A fabricated legal citation can turn an otherwise credible court filing into a professional liability problem. Filevine is adding citator and anti-hallucination capabilities to LOIS Legal Research, aiming to help attorneys identify nonexistent cases, inaccurate citations and overlooked negative treatment before legal documents are filed.
Generative AI has made legal research faster, but it has also introduced a particularly dangerous failure mode: an answer can look authoritative while pointing to a case that does not exist—or a real case that does not support the proposition attributed to it.
For lawyers, that distinction is consequential.
Filevine is attempting to address the problem directly with new citation verification and anti-hallucination capabilities inside LOIS Legal Research, its AI-powered legal research platform. The company says the new functionality is designed to identify problematic authorities while attorneys are researching and drafting, rather than leaving citation verification as a separate step at the end of the workflow.
The timing reflects a growing concern across the legal industry. Courts in the U.S. have sanctioned attorneys after filings contained AI-generated citations to nonexistent cases or misrepresented legal authorities. The issue has become one of the clearest examples of why generative AI in professional services requires human oversight.
LOIS does not eliminate that obligation. Lawyers remain responsible for checking the authorities they cite before filing. Filevine’s proposition is narrower and more practical: automate part of the verification process so attorneys can identify questionable citations faster.
Inside LOIS, citations receive a status, while unverified cases can be flagged before filing. Verification also takes place within the same environment where a lawyer conducts research and drafts a brief.
That workflow matters because citation checking can become tedious when performed manually across a large document. An attorney may need to confirm that a case exists, locate the relevant passage, determine whether the cited proposition accurately reflects the holding and check whether subsequent decisions have limited, distinguished or overturned it.
The problem becomes even harder when AI-generated research is involved.
Large language models such as those powering products from OpenAI, Google and Microsoft are designed to generate plausible language, not function as traditional legal databases. Without appropriate retrieval and verification mechanisms, a model can produce an apparently authoritative citation that has no underlying source.
Even when the case is genuine, the model can misinterpret its holding.
Filevine says its own legal team tested LOIS against 68 federal filings that courts had previously sanctioned for fabricated or misused authority. Across those filings, the team reviewed 2,073 citations and identified 174 severe errors that had been missed by the attorneys who originally submitted the documents.
Those results are significant, although they should be viewed as a company evaluation rather than an independent benchmark. The test set was specifically composed of filings already known to contain serious citation problems, so it does not establish how LOIS would perform across ordinary legal research.
Filevine also describes a separate comparison involving so-called silent negative treatment. These are situations in which a court decision has undermined or overruled an earlier holding without explicitly citing the earlier case.
According to Filevine, LOIS identified all 14 silent negative treatments in the test case set, while Westlaw’s KeyCite and LexisNexis’ Shepard’s did not identify any of them.
That claim is particularly interesting because conventional citator systems and AI research tools approach legal authority differently.
Westlaw’s KeyCite and LexisNexis’ Shepard’s have long been core tools for determining whether a case remains good law. Their value comes from large legal databases and extensive citation networks. AI research systems, meanwhile, are increasingly attempting to combine retrieval with natural-language reasoning.
The competitive landscape is therefore beginning to move beyond the question of whether an AI system can find a case. The more important question is whether it can establish that the case actually supports the legal argument being made.
For enterprise legal teams, that distinction could determine how AI is deployed.
A law firm might use generative AI to accelerate initial research, summarize opinions or draft portions of a brief. But those efficiencies do not remove the need for authoritative source retrieval and attorney review. Instead, AI increasingly needs to operate within a workflow that makes verification visible and repeatable.
Filevine’s approach is to place that verification directly inside LOIS rather than requiring lawyers to switch between an AI research tool, a conventional citator and their drafting environment.
That reflects a broader trend in legal technology: AI is moving from standalone assistants toward integrated systems that combine retrieval, reasoning, drafting and workflow controls.
The distinction is important as legal organizations evaluate products based not only on how much time they save, but also on how safely they can be deployed.
For legal departments and law firms, the relevant metrics will include citation accuracy, source provenance, auditability, false-positive rates and how effectively the system catches errors that conventional workflows miss.
The strongest legal AI systems are therefore unlikely to be those that simply produce the most convincing prose. They will be the systems that make it easier for attorneys to determine why an answer is correct, where the underlying authority comes from and whether that authority remains valid.
Filevine’s latest LOIS capabilities are aimed squarely at that problem.
The company’s message is not that AI can replace legal judgment. It is that technology can move some of the most time-consuming verification work closer to the point where legal arguments are created.
For an industry increasingly experimenting with AI-assisted research and drafting, that could prove more valuable than another model capable of producing a polished brief.
Market Landscape
Legal AI is moving into a more mature phase in which accuracy, provenance and verification are becoming as important as generative capabilities.
Traditional legal research platforms such as Westlaw and LexisNexis built their reputations around authoritative databases, citation networks and tools for determining whether precedents remain valid. Newer AI systems are adding conversational research, document analysis and drafting capabilities.
The emerging opportunity lies in combining both approaches.
AI can accelerate discovery and synthesis, but legal professionals still need reliable primary sources and mechanisms for checking whether generated claims accurately reflect those sources. This creates demand for retrieval-augmented legal research, citation verification, citators and audit trails.
The problem is especially acute in litigation, where a fabricated authority can result in sanctions, reputational damage and additional proceedings.
Filevine’s LOIS strategy illustrates where the market could be heading: AI-assisted legal work with verification built into the workflow rather than bolted on afterward.
For enterprise legal teams, however, vendor claims around hallucination detection should be evaluated carefully. Controlled benchmarks, independent testing and transparent methodology will become increasingly important as firms decide which AI systems can safely handle substantive legal work.
Top Insights
- Filevine added citation verification to LOIS, allowing lawyers to identify unverified or potentially inaccurate authorities within an AI-assisted legal research and drafting workflow.
- The technology targets AI hallucinations in legal research, including nonexistent cases and real decisions whose holdings have been incorrectly represented in court filings.
- Filevine reports finding 174 severe citation errors across 2,073 citations from 68 previously sanctioned federal filings, though the evaluation was company conducted.
- LOIS reportedly detected 14 silent negative treatments missed by KeyCite and Shepard’s in a separate test, highlighting a potential gap in conventional citation analysis.
- Legal AI adoption is shifting toward verification, with law firms increasingly requiring source provenance, auditability and human oversight alongside generative drafting capabilities.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
