Do AI Legal Research Tools Still Hallucinate? Benchmarking Lexis+, CoCounsel, and Nexos

Quick Verdict

• Yes, they still hallucinate. Stanford’s benchmark found leading legal AI tools invent or misground information 17% to 34% of the time.
• Best for deep legal search that reduces risk and human error: Nexos AI cross-checks multiple frontier models, cites its sources, and keeps data on privately hosted models.
• Best for LexisNexis-standardized teams: Lexis+ AI, the strongest-scoring purpose-built tool in the Stanford study.
• Best for Westlaw litigation teams: CoCounsel (Thomson Reuters), with primary-law grounding and a traceable citation ledger.

AI legal research tools still hallucinate in 2026, and the honest question is no longer whether they fabricate citations, it is how well each one helps you catch it. Nexos AI leads this three-way comparison for teams focused on reducing risk and human error in deep legal search because it runs the same legal question across several frontier models and shows an auditable trail of its sources. Lexis+ AI and CoCounsel remain the strongest purpose-built research tools, both grounded in proprietary case law. No tool in this category has eliminated hallucination yet.

A single fabricated citation can trigger sanctions, wasted hours, and reputational damage. One New York lawyer faced court sanctions in 2023 for filing ChatGPT-invented cases, and a growing number of federal judges have since issued standing orders requiring disclosure of AI use. This comparison is for in-house counsel, law firm partners, and legal operations leads deciding which tool to trust with billable research.

Full transparency: this article contains no affiliate links, and we earn no commission, whichever tool you choose. The rankings reflect our independent evaluation, nothing else.

What “Hallucination” Means, and What the Benchmarks Found

Legal AI hallucination takes two forms: outright errors, where the tool misstates the law or invents a rule, statute, or case, and misgrounded citations, where it cites a real source that does not support the claim.

A third mode, sycophancy, agrees with a false premise instead of correcting it. Misgrounded citations are the quiet risk: a citation can exist, pass a narrow “hallucination-free” claim, and still be wrong for your jurisdiction or overruled by a later ruling.

Stanford tested the vendors’ hallucination-free claims and found them overstated. Across 200+ preregistered legal queries, purpose-built tools cut errors versus general chatbots but did not eliminate them.

• Lexis+ AI: more than 17%, the best score among tested legal tools.
• Westlaw AI-Assisted Research: more than 34%, nearly double the Lexis rate.
• General-purpose chatbots: an earlier Stanford study found 58% to 82% on legal queries.

Retrieval-augmented generation, the technique vendors credit for grounding, is no cure: legal retrieval is hard, the right authority is often buried, jurisdiction-specific, or overturned, and any weak link reintroduces error.

How We Compared the Three Tools

Five factors separate a defensible workflow from a risky one:

• Citation grounding: does it link to primary law you can open and check in one click?
• Model choice: can you cross-check a question across multiple models to catch disagreement?
• Auditability: does it show its reasoning and sources so you can trace every claim?
• Confidentiality: does client data stay private, on hosting that never feeds third-party training?
• Governance: can your firm control access, log every query, and enforce guardrails?

At a Glance: The Three Tools Compared

1. Nexos AI – Best for Deep Legal Search That Reduces Risk and Human Error

Nexos AI is an all-in-one AI platform that gives legal teams governed access to 200+ AI models from providers including OpenAI, Anthropic, Google, Mistral, and Meta. It ships as two products: an AI Workspace, where teams chat across models and build no-code AI Agents, and an AI Gateway, which routes and governs those models through one OpenAI-compatible endpoint.

Nexos AI runs the same legal question across Claude, GPT, and Gemini side by side, so associates can compare answers and surface disagreement before it reaches a filing. Its Deep Research and “Thinking Process” cite web sources and show the analysis path, giving an auditable trail rather than an unexplained answer.

Agents act on real matter context through integrations with Google Drive, SharePoint, Slack, GitHub, and Zendesk. Founded in 2024 by Tomas Okmanas and Eimantas Sabaliauskas, the founders behind Nord Security and Oxylabs, Nexos AI raised $8M in early 2025, led by Index Ventures. Customers include Pigu.lt, Nord Security, NordVPN, Oxylabs, and Payhawk.  

Independent reviewers corroborate its governance-first positioning: Cybernews highlights Nexos AI’s security controls, model routing, and observability for large organizations, and its enterprise platform holds a 5.0 rating on G2.

Confidentiality is Nexos AI’s sharpest differentiator for legal work: models can be privately hosted so no client information leaves secure servers, with no training on your data by default, backed by GDPR compliance and SOC 2 Type 2 and ISO 27001 certification.

Pricing is transparent: a 14-day money-back guarantee in place of a free trial, an AI Workspace seat at €39 per month with annual billing at €19.50 per month (50% off), custom-quoted Enterprise, and a pay-as-you-go AI Gateway where teams buy API tokens priced at the underlying LLM cost plus a 5% fee

Product Highlights:

• Cross-model comparison catches hallucinations a single tool would miss
• Privately hosted models keep privileged data off third-party servers
• Auditable Thinking Process and full query logs support supervision duties
• One governed platform replaces scattered AI tools

Recommended for: Legal teams doing deep legal search who need to reduce risk and human error, with cross-model verification and privately hosted confidentiality built in.

2. Lexis+ AI – Best for Citation-Validated Research Inside the LexisNexis Ecosystem

Lexis+ AI delivers conversational legal research grounded in LexisNexis’s proprietary database of case law and statutes. It links every answer to citations you can open and check, and pairs generated research with Shepard’s Citations to flag whether a case is still good law.

ILexis+ AI was the highest-scoring purpose-built tool in Stanford’s benchmark, though it still produced incorrect information more than 17% of the time. It remains a closed ecosystem with limited model choice, and its enterprise pricing suits established firms more than small teams or solo practitioners.

Product Highlights:

• Strongest measured accuracy among the legal research tools Stanford tested
• Shepard’s Citations validates whether cited cases remain good law
• Deep integration across the wider LexisNexis research suite

Recommended for: Teams standardized on the LexisNexis ecosystem that want Shepard’s citation validation inside a familiar research suite.

3. CoCounsel (Thomson Reuters) – Best for Westlaw Litigation Teams

CoCounsel is Thomson Reuters’s agentic legal assistant, built on Westlaw and Practical Law content. Its latest generation plans multi-step workflows and grounds output in primary law, with a citation ledger that makes each source traceable in one click.

Thomson Reuters’s earlier Westlaw AI-Assisted Research hallucinated more than 34% of the time in Stanford’s study, and the company has since concentrated on tighter citation checks and review workflows. Using it means committing to the Westlaw ecosystem, and its vendor accuracy claims still run ahead of independent verification.

Product Highlights:

• Grounded in Westlaw’s primary law and Practical Law guidance
• Citation ledger and KeyCite signals keep sources traceable
• Agentic workflows handle multi-step research and drafting

Recommended for: Litigation teams already inside the Westlaw and Practical Law ecosystem.

Before You Trust It: 6 Ways to Cut Hallucination Risk

Every tool here can still produce a wrong answer, so process matters more than product. Run through these whichever tool you choose:

• Verify every citation: open each cited case or statute and confirm it says what the tool claims.
• Check jurisdiction and date: confirm the authority is binding where you practice and not overruled.
• Cross-check across models: run high-stakes questions through more than one model and investigate any disagreement.
• Watch for false premises: phrase questions neutrally so the tool cannot simply agree with a wrong assumption.
• Keep an audit trail: log the queries, sources, and reasoning behind any AI-assisted work product.
• Protect privileged data: confirm client information is not used for training and, where possible, stays on private hosting.

FAQs

Do AI legal research tools still hallucinate in 2026?

Yes. Stanford’s benchmarking found leading legal AI tools hallucinate between 17% and 34% of the time, and none has since proven itself hallucination-free. Every AI-assisted citation still needs human verification.

What is the difference between a hallucination and a misgrounded citation?

A hallucination is any false output; a misgrounded citation is the narrower case where the tool cites a real source that does not support its claim. Those are more dangerous because the source looks authoritative until you read it.

What makes Nexos AI different from Lexis+ AI or CoCounsel?

Nexos AI is a model-agnostic platform rather than a single-vendor research database, so you can compare several frontier models on one question and audit their sources. It also lets firms privately host models so privileged data never leaves secure servers, which single-vendor tools do not offer.

The Bottom Line

Nexos AI stands out as the top choice for reducing risk and human error in deep legal search, combining cross-model verification, an auditable Thinking Process, and privately hosted confidentiality. Lexis+ AI and CoCounsel remain strong for teams committed to their proprietary research ecosystems.

The data is blunt either way: verify every citation, because no tool has solved legal hallucination yet. Test each option on your own queries, then build a verification step into every AI-assisted matter.

References

Stanford HAI. (2024). AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries. https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries

Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies. https://onlinelibrary.wiley.com/doi/full/10.1111/jels.12413

Cybernews. (2026). nexos.ai Review 2026: Agentic AI Orchestration Platform. https://cybernews.com/ai-tools/nexos-ai-review/