We respect attorney-client confidentiality. No tracking pixels in our emails.
A practical step-by-step guide to AI legal research in 2026, including a 7-point citation verification checklist and tool selection by jurisdiction.
It is 11 PM. The brief is due at 9 AM. You need five circuit court cases on implied covenant of good faith in employment contracts. Your usual research takes three hours. With AI, the same search returns results in four minutes — but three of those results might be fabricated.
That is the tension at the center of every AI legal research workflow in 2026. This guide gives you the workflow to capture the speed without accepting the risk.
This is our step-by-step guide to using AI for legal research in 2026, written for practicing attorneys who need a structured, verifiable workflow rather than a general-purpose AI tutorial.
LawyerAI built this guide. We earn no affiliate revenue from these tools.
Here are the 4 rules we set for ourselves before writing this:
We re-review this list every quarter.
Short answer: AI legal research is fastest when your question is specific, your tool is corpus-grounded (not a general LLM), and your verification protocol is systematic. The Stanford RegLab 2024 study gives us the only independent accuracy data: Lexis+ AI and CoCounsel at 17% hallucination rate, Westlaw Precision AI at 33%. Use grounded tools, apply the 7-point citation checklist before any citation goes in a filing, and document your AI research process for the matter file. The workflow below takes the four-minute result and turns it into something you can use.
Every tool referenced in this guide is scored across five dimensions at /methodology. For legal research specifically, Accuracy is the most heavily weighted dimension given the professional responsibility implications of ai-hallucination in court filings.
| Tool | Hallucination Rate | Starting Price | Best For |
|---|---|---|---|
| Lexis+ AI | 17% (Stanford RegLab 2024) | Lexis base required | Federal and regulatory research |
| Westlaw Precision AI | 33% (Stanford RegLab 2024) | Westlaw base required | Westlaw-primary users |
| CoCounsel | 17% (Stanford RegLab 2024) | Westlaw base required | Research + drafting |
| Paxton AI | Not published (independent) | $65/seat/month | Budget-accessible research |
| Harvey AI | Not published (independent) | $140K+/year | Complex synthesis |
Before opening any AI tool, write out your research question in one sentence. Specificity is not just good research practice — it materially reduces hallucination risk. A narrow question constrains the AI's output to a well-defined domain. A broad question invites the model to range across topics where its training is thinner.
Compare these:
Broad: "What is the law on employment contracts?"
Narrow: "What is the standard in the Ninth Circuit for an implied covenant of good faith and fair dealing claim in an at-will employment termination context?"
The narrow question defines jurisdiction, doctrine, cause of action, and factual context. An AI tool given the narrow question is working in a constrained space. The broad question produces a survey that may be generally accurate but miss the specific controlling authority.
Write your question before you type it into any tool. Revise it until it could not be interpreted in more than one way.
Tool selection for legal research should follow the jurisdiction, not the other way around.
For US federal research — appellate circuits, Supreme Court, federal statutes and regulations — Lexis+ AI and CoCounsel are the highest-accuracy options per the Stanford RegLab 2024 benchmark. Both operate at 17% hallucination rate in that study. If you have an existing LexisNexis subscription, Lexis+ AI is the default choice. If you have an existing Westlaw subscription, CoCounsel (which requires that subscription) produces better accuracy than Westlaw Precision AI per the same benchmark.
For state-law-heavy research where the major platform pricing is prohibitive, Paxton AI at $65 per seat per month is accessible. Its corpus is smaller than Westlaw or LexisNexis, and no independent accuracy data has been published. Verification of every citation is mandatory.
For complex synthesis — multi-issue memos, cross-jurisdictional surveys, regulatory landscape analyses — Harvey AI produces more integrated output than database-search tools. Its synthesis capability is stronger than its search capability, but it requires a $140,000+ minimum contract.
For international law, none of the major US platforms have strong non-US coverage. vLex covers more jurisdictions and is worth investigating for cross-border work, though it is not on the LawyerAI directory currently.
After constructing your precise question and selecting the tool, run the query and evaluate the output before acting on it. Two things to check before proceeding:
First: does the AI's response directly address the specific question you asked, or has it drifted to a related but different question? Model drift is common — an AI asked about the Ninth Circuit implied covenant standard may answer with a general survey of implied covenant doctrine across all circuits. If the answer does not match the question, refine the query and run it again.
Second: how many citations did the AI generate? A response with 15 citations to a narrow question is a flag. Narrow research questions should produce a focused set of controlling or highly relevant authorities. A long list of peripheral citations may indicate the model is filling space rather than answering the question.
This evaluation step takes two minutes and saves the verification time of chasing irrelevant or fabricated citations.
Every citation an AI tool generates must pass all seven checks before use in a court filing or client advice. No exceptions.
Does the case exist? Search the citation in Westlaw, LexisNexis, or the official federal reporter. Do not assume existence because the citation looks correctly formatted.
Is the citation format correct? Volume number, reporter, first page, year, court. A real case with a wrong page number is still a citable error.
Is the holding accurately stated? Read the relevant portion of the actual opinion. AI tools frequently summarize holdings in ways that shift the actual legal conclusion, sometimes subtly.
Is the case still good law? Shepardize (LexisNexis) or KeyCite (Westlaw). A case that has been overruled or distinguished on the specific point you are citing is not citable for that point.
Is the jurisdiction correct? A Ninth Circuit case does not bind a Seventh Circuit court. An AI tool searching broadly may return the right legal principle from the wrong jurisdiction.
Does the case actually support the proposition you are citing it for? The hardest check. The AI may correctly identify a case as relevant while summarizing it in a way that supports your argument better than the actual holding does. Read the case.
Is it a direct quote or a paraphrase? If you are using quoted language from the case, verify the quote is verbatim. AI tools hallucinate quotations as confidently as they hallucinate citations. See our citation-validation glossary entry for the full protocol with examples.
After applying the 7-point checklist to your AI-generated citations, run one cross-reference check: take your central legal proposition and verify it against a secondary source — a treatise, law review article, or ALR annotation. This is not to duplicate the citation verification; it is to confirm that the legal framework the AI generated is accurate.
AI research tools can produce technically accurate citations that collectively support a flawed legal framework — citing real cases for real propositions that, assembled incorrectly, misstate the law. The secondary source check catches framework errors that citation-level verification misses.
This step is most important for complex multi-issue research where the AI is doing doctrinal synthesis rather than case retrieval.
Document that you used AI in your research process. This is not optional under the ABA Model Rule 1.1 competence framework, which courts and state bars have increasingly applied to AI use. See our ai-competency-lawyers entry for current court rules and state bar guidance.
The documentation does not need to be extensive. A one-paragraph entry in your matter file noting which tool was used, which queries were run, and that all citations were verified using the 7-point checklist is adequate for most purposes. If a court rule requires disclosure of AI use in filings — an increasing number do — your documentation provides the basis for that disclosure.
Find your closest match for the research task at hand:
How accurate is AI legal research today?
The best independent data comes from the Stanford RegLab 2024 benchmark. Lexis+ AI and CoCounsel tied at 17% hallucination rate on legal research tasks. Westlaw Precision AI measured 33%. GPT-4 without legal grounding measured 88%. No independent data exists for most other tools, including Harvey AI and Paxton AI. What 17% means in practice: for every 100 AI-generated citations, 17 will contain a material error before you catch it. That is not a reason to avoid the tools — it is the reason the 7-point checklist is non-negotiable.
How do I verify a citation in 30 seconds?
You cannot fully verify a citation in 30 seconds if you have not read the case. What you can do in 30 seconds is confirm existence and basic accuracy: search the citation in Westlaw or Lexis, confirm the case comes up, confirm the case name matches, and confirm the cited page contains a passage relevant to your proposition. That is the minimum check. The full 7-point protocol takes 5-10 minutes per citation the first time and gets faster with practice.
Which tool is safest for court filings?
"Safest" means lowest combined risk of hallucination and highest ease of verification. Based on available evidence, Lexis+ AI and CoCounsel are the safest options at 17% hallucination rate with direct Shepardizing and KeyCite integration that speeds up the good-law check. No tool is safe enough to skip manual verification. The Mata v. Avianca sanctions were not imposed because a tool failed — they were imposed because attorneys filed briefs without verifying AI output.
Do I need to disclose AI use to courts?
This varies by jurisdiction and is changing rapidly. As of mid-2026, a growing number of federal district courts have adopted local rules requiring disclosure of generative AI use in filings. Several state courts have similar requirements. Check the specific local rules for the court you are filing in. The ABA's Model Rule 1.1 competence framework applies regardless — you are responsible for the accuracy of everything in your filings, AI-generated or not.
What is the ABA rule on AI legal research?
ABA Model Rule 1.1 requires competence, which the ABA has interpreted to include understanding the benefits and risks of relevant technology. In the context of AI research, this means: understanding what hallucination is and how to verify against it, knowing the accuracy limitations of tools you use, and having a systematic verification process for AI-generated citations. ABA Formal Opinion 512 (2023) provides specific guidance on generative AI use. State bar ethics opinions have proliferated since 2023 and vary in specifics. The baseline obligation — verify your citations — is consistent across all guidance issued to date.
LawyerAI evaluations are independent. We do not accept payment that influences our editorial scores. Featured placements are clearly labeled and do not affect our 5-dimension methodology (Accuracy / Speed / Usability / Value / Security). We re-review tools every 6 months.
If you believe any information is inaccurate, contact editor@lawyerai.directory.