Legal AI's trust problem runs deeper than hallucinated cases
General-purpose AI and legal workflow platforms are competing to become the primary interface through which lawyers work, but substantive legal research still demands a different layer: specialist intelligence grounded in authoritative sources, domain-specific retrieval and auditable legal reasoning. The legal AI market is converging as providers from both directions race to occupy the same ground.
The most dangerous AI-generated answer in legal research is not the obviously wrong one. It is the one that is almost correct. Hallucinated authorities remain AI's most notorious failure mode in legal contexts, with reported incidents so prevalent that dedicated repositories now catalog them, including Damien Charlotin's AI Hallucination Cases database.
More insidious are AI's unexpected failure modes. A system may produce a coherent summary of relevant cases yet fail at the simpler task of quoting a judgment verbatim. Lawyers reviewing work delegated to junior colleagues know where to look for familiar risks, such as incomplete analysis or weak reasoning. AI's failures do not necessarily reflect the difficulty of a task and may vary from one output to the next: it can err in sophisticated analysis on one occasion and in basic transcription on another.
This unpredictability creates a difficult trade-off. Practitioners must either accept a broadly correct output that may conceal consequential errors or verify every proposition and quotation line by line, eroding the efficiency AI is meant to deliver.
Two layers, not three categories
Until recently, the legal AI market could be described in three categories: general-purpose systems such as Claude, ChatGPT and Microsoft Copilot; legal AI platforms such as Harvey and Legora; and specialist legal AI built for a defined field or jurisdiction. That model is becoming obsolete as general-purpose AI providers move directly into legal work.
Anthropic has introduced legal plugins and connectors for Claude; Perplexity has launched Computer for Counsel; and Google has introduced Gemini Enterprise for Legal. Meanwhile, legal-native platforms are broadening into end-to-end environments for research, drafting, review and collaboration. The boundary between general-purpose systems and legal AI platforms is blurring as both compete to become the primary interface through which lawyers work.
When assessing AI systems for legal research, it is more useful to distinguish between two layers. General AI, including legal-native platforms, provides the interface, broad reasoning capabilities and workflow environment. Specialist AI provides authoritative domain data, specialist retrieval, methods adapted to the field, auditable citations, guardrails and quality assurance.
Why specialist retrieval matters
Access to legal authorities is only the starting point. A reliable answer must identify applicable sources, distinguish current from superseded law, understand hierarchies of and relationships among authorities, and support each conclusion with verifiable sources. General AI without specialist research capabilities may confidently paraphrase an outdated statute without flagging that it has been replaced. Specialist AI should surface decisions under both the old and new regimes, explain where they diverge and cite the authority supporting each proposition.
These challenges are especially acute in international law and arbitration, where sources span jurisdictions, languages, institutions and legal traditions. Much of the relevant material is fragmented or inconsistently structured. General AI systems do not acquire the methods of arbitration research simply by receiving more documents. Meeting those demands requires specialist systems purpose-built for the sources, structures and methods of the field. This is a core focus of AI for Legal training, which addresses how domain-specific retrieval changes research workflows.
Jus Mundi's arbitration agent, Jus AI, applies this approach to international law and arbitration research. It combines a specialist corpus with Tenet v5, a proprietary model designed and trained by Jus Mundi specifically for the field. Tenet v5 interprets legal text alongside the metadata and surrounding context that determine an authority's relevance, rather than relying on textual similarity alone. In an independent evaluation by 20 leading arbitration experts, Jus AI received an average score of 4.28 out of 5 across reasoning, correctness, faithfulness to sources and completeness, with 80% of evaluations rated Good or Excellent.
The performance does not come from the model alone. A specialist agentic architecture scopes the question, breaks it into legal research tasks, retrieves from a curated corpus, checks the evidence and produces an answer grounded in the sources consulted. A dedicated quality-assurance protocol adds further controls. For paralegals and legal support staff adapting to these tools, an AI Learning Path for Paralegals covers how citation-backed research systems fit into document review and case preparation.
Integration without losing the specialist layer
The case for specialist AI is not that it is better suited to every legal task or that it will replace general AI platforms. The two layers' greatest value lies in combining their strengths. Specialist AI can still be accessed on a standalone basis, but the direction is increasingly toward integration with the environments in which lawyers already work.
One route is partnership and interoperability. On July 30, 2026, Jus Mundi launched its MCP connector for Claude, having previously introduced its agent-to-agent integration with Legora. Practitioners can access citation-backed arbitration and international law intelligence without switching context or sacrificing the specialist infrastructure behind the answer. Dipen Sabharwal KC, Partner at White & Case, said after testing the Jus AI-Legora integration: "The adoption of Jus AI adds trusted, citation-backed arbitration intelligence from their extensive dataset and will enable our lawyers to work more efficiently, surface deeper insights and deliver even greater value to our clients."
Another route is to incorporate specialist capabilities into general AI. Legora's acquisition of Qura illustrates this approach, reflecting the growing recognition that general legal AI platforms need specialist research infrastructure. But it also raises another question: can specialist legal intelligence simply be absorbed into a general platform, or does its reliability depend on preserving the distinct data, methods and quality controls that made it specialist in the first place?
Whichever integration model prevails, the decisive issue is whether the integrity of the specialist layer is preserved. Interfaces may change, but reliable legal research will continue to rest on the same foundations: authoritative data, specialist retrieval, and sources lawyers can verify.
Why this matters for legal professionals
When evaluating any AI tool for substantive research, ask whether it can show its work at the source level. A system that cannot produce the exact authority behind each proposition forces you into a choice between blind trust and line-by-line verification. That choice erases the efficiency AI is supposed to deliver. The practical test is simple: if the tool cannot cite the specific paragraph or decision supporting its answer, treat the output as a starting point, not a work product.
Your membership also unlocks:
