Reimagining Legal Fact Verification with GenAI: Toward Effective Human-AI Collaboration

Sirui Han, Yuyao Zhang, Yidan Huang, Xueyan Li, Chengzhong Liu, Yike Guo

TL;DR

While Generative AI (GenAI) excels at low-stakes tasks like drafting and language optimization, its adoption in high-stakes legal fact verification is severely constrained by strict professional accountability, accuracy demands, and confidentiality risks. By interviewing 18 non-litigation lawyers, this paper maps out current legal workflows and outlines core design requirements for future AI systems, arguing that GenAI must act as a transparent, auditable "infrastructural intermediary" rather than an autonomous agent.

What Problem Does This Paper Address?

In non-litigation legal practice—such as corporate transactions, compliance, and due diligence—fact verification is a high-stakes, deeply interpretive process. A single oversight, like missing an undisclosed beneficial owner, can result in severe financial and reputational penalties. Unlike routine tasks that can be easily automated, fact verification is a dynamic sensemaking activity where lawyers must synthesize voluminous, often contradictory information from fragmented public registries and client disclosures.

While current Large Language Models (LLMs) show promise in understanding legal text, integrating them into fact verification introduces critical Human-Computer Interaction (HCI) challenges. Legal practice operates under a zero-tolerance threshold for error and strict liability. If an AI hallucinates a reference or subtly alters the meaning of a clause, the human lawyer remains entirely liable. This paper addresses the gap between GenAI's current capabilities and the rigorous epistemic, ethical, and legal structures required by legal professionals.

Methods

The authors conducted a qualitative empirical study consisting of semi-structured interviews with 18 experienced non-litigation lawyers across various practice areas (e.g., M&A, IPOs, corporate compliance).

The participants represented a wide range of experience levels (from 1 to 10 years) and varying degrees of prior GenAI familiarity to capture a comprehensive view of current practices and expectations. The research team analyzed the interview transcripts using inductive thematic analysis following Braun and Clarke’s six-phase framework to identify recurring challenges, workflows, and design desires.

Datasets

Not discussed in the paper (this is a qualitative interview-based study rather than a machine learning benchmark paper).

Key Insights and Ideas

  • AI as an Additional Object of Verification: Instead of simply accelerating workflows, GenAI often introduces a "backtracking" burden. Because AI-generated text can look highly confident and logically structured even when incorrect, lawyers must spend significant time tracing every automated citation back to its original source, sometimes taking longer than manual verification.
  • Epistemic Ambiguity: AI-generated summaries occupy an unstable status—they are neither raw data nor fully validated legal facts. There is a systemic risk that early-stage AI interpretations can algorithmically amplify specific connections while obscuring others, inadvertently hijacking a lawyer's strategic framing of a case.
  • The Erosion of Junior Expertise: Junior lawyers traditionally develop professional judgment and "evidentiary sensitivity" by performing the labor-intensive grunt work of reviewing documents and cross-checking registries. Automating these early-stage tasks risks removing the very conditions under which legal intuition is cultivated.
  • Infrastructure Over Autonomy: Lawyers do not want autonomous AI agents to make final decisions. Instead, they envision a "human-in-the-loop" infrastructure that handles data extraction, standardization, and timestamp alignment while leaving the interpretive, judgment-dependent evaluation entirely to humans.

Key Results

The study maps out a common three-stage legal fact verification workflow: the Preparation Stage, Fact Verification, and the generation of the Legal Deliverable.

Figure 1: Legal Fact Verification Workflow with GenAI. Legal fact verification operates as a dynamic sensemaking activity where humans combine retrieval, comparison, evaluation, and synthesis to construct a coherent, defensible narrative of the matter, while GenAI provides targeted support. It offers cognitive aid for client communication and legal framework identification in the preparation stage, assists with preparatory document work during fact verification, and contributes to generating due diligence reports, transaction documents and legal opinion letters for legal deliverables.

Qualitative analysis revealed two primary structural barriers to GenAI adoption:

  1. Accountability Frameworks (15/18 lawyers): The threshold for acceptable error is near zero. Subtle errors embedded in coherent text magnify professional risks.
  2. Information Control and Privacy (14/18 lawyers): Fact verification involves highly sensitive corporate and personal data. The perceived opacity of external GenAI platforms creates significant data leakage anxieties, meaning client consent frequently dictates whether AI can even be used.

To overcome these barriers, the authors translate these findings into three core interaction design requirements for future legal tech:

  • Traceability: Every AI output must be explicitly linked to an inspectable source (e.g., screenshots or direct links to official registries).
  • Layered Reasoning: AI explanations should mirror legal reasoning structures (e.g., navigable mind maps or issue trees showing "legal issues-corresponding situations-legal basis").
  • Specialization: Models must be fine-tuned to respect legal epistemic norms, such as statutory currency and authority hierarchies, rather than just optimizing for textual brevity or linguistic coherence.

Limitations and Future Work

The authors note that their study relies on self-reported interview data, which may occasionally diverge from actual, fast-paced daily practices. Additionally, the scope is strictly limited to non-litigation legal work.

For future work, the authors suggest conducting observational field studies to see how human-AI verification unfolds in real-time. They also recommend expanding research into litigation practices and conducting cross-domain comparative analyses with other high-stakes fields, such as healthcare or financial AI.

Summary updated

Abstract

Fact verification is a critical yet underexplored component of non-litigation legal practice. While existing research has examined automation in legal workflow and human-AI collaboration in high-stakes domains, little is known about how GenAI can support fact verification, a task that demands prudent judgment and strict accountability. To address this, we conducted semi-structured interviews with 18 lawyers to understand their current verification practices, attitudes toward GenAI adoption, and expectations for future systems. We found that while lawyers use GenAI for low-risk tasks like drafting and language optimization, concerns over accuracy, confidentiality, and liability are currently limiting its adoption for fact verification. These concerns translate into core design requirements for AI systems that are trustworthy and accountable. Based on these, we contribute design insights for human-AI collaboration in legal fact verification, emphasizing the development of auditable systems that balance efficiency with professional judgment and uphold ethical and legal accountability in high-stakes practice.
BibTeX
@inproceedings{han2026reimagining,
  author = {Sirui Han and Yuyao Zhang and Yidan Huang and Xueyan Li and Chengzhong Liu and Yike Guo},
  title = {{Reimagining Legal Fact Verification with GenAI: Toward Effective Human-AI Collaboration}},
  booktitle = {CHI},
  year = {2026},
  doi = {10.1145/3772318.3791088},
}
Figure