUS) Database reveals that LLMs can predict case outcomes with approximately 70% to 80% accuracy over centuries of data, and can attribute unsigned opinions to specific Justices with an astonishing 91% accuracy [cite: 21, 22, 23, 40]. 

This predictive success, however, does not contradict the models' failure in formal logic; rather, it underscores the exact nature of their statistical architecture. Supreme Court predictions—such as those utilizing Martin-Quinn scores—rely heavily on identifying the ideological alignment of the Justices, historical voting patterns, and specific semantic markers within the text [cite: 22, 40]. 

The success of LLMs in this arena suggests that the models excel at mapping the latent behavioral and political patterns of human judges. This aligns perfectly with the jurisprudential theory of *Legal Realism*, which posits that the law is not a formal system of deductions, but rather a reflection of the psychological and political realities of the judges deciding the cases. The LLM is highly adept at predicting human output based on statistical correlation and historical bias [cite: 22, 27]. It is predicting the *who* and the *what* based on ideological and semantic markers, not the *why* based on jurisprudential "correctness." Consequently, predictive accuracy is a measure of the model's ability to recognize human behavioral patterns, not its ability to execute formal legal reasoning [cite: 20, 22].

## The Positivist Machine: LLMs and Hart's Concept of Law

The observable behaviors of LLMs—their excellence at textual rule extraction, their predictive capabilities regarding human behavior, and their failure at structural and moral synthesis—offer a unique opportunity to map machine intelligence onto the foundational theories of jurisprudence. The architecture of the LLM is, by its very nature, the ultimate, albeit flawed, manifestation of legal positivism.

### Primary Rules and Statistical Memorization

In his seminal work *The Concept of Law*, H.L.A. Hart famously delineated the legal system as a union of primary and secondary rules [cite: 12, 41, 42]. Primary rules are rules of obligation; they directly govern behavior, proscribing actions like violence or theft, or prescribing the requirements for making a valid contract [cite: 42, 43]. 

LLMs are highly proficient at modeling primary rules. Because primary rules are frequently explicitly codified and heavily represented in training corpora, the model can easily ingest a contract, locate the primary rule (e.g., "payment is due in 30 days"), and flag a potential breach. The model operates by learning what "norm-concordant behavior looks like" from the massive linguistic observation of explicitly stated primary rules [cite: 12]. As long as the rule is textually explicit, the LLM's attention mechanism can retrieve and apply it [cite: 12, 44].

### The Failure of Secondary Rules

However, Hart argued that a primitive society relying solely on primary rules suffers from uncertainty, static rigidity, and inefficiency. To solve this, a mature legal system develops secondary rules, which confer power and determine the conditions under which other rules become valid [cite: 12, 41, 42]. The three core secondary rules are the *rule of recognition* (to determine what constitutes valid law), the *rule of change* (to alter laws), and the *rule of adjudication* (to empower institutions to resolve disputes) [cite: 41, 42, 43].

It is precisely at the juncture of secondary rules that LLM behavior catastrophically fails. 

1.  **The Rule of Recognition:** LLMs cannot independently verify the validity or authoritative status of a legal norm. If prompted with a repealed statute, a dissenting opinion, or a completely fabricated case (hallucination), the LLM treats it with the same mathematical weight as binding law [cite: 4, 10, 26]. The model possesses no internal "rule of recognition"; it cannot distinguish between valid legal authority and persuasive fiction beyond the parameters set by external retrieval systems (RAG). 
2.  **The Rule of Adjudication:** Adjudication requires institutional authority and the ability to definitively resolve conflicts between primary rules [cite: 41, 42]. LLMs lack agency, memory of their social role, and authoritative finality; they generate multiple, often conflicting, probabilistic outputs depending on minor variations in the prompt (prompting instability) [cite: 11, 13, 41]. 

Because secondary rules require an institutional and social reality to function—a collective enactment of personhood, authority, and community recognition—an LLM, which operates purely in the realm of syntax, cannot execute them. The failure modes of LLMs in legal settings (e.g., confidently hallucinating a nonexistent precedent) are essentially failures to comprehend the secondary rules that govern the recognition of valid law [cite: 13, 42, 45].

## The Epistemic Gap and the Incapacity for Dworkinian Interpretivism

If LLMs represent an incomplete and rigid form of legal positivism, they are entirely incapable of executing interpretivism. Critics of the push toward "generative interpretation"—the movement advocating for the use of LLMs to resolve real-world interpretive disputes in contracts, statutes, or constitutional law—argue that the technology is inherently biased toward a hyper-formalist methodology that completely ignores the deeper realities of adjudication [cite: 16].

### The Simulacrum of Interpretation

In the paper *Generative Misinterpretation*, legal scholars Grimmelmann, Sobel, and Stein argue that using LLMs for legal interpretation introduces a fatal "epistemic gap" [cite: 11, 16]. The epistemic gap asks a fundamental question: do these models actually measure linguistic meaning and legal intent, or do they merely measure statistical frequency?

The authors argue that the fluency of LLM-generated text is a deceptive facade—a "Potemkin interpretation" with nothing of substance behind it [cite: 11]. Traditional legal reasoning is considered legitimate precisely because a human judge engages in the hard, principled work of moving from source materials to a persuasive conclusion, taking institutional accountability for the outcome [cite: 11]. LLMs sever the vital connection between the *process* of legal reasoning and the *artifact* (the written opinion) [cite: 11].

[image delta #2, 0 bytes]

 An LLM can produce a plausible-looking, highly persuasive opinion through predictive pattern completion, a process fundamentally alien to human legal reasoning [cite: 11, 13]. Consequently, the superficial fluency of the text is stripped of any substantive guarantee of accuracy, justice, or methodical rigor. 



### The Incapacity for Dworkinian Principles

Ronald Dworkin’s theory of law directly challenged Hart's positivism by positing that the law does not consist solely of primary and secondary rules, but also of *principles*—moral, political, and equitable standards that possess a dimension of "weight" [cite: 14]. When rules conflict, or in "hard cases" where the statutory law is ambiguous or exhausted, judges must engage in constructive interpretation. They must balance competing principles to find the answer that best fits the institutional history and moral fabric of the community, acting as the metaphorical ideal judge, "Hercules."

LLMs are architecturally precluded from engaging in Dworkinian interpretivism. They possess no capacity to weigh moral principles or assess the social gravity and equitable dimensions of a legal decision [cite: 14, 15]. If a prompt introduces conflicting legal principles, the LLM does not resolve them through constructive interpretation or moral balancing; it resolves them through statistical averaging. It defaults to the most probable token continuation based on its training data, which often results in "multiple-choice bias" or extreme sensitivity to the phrasing of the prompt (prompting instability) [cite: 11, 15]. 

Proponents of LLM integration, particularly those aligned with strict textualist jurisprudence, occasionally view the technology as the "logical end result of the textualist project," believing it can analyze language with perfect mathematical objectivity and remove human bias from the bench [cite: 11]. However, this "mathematization of language" is a dangerous fallacy [cite: 11]. It ignores the reality that language is inherently ambiguous, context-dependent, and that legal texts were written by human beings operating within a specific social and normative context [cite: 14, 15]. By stripping the text of its human context and treating it as a pure mathematical variable within a high-dimensional vector space, LLMs enforce a rigid, unyielding pseudo-formalism that is entirely blind to the equitable and moral dimensions that legitimize the law.

### Table 2: Mapping Jurisprudential Concepts to LLM Behaviors and Failures

| Jurisprudential Theory | Core Concept | Observable LLM Behavior / Mechanism | Corresponding LLM Failure Mode |
| :--- | :--- | :--- | :--- |
| **Legal Positivism** (H.L.A. Hart) | **Primary Rules** (Rules dictating behavior/obligations) [cite: 12, 41, 42] | **High Proficiency:** LLMs easily memorize and extract codified obligations from input context via "queried-rule locating" attention heads [cite: 12, 35, 44]. | **Contextual Blindness:** Applies rules rigidly based on syntax, failing to recognize equitable exceptions or semantic nuance [cite: 36]. |
| **Legal Positivism** (H.L.A. Hart) | **Secondary Rules** (Rules of recognition, change, adjudication) [cite: 41, 42] | **Severe Deficiency:** LLMs lack institutional awareness and cannot independently validate the authority of a text [cite: 13, 45]. | **Hallucination / Sycophancy:** Inability to distinguish between valid binding precedent and fabricated or overruled law; acceptance of false premises [cite: 10, 24]. |
| **Interpretivism** (R. Dworkin) | **Constructive Interpretation** (Resolving ambiguity via community standards/morality) [cite: 14] | **Statistical Averaging:** LLMs resolve ambiguity by outputting the most statistically probable token, not by weighing moral gravity [cite: 14, 27]. | **Prompting Instability:** Outcomes flip drastically based on minor semantic phrasing changes; systemic failure in ambiguous "hard cases" [cite: 11]. |
| **Interpretivism** (R. Dworkin) | **Legal Principles vs. Rules** (Principles have a dimension of "weight") [cite: 14, 15] | **Mathematical Homogenization:** The model treats all text (statute, dicta, moral principle) as flat data points mapped in a vector space [cite: 11, 15]. | **Epistemic Gap:** Superficial fluency masking a total lack of normative reasoning; severance of process from artifact [cite: 11]. |
| **Textualism / Formalism** | **Objective Meaning** (Text contains inherent, mathematically discoverable meaning) [cite: 11, 15] | **Pattern Matching:** Operates exactly as a hyper-formalist machine, analyzing syntax over semantic intent [cite: 11, 36]. | **Syntactic Brittleness:** Reasoning collapses when grammatical structure is altered, even if semantic meaning is identical [cite: 36]. |

## Conclusion

The intersection of large language models and the legal domain provides a profound testing ground for understanding both the limits of artificial intelligence and the fundamental architecture of human legal reasoning. An exhaustive review of recent empirical data (2023–2026), coupled with insights from mechanistic interpretability, confirms that LLMs do not engage in genuine legal reasoning. While their proficiency in textual pattern matching allows them to excel at statutory rule extraction and appellate outcome prediction (validating Legal Realist assumptions regarding human behavioral prediction), they experience catastrophic failure when tasked with the structural, multi-step logic required by frameworks like IRAC, and the complex analogical reasoning inherent in common law and pluralistic systems.

Ultimately, the behavior of LLMs serves as a technological mirror reflecting the fundamental divides in legal philosophy. As perfect engines of shallow positivism, they can map Hart's primary rules with unprecedented speed and efficiency. Yet, their complete lack of normative grounding, institutional awareness (the failure of secondary rules), and capacity for Dworkinian constructive interpretation exposes a massive epistemic gap. The fluency of generative AI is a statistical artifact, wholly severed from the principled, contextual, and deeply human process of adjudication. Recognizing this strict boundary is essential to prevent the deployment of "Potemkin interpretation" in high-stakes legal environments, ensuring that generative technology serves strictly as an administrative tool for retrieving the law, rather than a normative engine trusted to reason through it.

**Sources:**
1. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGNlsd9ysRrgZtIOQkxmPS8kO4BZBbT7a8I5gCBIWmFwBpuKdeyjWts87GckJ8srm9AV9kCB_mHBgdVgfT4mbpRhrraj-xjDvuPqCXiUvY9mhB8SuVboIVIbz3rTlx6ufGGuCnLPg==)
2. [niklaus.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEStiKaEq9egOtQid-f0tu1ACbdXxbQPOl8-P0TF7uLGEYCOlAcy2Gf3caM7OgZKA2ZF4-EZePmG-23heR5NcklHv9B1vPGNQFH2RIiUnQWOsxZWdmRl-xAcxGUxN_2TYG6CY8X)
3. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGG8oisoHewM5FZZ-v7uBHo2WPVo3l9_x48w2ZyJUy2CxHM5CDnX0z0UmZttzELHk02FbO0c12YrZhr-ppfpkyLtq8u2dbA3NNYUswJi-m-YnYTeOXk85I8fyxqNouyZASVgO47gSslhlbu8SXZV2c0FOWr8z1tzpAv9uf8VWlKjU3L_RDaGSrkd4aEwoH3ihZWN0oa7EKqD_aLEdefaI4boLinERsjyqBgtYMFqkzSyp6Mr7O-Tqszzcg-zcw2CZack0omMM6plJH7ZuxcDlczq3XzceRIsWrSPxfBZCc=)
4. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQElbP1MtIbKlky7oIyggGwO3kPCaiuiVtWemJsI7hIS_XTr0Ijw582JWmD8_5kDX-n1zqsiWE6DjnRctjxKZXSNn3hnKaIf6jOhzwyTJwU32bSOYOf4mQ==)
5. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGtr5INsq1gmDlYMpbkNI0hEpX6p3ZVPpyKCsQC3Kni6AMODmby6Wy6rnZDwU8OvGdW7hmOOVNfEPKWDxbj9OKONCds4xBiZqRWQtRG9y2JfDJG7QsiAWs2WwMPXD2ccA==)
6. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEk6peK5tY8OuxK5OV9g5kM33KJA0ZYOI7iMmYS6cBABNz9GZllKyBasVBabXJQTqe3hVPLGcQoN8CT4gv2YkE-_2Ll9ZAHNyimb5aQQ_OYTyuAi7c1cWXz7g==)
7. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF_53TU-Zy20zb92NjiVnAxhWRraOygLwj3KX5EfI87J8j-xvcND4aWjjgz93lu3rWhgFBVCOkC44QqHBEFLRawz7J_VG3rE0xLcf_eIb9obhrKYxF-lMw=)
8. [oup.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEsyfYjjIUoHNCRjScYGG1UUb9khQ8WN4ps1uGflygR8X0ifbfJ38Fy7_d1H2NKrcn0e5W1MsNEcyIeFWmKWMJLs5uRClvkZ6GgfzaNGE3tm9Ys1FHBs0bOCnUP_w31dNgAVw_iPFLbDG1UrQ==)
9. [emergentmind.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQERNtZ_1jSktQByyrPMYxO5mWELZ9mnl0oDe9uqr4NQZZsFOe5s4EyJ9Umn7p9Hnvbt8rUgwiolMrU-KWSfRbVBZrO2BA8sZzPExazwfQpSYSappvwLxs0Hy5nJqBKQV2HF82UU)
10. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGPfIXeScoCnG06xdWFLTOr_9uG0X5tCporhsC7ndPwcmsiaiVegnLhfq7Mop5vL2ZGY0LEjPZ5kvPViSujpf-CrlBtYjQfBvwcn7myTLFIKCX-LUIcyg==)
11. [Link](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGNtGWjGe6d2EYTd6NeBasG5EFDrbjOzFtJ2P-uEFEDSa_XBspy69GAfJnDSGPgeoL7jD5fQIQAqwx-6EnLlrhTz6xTx2AE965QFLJ6ebA_dyJ_UmBCbjk-Hev169yCJFpgDM7mNaA7xyuYzihOfjdymPSy-59iJPZM6v9Zv6bppeb7rLPk5pMXwERuDdszXOYeFndnfDRH_MMvCci75i2Wy5kf0A==)
12. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEhIDPp7A_1ZygS_AwU8cQKqAMKU5cY2LmF3Ss6bUWit7Q5YXhZMO2WO9s5nAEo2thneew8w4iJd5szt_KTvC6VvFwg4w15nhpbCpjfqf7rOnK-fuzDLsWLBQ==)
13. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFBCjjvGngThhwJmna9-iRUWhu-Kk-4fkmdTpcdNKd7BOfzLRAwuUANfkNlHykYCx6zfv8lqzyNLjjTPcLqBvd8s3fjM3nb4XzRtmZdwJzdfbRoUYgvgpk3lg==)
14. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFReJW2cOlklC9FNnn9ZCDefXYSL6DnPN54l6ubRWwXaqdqoAbTzppHZafqbOlSWePFb0rPUuvU9CSSNy7fpjbOpD3aJJGbpf3GkVcC9A5OWk-uX68zNiLCLQ==)
15. [tum.de](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGUKbP3tnjGRlP1x-a1QNzwls_Iot6I_fqGZ-XRDcAOGNtCJj5t_68d3PMBv91SCfyDMrvGI-4MxZ-_WUwHiqiMbvLq9-sqBfFowAr1Fx4zxp_nso0-yYLgRtsrUsv10XW4ndlLGoKman9vVdg1x-4NvoJCf2LMCf4ciFutMLDJHmPJ9AkrOnt8rF0=)
16. [harvard.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFdy3RbTwxYrruYmxf4Aoli1xrlrrNLqUNEMHk8FDi2gG-N_deL09FNQDAOVK59-KljiIX6xMUZWBy9-rNxXOmumQXpVvJzuF8epcuaRKPfziNr2DbgfoYgxDpNIjgSS_7t-98s5lknpmhTNv2Ms5iSFvzJFUpO6HuwZCHbX7Q1BNWtsw==)
17. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEgjEjp1X4FfnR_mtNObLmDpyy5t_4VVwU_S1WHwxq51KJ0a3Yi4T13rtx4RjLaiUepAgjQdtEdVSfzIXQsieLgV-evO-iqFPburflT2RNMyI2hdWQzMmlNQi0uVQk1hn--20w_weMulElifZc4HYFBmD5Hg5tcq-g_sZ1UppLE3VpyNONJcqkstm54zRMhMUXqnysUrTnpgOvHmqeiqYdK-8tQhn585E8x4t9WyAycE197GJfvESpCT3GpdU9EutWZ7MAfieKije22detp)
18. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFbUZa-nc_snNFojZ1EGz7CKHCHPF52a-zdL-K2ERSDL_7vQb2V2RM5i1W5P7tID3yNiSD6AD_GvM1jIXx-BkAwGfrd9T5oQEdtEP9RsrqOs28E3itQ-g==)
19. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE2bMwfdRzWCanc0FYJH6IypjywsT7n3CaUy0RDwZv89aQKyDxaIXwKoDxlCShS2YPcYJplFR385se9E1DFXlLR8OhngQyn01LfBHoht6Ua8W6ySK4MAw==)
20. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEaNbvSUGlzUiuvJiz2kY2_KKidn8HsRyV83hthlnZx1P5EpeEY6VFMwGXuxBCKsd9dCnDXCn8bCKMQIe7URVSAojg7bsZpm3XASk1YMI6Et3Wc554RebnftA==)
21. [oup.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEHb7FLBfWbWInXlg_sCPY60ankSzC9K2Jc11_U3NR4f1PsBbnoberL9c3rJE6r2lqL9MLiH-B8Ila9rEqB8E4UzfwjDpJ1MTTUB-ImMc0WZOBp0WFCb7XcpBY6ffZ1N8f6N7e05xH50zY=)
22. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGnLcw07uINlCB4qTpBlH6QxhQOXLBcqBiUjshw1INRNIRNvhJppPhgNQ3DyHXtJSVHy9zZhW8pBG0DdgayWBzcxCZHhjmb5eQWF7sABnIHceAZh3DZ0VVv0Q01wmTTGQNfg638YEVBtvBE0WcDfgISK1jScXIl2iesVkkp3Zgpemfp4aIMI5r7hTbssjEbWyMG7TdNBCmxnYOBT_NYuKOu55otiGfaSIC0KqOVuZAZAUqES0AILH6M7vY=)
23. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG-OpeIBRTr9_nCQwLaX59efE7KO-Lxgcq6sD5K_5v_aCjqBOcXbi5tvk-Y0cf4CwkEpgxrqdouRKntuZuA0DCteXD85ZDHDKsCOdlODCRd6WPfKC4hfABAPakME0q6-bq5Me1Zpk10XgMLGUk=)
24. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFAIFxkwTl17bVAymS4fY7j_ytR_h6p1yqD3SToJ7yEMoR86k1pidnRIhUrpwyHUeAoFbh4xkwh1U4zGxKnYpKfzHkdrQkgWOG_X1VSnpuV1Dim90vHyzxJFqs0dHs0Vvam03R9SoWsOVK3oZVHuUuqRHNHEXk_VXEfFXOvF8aKt87-OZtCuRlCOg7xLxIEuScZ2qDey2Sf56XcX1IwWbiYehymtMIw5uQfXIqE_miVaoHjXB02xYTzP5-J6qGKqMeBLGnDNXUeO8kAgwlfrVeoI4mXQ_JtPlSny16Axy9fx_86cOD8FsYjwls=)
25. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH1ZL5-b0qhCD6BZvXXggIoj5hlDaKSuko6If3d2txMP7aRRVPahZShhFBPuF641bGNatbr-1C0JF4ZEUTCpUPclVNPGC6QKVwdfEJNtorrWtZH0XtR0fTZWh9SVumG)
26. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGXE0rdVA9ghvlBlt2i9B9QUrm4bwGVoLxQPpfGkiVKxeAn2Ipr1ZS270W9xMmgQUcZdALBBr1mdxjn2huJSUe5dduNpmDCEn0ZUf4eU3naOItsaY0knDdp7Q1GU3M8Cck6Pwgo6I6dpFxrERS7_QsWijkMj17npJ9I3T8YEZHsFu31JyxGzQG1g8BbQL-fl3O8fSuK78kddnVmKI9QMpc-eTPPzxT8i7go)
27. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH1DYQRz7NwUnyxR6fJydgnPOb2JQ3416yBumniFazB8Dh6OCVaaRnMnFFWkru6rj_DsrnyvSuoNhiZgWyw2t0ySt_8iriurIP4G3v9MpY0wkDBHvnl2VVtCAuvdRAzGU4xPB0OFMI4YJzn5Qrr4m66qA_eXIf-xOyZ4VnxmzDyrW6QxLxYeeCByCxJwRp72KZQlLHsf6Ff8Hcv965r7ZYoGj1S0eZ3qUBUjOQOLH8AQg==)
28. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE5pMs29QmXr4wZQdqdoiM9dyebbIRXrL5A6jYPYqAjoG7Zli14dITh7IkWTNKZbf7mcuRMwb6ul2AqkRuRj-xrXv4NRdBfKfbEce1ndJKHSHmQiGefxguwk1DGeSk_)
29. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGid6-6I9bu0jHqeDjRsT0Ctvwh_FSLnKg8PQlrrMIRxUShWVU71TeLBgdYX5HO6d0lUt_DmgpRAXNAKAEWKZNoK4wxNgEniHzlJueyPT6duvMEvg-ApjwLSl7OU6h9N80puwX8lWvftRUUoyGfP5HGnY3BybTn-kGnuc6fQjbaVw2YkDDf)
30. [preprints.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGFcm3HVYqS7542D6b4MFJD47mESpH67y4PMpkKyFQF0GqxgU4zbhTeS9unkjOqJ_KJUVGS7TNGD-L-p0DiRTDMySVWCAc0zyGenx-MDMm1RJLxVJi4xp5DkueA7vhm2qf7p3EtygZi-Qc=)
31. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEEjurqC-NIiFmQeeFg0Q1_guUW5yn-oBo4u1iZSELhlmhfDkldAPRF_H2CjzmoPVwrHtWfeW-xy2IwhE5MSCgItiFLbkUwm9Qc8tqvBLGYA91r8REI2Ck50UBpRgvuiua25axVW-U2kAgS3ABj2MP3QE5KItooMhfGYHCax1RN)
32. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHaA1RNTGQghtVKbbgOrg00AlShXmh2VwHUz9w7Se5W1dMPHWlZlL-rAZY5C-JNYF-5zghhMneH-yaFr3mTS4XpyExZ4k6ZdnE2wxdVOY9EM8jbnqBbUSr7PQ==)
33. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFMlR7ScUth-6CFEcHmiV94xK76Icl7crVBsjoEWCn6z3NW5AHyq8iwi6US8YMi22pR4sLfm_e0nrM6NlS7Y_U1rTLb1Bjo_4cfjePBVAxY2iCVmUmk0cpUtQ==)
34. [unibo.it](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFHpkfvb7e2KiZcKLmc19hkzIX0EULPYKLD_ecJTHeizhz8-fjtJCG5k_LUOUjj7ep_AyGxbie6f2UOzbvLydAEl5S_wGKuPIOg702XNLd2EEZ38lnyo7GMS2LYxeM5NzXRJ0EjopnA4VSWL19MEoHTpGzdRnhcq8FXnbDbSw==)
35. [neurips.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHk5e7zuHMNueub71SXRwJmK1or5UWnoTlKx88xNYtH8tqYN6PDKFcefbppiDzv_fpsDzCkHUCKxyjojp3XilgPEY9CTTKGQjhpxMHsfbCoD0vPpeDK1eigLn2fa55_F3dwWno=)
36. [mit.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEn3fqkXoOjEZb20P1yunqPx640R_BWkpzBGIMG6oFXU8mChBaG_4zutJq7hnuvIwLCq76g2jfWqXGm2IDf5rmljho4_HWM34qRjc1vzBHU2YlNGByWtD64G0CJHueEixxb5LNabBcjrOz2W2cYzTzwvuhCW3ZHPA3t0uwgTshKkxL_dgudcSk8a3GK1NpzHg==)
37. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHTwG60B1V-kjYWLhE36e2vyaZolMXnhJ17GbnVZcyuCwJwPq3kif7JG0K-EvYjDXfmayl-kuJJHbzrcsqCU5m5oLMVh3OzyiVUQr4ur_uOd9at09lRHNS6QIvv8N0CggZwJ-U_DK5PVZI0)
38. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHo1Na2yzU8DEj1IXZSN5WE0i9t6G_e3xtwQ7-p_rDYXAnq8GfCgSuMJncI71vhr5Vo6T8Dk7cw7JUO-M5s-5eUZSwueZRCCdHtnwPQQcSXQN6EuAF9bw==)
39. [scholaris.ca](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHvt5CXNsB2YnrrgKI2oj0EyQEDAlRARaEmmBKa1LkJUt8ClDk__7Bb6uq5POX9tudFRRPvkUMfhDYrElDFfhKkPDToRpkgcw8QSXwEYPAVvggi6erTiTm8T6PATAPcY67U1kxQivnJ9JU48JnijgSMZGT4FUTAzuHURR-JL13rA9US6dITklnTotK-eg==)
40. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEXL29Cet9rxta5TvBpxNoYdSw8slODBBOd-Zd18USKnLhZ4_3k64wwzhAdn7J6jGrCjEIAGMZ4bCwLRxfSYUmlPudFEh1YdiDZayfiSU1oDo4mae_T0k3O0IlR-0KTXCie1rv3ZIbYs-EChfwxnVeaGQ9D41saNAnq)
41. [iit.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHD0WunjlET8iwXYmTnhF3RNt7NAYP4Vt2ItBS91DM6oTrJ8dM-FOepI7nqyl8wUNkZ4rt8v_9oyWWsX23NOUT6kvtVdDJZHMaL8so7-Ky2a5x6vMhYHXuJN9okN_9Ute9DNhEHZt74R8F9MUvZYcTwjHfBzWVdQh86TvE_3tQm4pce_DTQyQyWfuakp5tz)
42. [brandeis.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG4aBpz9cZqFdWZHJseoWkRUxT7Zo-lQbokDpIhUIjN5ot4vtqUQgPX484mMtpG_JVW9WoYe6zbPNUHYvXXVL8GVZfFiweGOo70dOjbBDCAfhU-yJaSpYFoHp1ngK9I4ej-tYa81irouSwE9EZJ3YcWjjfnig==)
43. [scribd.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFpmru97rmNMv7Ltm2Et1YxG12gcIntN_Ee1TQXiI4bYgQMnwRAfyip-_BDSkag5Vcomo6HVpqltOoY3fWqOQSBDsPqrCi100ekKxlGNuwjZCwO0h1dcBeT0oKAFYmbiZuY7bKA1ZxvaP9cI_n2SOcr7z-u-jXYWZUcI79Ri7ACK_shamj935duC2t_oDLXy1IOpyHTbbXJvZ8qiTrO2A==)
44. [cecan.ac.uk](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEnL9pvqECVoOaDVyq6XdaQpOUvB2swP_CHre-R93AmP57Kcq1wH4YKQ5kBwB08LLsPubqUYhrgwgp3lsZ47OoCrwVBTYKSlqLb03VPZhDzmKe0HtCBZF_D1LYisDIx8XGwtT-upGiPnFSr0mQum0Q9uA1BQTyuYaHVtwQ5Fdbr4thOV1Eqpd2Wu18GtNyVeA50VPUA6bvIyJDirnGVcSZfP0LRyQ_NHFs=)
45. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEsJRrnAYHO72TmSAtgggCa8n6zohxcFcBNk4JKpv5Gz5uj1FQXCuArU82X8KpTuqDHkbl8PryGJE4K4ndM8cLryuIqOH922yqrj-7WNaXZuEXEWkkwHNLBZA==)