CS Paper Radar

LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context

Anastasiia Kucherenko, Francois Brouchoud, Dimitri Percia David, Andrei Kucharavy

cs.AI ·submitted 2026-10-19 11:38 UTC · v1

Abstract

Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation. Following standard prompting guidance, a user naturally states their role (``I am a lawyer...''). We show that this innocuous habit has an unexpected cost: such authority framing can make the model selectively refuse legitimate requests, introducing bias through uneven access to assistance. We evaluate four modern small open-weight LLMs on benign legal-domain prompts to measure this over-refusal. Surprisingly, we find that authority-style prefixes (''I am legal counsel working for a national supreme court'', ''I'm a defense lawyer'') systematically increase refusal rates by 2--20x over the no-prefix baseline, while a known role-play jailbreak prefix shows mixed effects, sharply increasing refusals in some models and barely shifting them in others. The finding suggests that small on-prem deployable LLMs are unstable under contextual framings that a real institutional user might naturally introduce, and further investigation is essential to minimize opportunities for bias.

Closest papers

Nearest neighbours by meaning across the whole index.

HireWatch: Evaluating LLM Compliance with U.S. Employment Law Under Contextual Pressure

cs.CL cs.AIcs.CY ·2026-11-16 ·Huy Nghiem, Maria Kostylew, Noam Kolt
publishedcode/data similarity 0.61
Abstract · NeurIPS 2026 Evaluation and Dataset Track. Law and AI

The deployment of large language models (LLMs) in hiring pipelines is growing, yet existing evaluations generally overlook their compliance with legal rules. To fill this gap, we introduce HireWatch, an evaluation that assesses whether LLMs comply with U.S. laws that govern the questions that job candidates may be asked. We construct a 200-item bank of lawful and unlawful interview questions, whose legal classification is based on U.S. federal statutes, regulations, and agency guidance. We instruct LLMs to select questions from this bank in preparing for job interviews, across a set of various occupations and contexts, and report the fraction of unlawful questions that LLMs select. Evaluation of 18 closed and open-weight LLMs across ~100,000 trials under different types of contextual pressure reveals that LLMs' legal compliance is brittle. While LLMs generally avoid selecting unlawful questions in low-pressure contexts, legal violations increase sharply under higher pressure. Explicit prompts to comply with law can improve compliance rates, but do not altogether eliminate legal violations. We publicly release our code and data at https://github.com/hnghiem-nlp/hirewatch_repo.

0.41

Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts

cs.AI cs.CR ·2026-09-01 ·Rui Yang, Yang Hong, Yichao Xu, Zhengyu Liu, Yinzhi Cao +1 more
benchmarkhot direction similarity 0.61
Abstract · 12 pages, 5 figures

Cybersecurity assistance is dual use, requiring decisions about when to answer, limit, or refuse a request. Prior work studies refusal evaluation and conversational jailbreaking, but how conversational histories shift assistance boundaries in both directions for identical cybersecurity requests remains underexplored. We introduce 3R-Bench, a benchmark of 150 real-world cybersecurity requests evaluated across eight LLMs under direct prompting and two predefined conversational settings. We separately code refusal, bounded assistance, compliance, and provider blocking to characterize delivered assistance. On 400 matched ambiguous request-model pairs, compliance is 75.3% directly, 57.8% after refused precedent, and 81.0% after accepted precedent. A smaller repeated-generation panel reproduces the refusal-acceptance gap. On a balanced 75-request panel reusing the direct baseline, compliance decreases from 74.2% directly to 57.2% after neutral task-decomposition dialogue and 60.3% after failure-ended dialogue. These findings show that cybersecurity assistance varies with conversational history and model deployment, motivating evaluation beyond direct-request refusal rates.

0.60

Auditing Chatbots for Incursion into Professional Legal Practice

cs.CY ·2026-11-13 ·Nel Escher, Megan Elliott, Isabelle Qian, Yana Sharifullina, Svenja Guhr +2 more
benchmarkposition similarity 0.60
Abstract

The provision of legal advice has long been in the exclusive jurisdiction of licensed attorneys-a boundary that unauthorized practice of law (UPL) doctrine maintains to protect the public from incompetent counsel. We argue that this boundary is being redrawn, not by courts or bar authorities, but by frontier AI companies. While early chat models were trained to avoid dispensing legal advice, modern models have been reconfigured to reduce refusals and prioritize helpfulness. We ask whether chatbots increasingly do tasks that require a law degree and bar admission, and whether this is harmful in high-stakes legal domains. We develop UPLBench, a benchmark derived from court decisions finding UPL in high-stakes legal domains (e.g., immigration, family law). We then audit 23 chatbot models dating from 2023 to present, evaluating whether chatbots refuse or provide substantive legal advice, and whether they refer users to lawyers or legal aid. Unlike prior work that aims to establish the legal capabilities of chatbots (e.g., bar exam performance), our approach converts legal doctrine into a testable benchmark for evaluating chatbot adherence to law. We find modern chatbots increasingly perform tasks courts reserve for lawyers and decreasingly recommend that users consult lawyers. Subject matter expert attorneys reviewing our transcripts identified interaction patterns, such as failure to elicit relevant facts, that would be considered inadequate counsel. If acted upon, such advice could lead to deportations, loss of homes, and family separations-the very harms that UPL rules are designed to protect against. We weigh paths forward to protect the public, including compelling models to comply with existing law. We call upon the legal profession to respond to this incursion into its domain of expertise as there are insufficient protections and remedies for users who rely on chatbot legal advice.

0.26

Before Retrieval Begins: Safety and Tool-Use Decisions in LLM Agents

cs.CL ·2026-10-16 ·Aditya Nawal, Manit Baser, Mohan Gurusamy
hot direction similarity 0.58
Abstract · 19 pages, 5 figures. Preprint version of a manuscript submitted to Transactions of the Association for Computational Linguistics (TACL)

Large language models increasingly act as agents that call external tools, yet most tool-safety research examines only what follows a tool call. The prior decision of whether to call the tool remains largely unexamined and can influence downstream safety. Across the seven open-weight models, the unweighted model-average harm rate after admitted harmful requests is 74.5%. To identify what governs this decision, we apply a factorial intervention to the input request. Within our generated construction, neutralizing the requested-operation component produces the largest admission contrast (19 pp). A linear direction estimated from this contrast influences held-out admission in four of seven models; effects on operation-neutralized requests in some models show that it is not an operation-specific mechanism. We then test whether written refusal is behaviorally and representationally coupled to the same decision. Written refusal accompanies 33.6% of admitted harmful calls. Plain-text refusal training reduces unsafe admission in models that retain benign tool use, showing that response-level safety training can transfer to the call decision. Together, these results establish a partial, model-dependent relation between written refusal and tool admission. They motivate direct evaluation and supervision of tool admission rather than treating text refusal as its proxy.

0.29

PrivacyEvalClaw: A Tool-Augmented Deliberative Agent for Auditing Privacy Leakage in LLM Jailbreak Responses

cs.CR ·2026-10-27 ·Krishna Kanth Nakka, Yujun Wang, Shijie Zhou, Tu Nguyen, Gabriele Gelardi +1 more
published similarity 0.58
Abstract · Accepted to AACL-IJCNLP 2026

Reliable evaluation is a critical bottleneck in privacy red-teaming of large language models (LLMs): determining whether a jailbreak response actually reveals personal information can be difficult when disclosures are indirect, contextual, partial, or obfuscated. Existing evaluation approaches increasingly rely on guard models and LLM-as-Judges, yet these evaluators may interpret the same response differently. To better characterize this evaluation setting, we conduct a systematic study of privacy evaluators and find substantial disagreement, particularly in ambiguous cases that require additional evidence or targeted analysis to resolve. Motivated by these findings, we introduce \textsc{PrivacyEvalClaw}, a training-free deliberative agent for evaluating privacy leakage in LLM jailbreak responses. Rather than relying on isolated single-pass judgments, PrivacyEvalClaw treats evaluators and analysis modules as tools, adaptively gathers evidence, investigates uncertain disclosures, and reconciles conflicting signals before producing a final judgment. Our primary experiments use pairwise LLM-as-Judge evaluation of complete trajectories, showing that PrivacyEvalClaw achieves favorable win rates against competing baselines across outcome, process, and overall quality, while producing explicit reasoning traces that make the evaluation process more transparent and auditable.

0.21

Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?

cs.CY cs.AIcs.CL ·2026-08-10 ·Lorenzo Molfetta, Alessio Cocchieri, Luca Ragazzi, Ilaria Bartolini, Marco Patella +1 more
top venuebenchmarkhighly cited similarity 0.57
Abstract · Please cite the definitive, peer-reviewed version of this article published in the Proceedings of the 64th Annual Meeting of the Association…

In medicine, claims remain valid when supported by empirical evidence grounded in stable biological reality. In law, by contrast, truth is contingent, defined by jurisdiction, temporal validity, and the hierarchy of authoritative sources. The recent success of large language models (LLMs) on medical licensing examinations has encouraged an expectation of comparable legal competence. This analogy, however, obscures a critical distinction between domains. Unlike in medicine, legal performance often depends less on inference than on determining when external authority is applicable, valid, and non-contradictory. We introduce a comparative diagnostic framework evaluating legal reasoning against medical baselines along four axes (knowledge recall, grounding, confidence, and robustness), uncovering a sharp domain asymmetry when applied to a new benchmark that encodes temporal validity and normative relationships. While medical LLMs reliably benefit from verified sources, legal LLMs struggle to assess when retrieved citations are useful or misleading, exhibiting overconfidence in perturbed contexts and sensitivity to superficial formatting cues. Increased model scale amplifies this tendency, revealing that stronger instruction following can coincide with weaker resistance to authoritative perturbations. These findings show that LLMs treat law as unstructured text rather than binding precedent, while revealing a tendency to over-trust authoritative but false information when external references conflict with a model's internal knowledge.

0.90

Credible, Not Always Correct: How Reddit Users Verify AI-Generated Legal Advice

cs.CY ·2026-08-13 ·Rebecca Owens, Yusuf Mücahit Çetinkaya, Stergios Aidinlis, Dhyey Mehta, Tuğrulcan Elmas
similarity 0.56
Abstract

Large language models (LLMs) are increasingly used by laypeople to resolve real legal problems, against a backdrop of persistent access-to-justice deficits. This article presents evidence that the practical force of AI-generated legal advice depends not on its accuracy but on the social production of its credibility. While existing research has assessed the accuracy of legal AI, less is known about how machine-generated guidance is verified and made credible enough for lay users to act on. Drawing on a dual-method analysis of 153 Reddit narratives and 5,341 community reactions, this article maps a spectrum of verification practices. At one end, a minority of users verify AI-generated legal advice by triangulating across models, and some submit AI-generated guidance to platform communities for evaluation before acting, a configuration we term distributed counsel. Far more commonly, however, narratives are silent on verification. AI-generated legal advice is acted on the strength of its lawyer-like form and emotional reassurance alone. These findings show that AI-assisted legal self-help operates within an emerging informal infrastructure which redistributes the work of verification to those least equipped to bear it.

0.04

Privacy Conductor: An Architectural Enforcement of Contextual Integrity in LLM-Based Document Retrieval

cs.CR ·2026-10-11 ·Ryan Thornton, Mir Mehedi A. Pritom, Maanak Gupta, Jessee Roberts
similarity 0.56
Abstract

Large language model (LLM) that retrieve internal documents within an organization must decide who may receive them. Existing approaches based on Contextual Integrity ask the model to make that decision, so access depends on probabilistic reasoning. We present Privacy Conductor, an architecture that removes the decision from the model: a deterministic gate enforces a policy written by an authority, and the model receives only what the requester may already see. On a synthetic clinical corpus of 2,520 requester--document flows, the gate reproduces its policy exactly, so the remaining risk lies in the document labels it reads. As label source is of primary importance, we compare LLM and human generated labels. Across seven configurations of six models from two model families, LLM-written labels degrade the policy in every case. Models assign a document's sensitivity tier correctly on 78--86% of documents, but reproduce its access-control list on as few as 23%. Relabeling the same documents five times with an identical request changes 13 to 124 access decisions per configuration, a result that does not depend on our own labels. On 214 flows from HIPAA court cases, the most accurate configuration (87.8%) still permits 11 of the 20 disclosures the law prohibits. LLMs can describe what a document contains, but not decide who should receive it. Hence, we recommend privacy policy should be authored by accountable personnel and enforced outside the model.

0.16