Can Harvey AI Hallucinate Case Citations Like ChatGPT?

Can Harvey AI hallucinate case citations like ChatGPT?

Yes, Harvey can still hallucinate. The risk is lower than with general-purpose ChatGPT when Harvey answers from LexisNexis primary law, but no independent benchmark has measured Harvey's citation error rate. Courts sanction fake citations no matter which tool produced them.

Here is the verdict for each tool:

• Harvey (Harvey AI, San Francisco): fits large and mid-size firms that want a legal-specific drafting and research platform with Lexis-grounded answers, and that will pay enterprise prices for it.
• ChatGPT (OpenAI): fits general drafting, summarizing and brainstorming. It is the wrong place to look for case law you plan to cite.
• OpenAI for Attorneys (Lexi): fits firms whose real bottleneck is operations (intake, engagement letters, document collection, filings prep, billing) rather than case-law research.

One disclosure: we build one of these tools. Lexi is the AI paralegal in OpenLaw for Attorneys. It is not the product at getlexi.io, which comes from a different company that also uses the name Lexi. Lexi gets one entry and one honest take here, same as the rest.

This page is for firm owners, partners and administrators deciding which AI to trust with research and client work. Prices and vendor terms were checked as of October 2026.

Why do legal AI tools hallucinate case citations?

Large language models predict likely text, so they can write a case name, reporter cite and quotation that look right and don't exist. Retrieval, meaning answering from a real legal database, cuts that risk but doesn't remove it.

The best public evidence comes from Stanford RegLab and Stanford HAI. Their study, published in May 2024, ran more than 200 legal queries through the leading research tools:

• Lexis+ AI: incorrect or misgrounded answers more than 17% of the time.
• Ask Practical Law AI: more than 17%.
• Westlaw AI-Assisted Research: more than 34%.
• General-purpose chatbots such as GPT-4: between 58% and 82% in the researchers' earlier testing.

The researchers counted two kinds of error. One is an answer that misstates the law. The other is a misgrounded citation, where a real case gets cited for something it doesn't say. Real-but-wrong citations are harder to catch than invented ones, because the cite checks out when you look it up.

Harvey was not in the study, and as of October 2026 Harvey has not published an independent hallucination benchmark. The honest answer is that grounded tools beat raw chatbots, and none of them is error-free. For a wider look at the research tools, see how AI is transforming legal research.

Does Harvey AI have access to live case law like Lexis+ AI?

Yes, through LexisNexis. On June 18, 2025, LexisNexis and Harvey announced an alliance that brings LexisNexis primary law and Shepard's Citations into Harvey. Harvey launched Ask LexisNexis in open beta on August 12, 2025.

According to Harvey, Ask LexisNexis answers natural-language questions from U.S. federal and state case law, statutes, administrative codes and agency decisions. It returns pinpoint citations linked to the full text, and Shepard's shows whether a cited authority is still good law. That is a different architecture from ChatGPT, which doesn't check a citator before it writes a cite.

There are limits:

• Feature, not default: grounding applies when you use the LexisNexis feature. A general Harvey prompt or draft isn't the same as a Shepardized research answer.
• Misgrounding still happens: a real case can still be cited for a proposition it doesn't support. That is the error Stanford found in Lexis+ AI itself.
• Beta status: Harvey described the feature as open beta at launch, so check its current status in your contract.

What happens if a hallucinated citation reaches the court?

The lawyer who signed the filing answers for it. Courts haven't excused fake citations because they came from a premium tool.

The case most people know is Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y.). On June 22, 2023, U.S. District Judge P. Kevin Castel fined two lawyers and their firm, Levidow, Levidow & Oberman, $5,000. They had filed opinions invented by ChatGPT and kept standing by them after the court questioned whether they existed. Sanctions have kept coming since. In March 2026, the Sixth Circuit's order in Whiting v. City of Athens, Tennessee imposed $15,000 in punitive sanctions on each of two lawyers over more than two dozen fake or incorrect citations.

ABA Formal Opinion 512, issued July 29, 2024, sets the professional standard. Lawyers must understand what a generative AI tool can and can't do, and they can't rely on its output without checking it independently. The opinion also covers confidentiality, candor toward the tribunal, supervision and fees. None of those duties changes with the vendor.

How much does Harvey cost compared to ChatGPT?

Harvey doesn't publish prices. ChatGPT does: as of October 2026, ChatGPT Business costs $25 per user per month billed monthly, or $20 billed annually. Treat any number as a starting point for negotiation.

Harvey sells through annual contracts quoted by its sales team. Third-party estimates of its per-seat cost vary widely and Harvey hasn't confirmed them, so we don't repeat them here. Get a written quote that spells out seat minimums, the LexisNexis content, and whether Ask LexisNexis is included or billed as an add-on.

The price gap reflects different products. A ChatGPT seat buys a general assistant. A Harvey seat buys a legal platform with legal content behind it. Neither price covers the attorney time needed to verify citations. To weigh that time against the subscription, see the ROI of AI for law firms.

Is Harvey AI safe for confidential legal work?

Harvey's published terms are strong. Its security page says it doesn't use customer inputs, outputs or uploaded documents to train models, requires zero data retention from its model providers, and holds SOC 2 Type II and ISO 27001, 27701 and 42001 certifications.

Privilege is a separate question from security. In United States v. Heppner (S.D.N.Y.), Judge Jed S. Rakoff ruled from the bench on February 10, 2026, and issued a written opinion on February 17, 2026. He held that documents a defendant created with the consumer version of Anthropic's Claude were protected by neither attorney-client privilege nor work product. The reasons were that the provider's terms allowed it to collect and train on the data, and that counsel hadn't directed the use. The court left open whether enterprise tools that exclude training and promise confidentiality would come out differently.

In practice, keep client matters out of consumer ChatGPT accounts. Use enterprise terms with any vendor, and have lawyers direct the AI work. For the checklist to put to a vendor, read what law firms should demand on SOC 2.

Where does each tool still win?

Each tool wins at a different job, and ChatGPT's job isn't citing cases.

Where Harvey still wins

Harvey is the strongest of the three for legal research and drafting at scale. It has Lexis-grounded answers with Shepard's, co-developed workflows for motions to dismiss and summary judgment, document review across large sets, and the security certifications that big-firm risk committees ask for. It is also the only one of the three built around case law.

Where ChatGPT still wins

ChatGPT is cheap, quick to roll out and good at plain-language work: rewriting a client letter, summarizing a non-confidential article, outlining a presentation. On a Business plan it fits non-research tasks well.

Where Lexi wins

Lexi takes on the back-office work that eats a small firm's week. It books intake recalls, drafts engagement letters and discovery responses, collects documents, prepares filings, processes payments and updates matter files. It works with Clio Manage, Thomson Reuters tools and Google Calendar. Lexi isn't a case-law research database, and it doesn't claim to be one. The stack it consolidates is set out in what Lexi replaces.

Can firms use Harvey and ChatGPT, or Harvey and Lexi, together?

Yes, as long as each tool has a defined job. Firms most often run a grounded research tool for anything that will be cited and keep general assistants and operations tools away from citations.

A workable split looks like this:

• Research and briefs: Harvey with Ask LexisNexis, or Lexis+ AI or Westlaw, followed by a human cite-check.
• General writing: ChatGPT Business, with a written policy that no case citation from it goes into a filing.
• Firm operations: Lexi for intake, engagement, documents, deadlines and billing.

Overlap costs money, so map the tasks before you buy seats. Our guide on how to choose the right AI tool for your law firm has a task-mapping worksheet.

How do you check AI-generated case citations before filing?

Treat every AI citation as unverified until a person has read the case. The same steps apply whether the draft came from Harvey, ChatGPT or anything else.

1. Pull each cited case from Lexis, Westlaw or the court's own site. Never rely on the AI's summary of it.
2. Read the pinpoint page and confirm it supports the exact proposition in the brief.
3. Run the cite through Shepard's or KeyCite to confirm it is still good law.
4. Check every quotation word for word against the opinion.
5. Record who verified each citation and when, so supervision under ABA Formal Opinion 512 can be shown later.

Choose Harvey if

Your firm does heavy research and drafting, has the budget for enterprise seats, and wants Lexis-grounded answers inside one platform.

Choose ChatGPT if

You need low-cost help with general writing and will keep it out of case research and client-confidential consumer accounts.

Choose Lexi if

Your hours go to intake, paperwork, follow-up and billing rather than research, and you already pay for a research database.