[Legal Services] When "Hallucinated Precedents" Cross the Malpractice Line: Source-Grounded Retrieval, Approved Computer Actions, and Verifiable Legal Evidence Memos

Why Plausible-Looking Draft Citations Cross the Malpractice Line

Imagine a litigation team preparing a pre-trial brief under tight deadlines. A junior associate uses a generative text assistant to synthesize case law on a procedural dispute. The system returns a polished draft memorandum complete with case names, docket numbers, court jurisdictions, and seemingly authoritative holdings. Yet when the supervising partner checks those citations against official court reporters and statutory databases, several pivotal decisions do not exist at all, while others hold the opposite of what the draft asserts. Submitting unverified, fabricated authority to a court or client crosses a bright line of professional ethics, candor to the tribunal, and malpractice liability.

Replacing a general-purpose chatbot with a specialized legal search product does not eliminate this risk. Empirical research shows that even commercial AI legal research tools using retrieval-augmented generation (RAG) still fabricate legal facts and citations; a legal workbench must bind every sentence to verifiable source texts and require item-by-item attorney review and sign-off, without issuing automated legal opinions.

From General Models to Proprietary Legal RAG: Empirical Evidence on Hallucinations

Large language models are increasingly adopted across legal practice, education, and research, yet their potential is threatened by “hallucinations”—textual output inconsistent with legal facts^[1]. In the first systematic evaluation of legal hallucinations across public-facing models, Dahl, Magesh, Suzgun, and Ho (2024) documented trends across jurisdictions, courts, time periods, and cases^[1]. They show that public-facing models hallucinate at least 58% of the time, struggle to predict their own hallucinations, and often uncritically accept users’ incorrect legal assumptions^[1]. Proposing a typology of legal hallucinations, the authors caution against rapid, unsupervised integration of popular models into legal tasks^[1].

Certain legal research providers have touted retrieval-augmented generation (RAG) as “eliminating” or “avoid[ing]” hallucinations, or guaranteeing “hallucination-free” legal citations^[2]. Because proprietary systems are closed, systematically assessing these claims is challenging^[2]. Magesh, Surani, Dahl, Suzgun, Manning, and Ho (2025) designed and reported on the first preregistered empirical evaluation of AI-driven legal research tools, assessing RAG-based proprietary systems^[2]. They demonstrate that providers’ claims are overstated: while hallucinations are reduced relative to general-purpose chatbots, leading commercial AI research tools from major legal information providers still hallucinate between 17% and 33% of the time, with substantial differences across systems in responsiveness and accuracy^[2]. Alongside a preregistered dataset and a typology differentiating hallucinations from accurate legal responses, the study provides evidence to inform legal professionals’ responsibilities in supervising and verifying AI outputs—a central open question for integrating AI into law responsibly^[2].

Why does retrieval augmentation alone fail to guarantee reliable legal reasoning? Guha et al. (2023) introduced LegalBench, a collaboratively constructed benchmark built with legal professionals^[3]. LegalBench consists of 162 tasks covering six different types of legal reasoning^[3], showing how popular legal frameworks correspond to benchmark tasks and giving lawyers and developers a common vocabulary^[3]. Evaluating 20 open-source and commercial models^[3] illustrates how varied legal reasoning is. Together, these studies show that neither general chatbots nor proprietary RAG tools can operate as autonomous authorities; source-passage anchoring and lawyer verification are mandatory architectural constraints.

Source Grounding, Approved Actions, and Evidence Memos: Mapping the Legal Workbench

Translating these findings into software architecture requires replacing open-ended chat with structured workflows. A workbench for legal services must decompose research and case preparation into scoped retrieval, adversarial evidence comparison, controlled desktop operations, and human sign-off gates. Within rlab.works, this design maps onto skill packs across the AI Gateway, Intelligence, and Agent OS modules.

At the clause lookup layer, Connect information search (S17) in the AI Gateway module starts from an explicit input contract: Question, authorized sources and search scope. It collects permitted material with source attribution, builds ranked retrieval candidates from authorized passages, and prepares a summary draft with passage citations before halting at a human review gate—Check the sources. A reviewer confirms the summary and evidence package, while rejected output returns for revision and does not advance. In legal practice, this connects an authorized contract repository to clause search while preserving original passage locations so legal staff can inspect sources line by line.

For full issue briefings, skill packs in the Intelligence module govern drafting and argument analysis. First, [Research and write a report (S24)](/skills/sources-to-report/) takes a *Research question, permitted sources and report outline*, plans the search scope, collects sectioned material with a source index, and organizes a cited draft report before pausing at the *Review the evidence* human gate. Only after legal verification does it format and export the report. Underlying this step, [evidence retrieval, reranking and answer synthesis](/capabilities/evidence-search-answers/) fetches selected source pages within time and size budgets and synthesizes cited answers. Second, because models often accept users' incorrect legal assumptions^[1], [Keep an evidence memo (S25)](/skills/research-evidence-memo/) structures dispute analysis around *Research entities, claims and existing evidence*. It collects dated source material, resolves entity relations and evidence indices, and uses investment thesis evidence and verdict tracking capabilities to record supporting, opposing, and uncertain points side by side, saving a traceable memo only after the *Review and save the memo* human gate.

For local docket organization, Approve computer actions (S03) in the Agent OS module takes an Action goal, account and permitted materials, checks readiness, and pauses at Approve the action list. Once approved, it acts via native computer control with permission gates, retaining screenshots, file diffs, and execution receipts before a final Check the changes human review gate.

Architectural honesty requires stating clear boundaries. In the rlab.works catalog, legal services has no skill packs rated as source-supported industry; all packs discussed here are designated for transferable use. Every skill pack is a proposed composition demonstrated with synthetic data and not yet connected to live services, and component maturity reflects static source inspection rather than production deployment. Adapting them to legal practice requires domain engineering:

• Connect information search (S17): wire in document access permissions and clause locations; summary outputs must never stand in as legal opinions.
• Research and write a report (S24): replace academic paper connectors and news topic models, and verify statutory validity and case applicability; the workflow never makes automated legal conclusions.
• Keep an evidence memo (S25): replace disclosure, market-series, and company models entirely with case files, statutes, parties, and timelines; original scores cannot be transferred as legal judgments.
• Approve computer actions (S03): restrict readable and modifiable directories item by item; external submissions remain outside default permissions.

Pre-Pilot Checks for Managing Partners and General Counsel

Before funding a legal AI pilot, a managing partner or General Counsel can apply concrete checks to ensure compliance with professional responsibility:

• Are claims anchored to exact locations in authorized primary sources? Confirm that search is restricted to approved statutory and case databases or permissioned contract stores, and that every drafted sentence links to an inspectable passage.
• Does the workbench place supporting and contrary authority side by side? Because models readily accept incorrect legal assumptions^[1], require explicit separation of supporting precedents, adverse authority, and unresolved facts.
• Do human gates enforce attorney sign-off and block external filing? Ensure unapproved summaries, memos, and file edits cannot advance without human review, and keep external submissions disabled by default.