Find What Matters: Smarter Legal Research
Legal research used to mean hours in Westlaw, stacks of printed cases, and a billable rate that made clients wince. AI hasn't replaced that process, it has compressed it. Attorneys, paralegals, compliance officers, and in-house counsel are now using AI tools to find relevant case law in minutes, summarize 80-page decisions in seconds, and draft research memos before lunch. This lesson shows you exactly how to do that, which tools to trust, and where the real risks live.
7 Things You Need to Know About AI Legal Research
- AI legal research tools fall into two categories: general-purpose AI (ChatGPT, Claude) and legal-specific AI (Westlaw AI, Lexis+ AI, Casetext CoCounsel). They behave very differently and carry different levels of risk.
- General-purpose AI tools hallucinate citations. They will confidently invent case names, docket numbers, and holdings that do not exist. Never cite a case you found through ChatGPT without verifying it in an official legal database.
- Legal-specific AI tools (Westlaw AI, Lexis+ AI) are trained on verified legal databases and can link directly to real cases, dramatically reducing hallucination risk for citation-dependent tasks.
- Prompt quality determines output quality. A vague question gets a vague answer. A well-structured prompt with jurisdiction, legal issue, and context gets a usable research memo draft.
- AI cannot replace attorney judgment. It can surface relevant precedent and summarize holdings, but assessing how a judge in the Southern District of New York will weigh those cases requires human expertise.
- Client confidentiality rules apply. Inputting client names, case details, or privileged information into a public AI tool (like the free version of ChatGPT) may violate your professional responsibility obligations.
- AI research tools are moving fast. Casetext CoCounsel (acquired by Thomson Reuters) costs approximately $100/month per user. Lexis+ AI is available as an add-on to existing Lexis subscriptions. Westlaw AI is bundled into select Westlaw plans.
Two Types of AI Research Tools, and Why the Difference Matters
Think of general-purpose AI tools like a very well-read research assistant who studied everything, novels, news articles, law school textbooks, Wikipedia, and millions of court opinions scraped from the web. They know a lot. But their knowledge has a cutoff date, they can't access live databases, and they sometimes misremember details with total confidence. Using ChatGPT to find case law is like asking that assistant to quote a case from memory, they'll give you something that sounds exactly right, but the citation may be completely fabricated.
Legal-specific AI tools are different because they're built on top of verified, continuously updated legal databases. When Westlaw AI or Lexis+ AI cites a case, it's pulling from the same authenticated database you'd search manually. The AI is acting as a search and synthesis layer on top of a trusted source, not generating text from memory. This distinction is the single most important concept in AI-assisted legal research. It determines whether your research is a starting point or a liability.
- General-purpose AI (ChatGPT, Claude, Gemini): Good for summarizing documents you paste in, drafting research memos from verified notes, brainstorming legal theories, and explaining concepts in plain language.
- Legal-specific AI (Westlaw AI, Lexis+ AI, Casetext CoCounsel): Good for searching case law, finding statutes, checking whether a case is still good law, and generating research memos with real citations.
- The hybrid approach: Use legal-specific AI to find and verify cases, then use a general-purpose tool to help draft the memo or client summary, keeping all confidential details out of public tools.
- Paralegal and in-house teams without Westlaw or Lexis subscriptions: Use Google Scholar (free) to verify any case a general AI surfaces before relying on it.
- Bar association guidance is evolving: As of 2024, multiple state bars (California, New York, Florida) have issued formal guidance on AI use in legal practice. Check your jurisdiction's current rules.
Use Claude for Document Summarization
| Tool | Type | Best For | Hallucination Risk | Approx. Cost |
|---|---|---|---|---|
| ChatGPT Plus (GPT-4o) | General-purpose | Drafting, summarizing pasted text, brainstorming legal arguments | HIGH for citations | $20/month |
| Claude Pro | General-purpose | Summarizing long documents, drafting memos from your notes | HIGH for citations | $20/month |
| Microsoft Copilot (M365) | General-purpose | Summarizing Word docs, drafting emails, meeting notes | HIGH for citations | Included in M365 Business plans |
| Westlaw AI (Thomson Reuters) | Legal-specific | Case law search, statute lookup, KeyCite verification | LOW | Bundled in select Westlaw plans |
| Lexis+ AI | Legal-specific | Case law, secondary sources, Shepard's citation check | LOW | Add-on to Lexis subscriptions |
| Casetext CoCounsel | Legal-specific | Case law search, deposition prep, contract review | LOW | ~$100/month per user |
| Google Scholar | Free legal database | Verifying case citations, reading opinions (no AI synthesis) | N/A, manual search | Free |
The Hallucination Problem. What It Is and How It Happens
Historical Record
Steven Schwartz
In early 2023, attorney Steven Schwartz submitted a brief to federal court citing six cases, all of which were fabricated by ChatGPT. The court sanctioned him.
This case demonstrates the hallucination problem in AI legal research and the serious professional consequences of submitting unverified AI-generated citations.
Here's why it happens: general AI tools like ChatGPT are trained to produce fluent, plausible text, not to retrieve verified facts. When you ask for a case about employer liability for remote work injuries, the model doesn't search a database. It generates text that statistically resembles what an answer to that question would look like, based on patterns in its training data. Sometimes that output matches a real case. Sometimes it constructs a convincing-sounding fiction. The model has no mechanism to flag the difference, which is why it presents both with identical confidence.
- Never ask ChatGPT, Claude, or Gemini to find cases for you by topic and then cite those cases directly.
- Always verify any case name, citation, or holding in Westlaw, Lexis, or Google Scholar before it goes into any document.
- Use general AI only when you supply the source material, paste the case in, ask the AI to summarize it.
- If you use a legal-specific AI tool (Westlaw AI, Lexis+), still spot-check a sample of citations, these tools are better but not infallible.
- Document your verification process. If a client or court ever questions your research, you need to show that a human confirmed every citation.
| Task | Safe to Use General AI? | Recommended Approach |
|---|---|---|
| Finding cases by topic or legal issue | NO | Use Westlaw AI, Lexis+ AI, or Casetext CoCounsel |
| Summarizing a case you've already found | YES | Paste the full text into Claude or ChatGPT |
| Explaining a legal concept in plain language for a client | YES | Use ChatGPT or Claude, verify accuracy before sending |
| Drafting a research memo from your verified notes | YES | Use ChatGPT, Claude, or Copilot with your notes as input |
| Checking if a case is still good law (not overruled) | NO | Use Westlaw KeyCite or Lexis Shepard's only |
| Identifying relevant statutes by jurisdiction | CAUTION | Use legal-specific AI or verify manually in official code |
| Summarizing a contract or agreement you received | YES (with caution) | Use Claude Pro, remove client names first if using public tool |
| Drafting client-facing summaries of legal outcomes | YES | Use AI for drafting, attorney reviews before sending |
What 'Good Law' Means, and Why AI Can Get It Wrong
How to Write Prompts That Get Useful Legal Research Output
Prompting an AI tool for legal research is not like typing a search query into Google. A Google search for 'employer liability remote work injury' returns a list of links you evaluate. An AI prompt for the same topic returns a synthesized answer, and the quality of that answer depends almost entirely on how much context and structure you provide. Think of it like briefing a junior associate: the more specific your instructions, the more useful their work product. Vague instructions produce vague output. Specific instructions, with jurisdiction, parties, legal issue, and desired format, produce something you can actually use.
The single most effective upgrade to any legal research prompt is adding the desired output format explicitly. 'Summarize this case' produces a paragraph. 'Summarize this case in three sections: (1) key facts in three bullet points, (2) the court's holding in one sentence, and (3) the practical implication for employers in two sentences' produces something you can drop directly into a memo. You're not doing more work, you're front-loading the structure so the AI doesn't have to guess what you need. This is especially powerful when you're using AI to process multiple documents quickly.
Prompt
I'm an employment attorney preparing a client memo for a mid-size retail company. Below is the full text of a recent federal circuit court decision on employer obligations for remote work injury compensation. Please summarize it in the following format: 1. Key Facts (3 bullet points, plain language) 2. Court's Holding (1-2 sentences) 3. What This Means for Employers (2-3 sentences, practical focus) 4. Open Questions (any issues the court left unresolved) Keep the language clear enough for a non-lawyer HR director to understand. Do not add any cases or legal citations I haven't provided. [PASTE FULL CASE TEXT HERE]
AI Response
**1. Key Facts** - A warehouse employee who transitioned to remote work during a company restructuring sustained a back injury while setting up her home office workstation during her scheduled shift. - The employee filed a workers' compensation claim; the employer denied it on the grounds that the injury occurred outside company premises and outside the employer's control. - The district court sided with the employer. The employee appealed, arguing the injury occurred during work hours in a space the employer had implicitly approved. **2. Court's Holding** The Ninth Circuit reversed the district court, holding that injuries sustained by remote employees during scheduled work hours, in a designated home workspace, fall within the scope of workers' compensation coverage, provided the employer had knowledge of and acquiesced to the home office arrangement. **3. What This Means for Employers** Retail and other companies with remote or hybrid workers should assume that home-office injuries during work hours may be compensable under workers' compensation. Employers who have issued remote work policies, provided equipment, or formally approved home workspaces are particularly exposed. HR teams should review current remote work agreements and consult with employment counsel about updating injury reporting procedures. **4. Open Questions** The court did not address whether employers can limit liability by requiring employees to use company-approved ergonomic setups, or how this ruling applies to employees who work remotely on an informal or ad hoc basis without a written agreement.
Confidentiality Rules and AI. The Risk Most Attorneys Overlook
Professional responsibility rules, specifically Model Rule 1.6 on confidentiality and Rule 1.1 on competence, apply directly to how attorneys use AI tools. When you paste a client's name, case details, settlement figures, or privileged communications into a public AI tool, that information may be stored, used to train future models, or accessible to the tool's developers depending on the platform's terms of service. The free version of ChatGPT, for instance, historically used conversations for model training unless users opted out. This is not a hypothetical risk, multiple bar associations have issued explicit warnings about it.
The practical solution is not to avoid AI, it's to use AI tools that offer enterprise privacy terms, and to develop a firm habit of anonymizing inputs before they go into any public tool. Replace client names with placeholders ('Client A,' 'the employer,' 'the plaintiff'). Remove identifying case numbers, jurisdiction details, and dollar figures when drafting example prompts. Many firms are also negotiating enterprise agreements with OpenAI (ChatGPT Enterprise), Anthropic, or Microsoft Copilot for Legal, versions that explicitly prohibit using your data for model training and offer stronger data security guarantees.
| AI Tool Version | Data Used for Training? | Enterprise Option Available? | Recommended for Client Data? |
|---|---|---|---|
| ChatGPT Free (GPT-3.5/4o) | Yes, by default (opt-out available) | No | NO |
| ChatGPT Plus ($20/month) | Yes, by default (opt-out available) | No | NO without opt-out |
| ChatGPT Enterprise | No, contractually prohibited | Yes | YES, with firm IT approval |
| Claude Pro ($20/month) | Anthropic states conversations not used for training by default | Claude for Enterprise available | CAUTION, verify current terms |
| Microsoft Copilot (M365 Business) | No, covered under Microsoft's data protection terms | Yes, included in M365 plans | YES, with IT configuration |
| Westlaw AI / Lexis+ AI | No. Thomson Reuters / LexisNexis enterprise terms apply | Subscription-based | YES, designed for legal use |
Check Your Bar's Current AI Guidance Before You Start
Goal: Build a repeatable, defensible AI research workflow: find and verify the source first, then use AI to synthesize, never the other way around.
1. Open Google Scholar (scholar.google.com) and search for a real case relevant to a legal issue you work with regularly, for example, search 'at-will employment wrongful termination' filtered to your circuit. 2. Select one case from the results. Click through to read the full opinion and copy the official case name, citation, and a 2-3 sentence summary of the holding in your own words into a blank document. 3. Open Claude Pro or ChatGPT Plus. Create a new conversation. Do NOT paste any client names or real case details, use this as a practice run with the public case you found in Step 1. 4. Paste the full text of the court opinion into the chat (Google Scholar provides the full text). Use the prompt structure from the prompt example above: request key facts, holding, employer implications, and open questions. 5. Compare the AI's summary to your own notes from Step 2. Note any differences, oversimplifications, or missing nuances. This comparison is your quality-check habit. 6. Copy the AI output into your document below your verified notes. Label it clearly: 'AI-generated summary, verified against source text [date].'
Part 1 Cheat Sheet. AI Legal Research Essentials
- Two tool types: General AI (ChatGPT, Claude) = summarizing documents you provide. Legal AI (Westlaw, Lexis+, Casetext) = finding and citing real cases.
- Hallucination rule: Never cite a case from general AI without verifying it in Westlaw, Lexis, or Google Scholar.
- Safe general AI uses: summarizing pasted documents, drafting memos from your verified notes, explaining concepts in plain language.
- Unsafe general AI uses: finding cases by topic, checking if a case is still good law, identifying statutes by jurisdiction.
- Prompt structure: always include jurisdiction, legal issue, parties/context, and desired output format.
- Confidentiality default: anonymize all client details before using any public AI tool. Use enterprise versions (ChatGPT Enterprise, Copilot for M365) for anything involving real client data.
- Citation verification tools: Westlaw KeyCite (red = overruled, yellow = caution), Lexis Shepard's Citations, no general AI can replicate this.
- Bar obligation: Check your state bar's current AI guidance. Competence under Rule 1.1 includes knowing how these tools work and where they fail.
- Free verification option: Google Scholar (scholar.google.com) is free, shows full opinions, and lets you confirm any case a general AI surfaces.
- Best document summarization tool: Claude Pro, 200,000-token context window handles full contracts and long opinions in one paste.
Key Takeaways from Part 1
- The most important distinction in AI legal research is general AI vs. legal-specific AI, they carry fundamentally different reliability profiles for citation work.
- AI hallucination of legal citations is not a bug that will be fixed soon, it is an inherent behavior of how large language models work. Your verification habit is the fix.
- The safest, highest-value use of general AI in legal work is summarizing documents you already have and trust, not finding documents for you.
- Confidentiality obligations follow you into AI tools. The platform's terms of service determine whether your inputs are private, and many default to using them for training.
- A well-structured prompt, with jurisdiction, issue, context, and output format, is the difference between an AI output you can use and one you have to rewrite from scratch.
Now that you understand what AI legal research tools can and cannot do, the real skill is knowing how to use them precisely. Vague inputs produce vague outputs. The professionals getting the most from tools like Westlaw AI, Lexis+ AI, and ChatGPT are the ones who treat these tools like a sharp but literal-minded research clerk, one who needs clear instructions, specific context, and explicit scope.
7 Things Every Legal Professional Must Know About AI Research
- AI legal tools summarize and surface, they do not advise. The analyzis and professional judgment remain yours.
- Jurisdiction matters enormously. Always specify the state, federal circuit, or country in your prompt. An AI answer about employment law in California can be dangerously wrong if applied in Texas.
- Cutoff dates are real. ChatGPT's training data has a knowledge cutoff. Westlaw AI and Lexis+ AI pull from live databases, but you must verify currency for any fast-moving area like tax or immigration.
- Hallucinations happen. AI tools can fabricate case citations that look real but don't exist. Every cite must be independently verified before it appears in any document you sign or submit.
- Secondary sources first, primary sources second. Use AI to understand the landscape of an issue, then drill into actual statutes and case law.
- Confidentiality is non-negotiable. Do not paste client names, case numbers, or identifying facts into free consumer AI tools. Use firm-approved platforms with data security agreements.
- AI research is a starting point, not a finishing point. It compresses the first two hours of a research task. The last hour, judgment, strategy, application, still belongs to you.
Structuring Prompts for Legal Precision
The single biggest difference between a useful AI research result and a useless one is prompt structure. Think of it like briefing a new associate. You wouldn't say, 'Research non-compete stuff.' You'd say, 'I need a summary of enforceable non-compete clauses under Illinois law, focusing on cases decided after 2018, for a mid-level software engineer earning $95,000.' That specificity, jurisdiction, time frame, context, role, is exactly what AI tools need to return targeted, usable results.
A reliable prompt structure for legal research follows four components: the legal issue, the jurisdiction, the relevant facts or context, and the desired output format. Skipping any one of these forces the AI to guess, and it will guess broadly. When you specify output format, 'give me a bullet list of key factors courts weigh' versus 'write a paragraph summary', you control how the information lands and how quickly you can scan it for relevance. Format is not cosmetic; it determines how fast you can work.
- Legal issue: Be specific, 'wrongful termination under at-will employment doctrine' not just 'firing'
- Jurisdiction: State, federal circuit, or country, always included, never assumed
- Relevant facts: Industry, contract type, dollar amount, timeline, party type (individual vs. corporation)
- Output format: Bullet list, numbered factors, comparison table, plain-language summary, or draft language
- Scope limiter: 'Focus only on case law after 2015' or 'exclude federal cases' narrows noise
- Role framing: 'You are summarizing this for a non-attorney client' or 'for a litigation memo' shifts tone and depth
The Four-Part Prompt Formula
Legal Research Prompt Structures: Quick Reference
| Research Scenario | Weak Prompt (What Not to Do) | Strong Prompt (What to Do) |
|---|---|---|
| Contract dispute | 'What happens when someone breaks a contract?' | 'Summarize the key elements for proving breach of contract in New York commercial disputes, focusing on cases involving service agreements over $50,000, decided after 2016.' |
| Employment law | 'Tell me about wrongful termination.' | 'List the legal standards courts apply to wrongful termination claims under California FEHA for employees terminated during a reduction in force. Format as numbered factors.' |
| Intellectual property | 'Can I use someone else's photo?' | 'Summarize fair use doctrine under U.S. copyright law as applied to business marketing materials. Include the four-factor test with a brief explanation of each factor.' |
| Real estate | 'What are landlord obligations?' | 'What are a commercial landlord's implied warranty of habitability obligations under Texas law for office space leases? Cite relevant statutes if known.' |
| Privacy/data | 'What is GDPR?' | 'Summarize GDPR Article 17 right-to-erasure obligations for a U.S.-based SaaS company serving EU customers. What triggers the obligation and what are the documented exceptions?' |
| Litigation research | 'Find cases about negligence.' | 'Identify the key factors courts in the Seventh Circuit use to determine whether a duty of care exists in premises liability cases involving business invitees. Summarize in bullet points.' |
Verifying What AI Gives You
AI tools, including premium platforms like Westlaw AI, can produce outputs that sound authoritative but contain errors. The most dangerous type is a fabricated case citation: a plausible-looking case name, court, and year that simply does not exist. This has happened in real court filings, resulting in sanctions and professional embarrassment. Verification is not optional. It is the professional standard. Every case name, statute reference, and regulation number produced by AI must be independently confirmed in a primary source before use.
The verification workflow does not have to be slow. For case citations, run the name directly through Westlaw, LexisNexis, or Google Scholar. For statutes, cross-check against the official government source, state legislature websites or the U.S. Code at uscode.house.gov. For regulatory information, check the relevant agency's official site (IRS, EEOC, SEC, etc.). Build this two-step habit: AI generates the map, primary sources confirm the territory. Once you internalize this rhythm, AI research becomes dramatically faster without sacrificing accuracy.
- Copy the case name exactly as the AI provided it
- Search it in Westlaw, LexisNexis, or Google Scholar (scholar.google.com is free)
- Confirm the court, date, and holding match what the AI described
- Check that the case has not been overruled using Shepard's (Lexis) or KeyCite (Westlaw)
- For statutes, go directly to the official government source, not a legal blog or summary site
- For regulations, check the relevant federal agency website or the Code of Federal Regulations at ecfr.gov
- Note the verification date on your research file so you can re-check if the matter extends over months
| Source Type | AI Can Help With | Must Verify At | Free or Paid |
|---|---|---|---|
| Federal case law | Summaries, key holdings, relevant factors | Google Scholar, Westlaw, LexisNexis | Free (Scholar) / Paid (WL, Lexis) |
| State case law | Issue spotting, leading cases by state | State court websites, Westlaw, Fastcase | Varies by state / Paid |
| Federal statutes | Plain-language explanations, section summaries | uscode.house.gov | Free |
| State statutes | Overview of statutory framework | State legislature official websites | Free |
| Federal regulations | Regulatory summaries, key definitions | ecfr.gov, agency websites (IRS, EEOC, SEC) | Free |
| Secondary sources (law reviews, treatises) | Background understanding, issue framing | Law school library databases, SSRN | Free (SSRN) / Institutional access |
| International law | High-level summaries only, use with caution | Official treaty databases, foreign government sites | Varies |
Google Scholar: The Free Verification Tool You Already Have
Sample Prompt: Employment Law Issue Spotting
Prompt
I'm an HR manager at a 200-person technology company in Illinois. A senior engineer filed a complaint alleging they were passed over for promotion three times because of their age (they are 58). They have 14 years of tenure and strong performance reviews. The promotions went to employees aged 28, 31, and 34. I need to understand the legal framework our legal team will be working under. Please summarize the key legal standards under the Age Discrimination in Employment Act (ADEA) that apply to promotion decisions, the types of evidence courts consider most significant in these cases, and what documentation an employer typically needs to demonstrate a legitimate, non-discriminatory reason for its promotion decisions. Format your response as three labeled sections.
AI Response
**ADEA Legal Standards for Promotion Decisions** The Age Discrimination in Employment Act (ADEA) prohibits employers with 20 or more employees from discriminating against workers aged 40 and older in any term or condition of employment, including promotions. Under the McDonnell Douglas burden-shifting framework, the standard courts apply in most ADEA cases, an employee establishes a prima facie case by showing: (1) they are 40 or older, (2) they were qualified for the position, (3) they were not promoted, and (4) a substantially younger person received the promotion. The burden then shifts to the employer to articulate a legitimate, non-discriminatory reason. **Evidence Courts Consider Most Significant** Courts examine the comparative qualifications of candidates, the consistency of the promotion criteria applied, statements made by decision-makers during the process, statistical patterns in promotion outcomes by age, and whether the stated reason for the decision is supported by contemporaneous documentation. In the Seventh Circuit (which covers Illinois), courts also scrutinize whether the employer's stated reason changed over time, called 'shifting justifications', as this can suggest pretext. **Documentation the Employer Needs** To defend a promotion decision, employers should maintain: written job postings with defined criteria, scoring rubrics or evaluation forms completed at the time of the decision, interview notes, contemporaneous records of who participated in the decision, and documented business reasons tied to specific competency gaps. Post-hoc justifications, reasons constructed after a complaint is filed, are given significantly less weight by courts. All documentation should be preserved immediately upon learning of the complaint.
Using AI Across Different Legal Research Contexts
AI research tools are not one-size-fits-all. The way you use ChatGPT to prepare for a client intake meeting is different from how you use Westlaw AI to pull case law for a brief, and different again from how you use Lexis+ AI's contract analyzis feature to flag risk in a vendor agreement. Matching the tool to the task, and understanding what each tool is optimized for, is what separates professionals who get real time savings from those who get frustrated and abandon the technology after two tries.
For background research and issue spotting, general AI tools like Claude Pro or ChatGPT Plus work well because the task is conceptual, you're building a map of the legal landscape before you need precision. For primary source research and cite-checking, you need a legal-specific database tool like Westlaw AI or Lexis+ AI because they are connected to live, curated legal databases. For contract review, tools like ContractPodAi, Ironclad AI, or Lexis+ AI's contract features are specifically trained on contract language and clause structures. Use the right tool for the right job.
| Task Type | Best Tool(s) | What AI Does Well Here | What You Still Do |
|---|---|---|---|
| Background issue spotting | ChatGPT Plus, Claude Pro | Explains legal concepts, identifies relevant areas of law, frames the issue | Confirm jurisdiction-specific accuracy, add case context |
| Primary case law research | Westlaw AI, Lexis+ AI | Surfaces relevant cases, summarizes holdings, identifies key factors | Verify citations, check for overruled cases, apply to facts |
| Statute and regulation lookup | Westlaw AI, ChatGPT Plus + verification | Summarizes statutory language in plain English, identifies key provisions | Confirm against official government source, check amendment history |
| Contract review and redlining | Lexis+ AI, ContractPodAi, Ironclad AI | Flags unusual clauses, identifies missing standard provisions, compares to market norms | Evaluate risk tolerance, negotiate positions, make final call on acceptance |
| Client-facing plain language summaries | Claude Pro, ChatGPT Plus | Translates legal concepts into accessible language for non-lawyers | Review for accuracy, add jurisdiction-specific caveats, sign off on advice |
| Deposition or hearing prep | ChatGPT Plus, Claude Pro | Generates potential questions, identifies factual gaps, summarizes prior testimony | Apply case strategy, assess witness credibility, make tactical decisions |
| Compliance checklists | ChatGPT Plus, Claude Pro, Copilot | Drafts structured checklists based on regulatory frameworks | Verify against current regulations, customize to client industry and size |
Never Use Free Consumer AI Tools With Client Information
Practice Task: Research a Legal Issue Using AI
Goal: Produce one verified AI-assisted research summary on a real legal issue, with at least one citation independently confirmed and one identified research gap, ready to hand to a supervising attorney or use as a starting point for a client memo.
1. Choose a real or hypothetical legal issue relevant to your work, for example, a non-compete clause dispute, a data breach notification obligation, or a contract termination question. Write it down in one sentence. 2. Open Claude Pro or ChatGPT Plus (use a firm-approved account, or anonymize all details completely). 3. Write a prompt using the four-part formula: state the legal issue clearly, specify the jurisdiction (pick a real state or federal context), include 2-3 relevant facts, and specify your desired output format (bullet list, numbered factors, or labeled sections). 4. Submit the prompt and read the full response. Highlight or note any specific case names, statute numbers, or regulatory citations the AI provides. 5. Take one of those citations and verify it using Google Scholar (scholar.google.com) or the relevant government website. Confirm the source exists and the AI's description of it is accurate. 6. Identify one gap or limitation in the AI's response, something it couldn't answer, a jurisdiction it didn't address, or a caveat it added. Note what additional research step would be needed to fill that gap.
AI Legal Research: Quick Reference Cheat Sheet
- Always include: issue + jurisdiction + facts + output format in every research prompt
- Best for background research: ChatGPT Plus, Claude Pro
- Best for primary source research: Westlaw AI, Lexis+ AI
- Best for contract review: ContractPodAi, Ironclad AI, Lexis+ AI contract features
- Free citation verification: Google Scholar (scholar.google.com)
- Free statute verification: uscode.house.gov (federal), state legislature websites (state law)
- Free regulation verification: ecfr.gov, relevant federal agency websites
- Never paste client names or case numbers into free consumer AI tools
- Hallucination check: verify every case name before it appears in any document
- Overruling check: use Shepard's (Lexis) or KeyCite (Westlaw) for any case you plan to cite
- Cutoff dates: ChatGPT and Claude have training cutoffs, use Westlaw/Lexis for current law
- Output formats that work best: numbered factors, labeled sections, bullet lists, comparison tables
- Scope limiters that improve results: date ranges, specific courts, party types, dollar thresholds
- AI does not replace judgment: strategy, client advice, and professional sign-off remain yours
Key Takeaways From This Section
- Prompt quality is the primary driver of AI research quality, structure every prompt with issue, jurisdiction, context, and format
- Different tools serve different tasks: general AI for background understanding, legal-specific platforms for primary sources, specialized tools for contract review
- Every AI-generated citation must be independently verified before it appears in any professional document
- Free consumer AI tools are not appropriate for confidential client matters, use firm-approved platforms or fully anonymize inputs
- AI compresses the early stages of research dramatically; professional judgment on application and strategy remains entirely your responsibility
AI legal research tools are only as good as the professional using them. Knowing which tools to trust, how to verify what they produce, and where human judgment is non-negotiable separates cautious, effective adoption from costly mistakes. This section is your reference sheet, keep it open the first time you use AI for a real research task.
7 Things Every Legal Professional Must Know About AI Research
- AI tools hallucinate citations, they can generate plausible-sounding but completely fictitious case names, docket numbers, and holdings.
- No AI tool replaces Westlaw or Lexis for authoritative, up-to-date primary source verification.
- General-purpose AI (ChatGPT, Claude) is better for drafting and summarizing than for cite-checking.
- Legal-specific tools (Westlaw AI, Lexis+ AI, Harvey) are trained on verified legal databases and carry far lower hallucination risk.
- Jurisdiction matters, always specify the jurisdiction in every research prompt.
- AI output is a starting point, never a filing-ready product.
- Confidentiality rules apply, never paste client names, case facts, or privileged information into a consumer AI tool.
Understanding AI Hallucination in Legal Contexts
Hallucination is the term for when an AI generates content that sounds authoritative but is factually wrong. In legal research, this is acutely dangerous. A 2023 federal court case. Mata v. Avianca, saw a lawyer sanctioned after filing a brief containing six AI-generated citations that did not exist. ChatGPT had invented the case names, courts, and holdings with complete confidence. The court fined the attorneys $5,000 and required them to notify the judges whose names had been falsely cited.
Hallucination happens because large language models predict plausible text, they are not running a live database query. When asked for a case on a narrow legal issue, the model may construct a citation that fits the pattern of real citations rather than retrieve one that actually exists. The solution is simple but non-negotiable: every case cite produced by a general AI tool must be verified in Westlaw, Lexis, or Google Scholar before use.
- Always paste a generated citation into Westlaw, Lexis, or Google Scholar to confirm it exists.
- Check that the holding matches what the AI described, paraphrasing errors are common.
- Confirm the case has not been overruled using KeyCite or Shepard's Citations.
- Be especially skeptical of very specific citations (exact page numbers, narrow holdings), these are higher hallucination risk.
- If you cannot verify a citation, delete it and find a real one manually.
The 30-Second Verification Rule
| AI Tool | Best For | Hallucination Risk | Verify Citations? | Cost |
|---|---|---|---|---|
| ChatGPT Plus | Summarizing, drafting memos, plain-language explanations | High for citations | Always | $20/month |
| Claude Pro | Long document analyzis, contract review, research memos | High for citations | Always | $20/month |
| Westlaw AI (CoCounsel) | Case law research, brief drafting, cite verification | Low, database-grounded | Still recommended | Subscription |
| Lexis+ AI | Case research, Shepardizing, statutory analyzis | Low, database-grounded | Still recommended | Subscription |
| Google Scholar | Free case law lookup, citation verification | Not generative, lookup only | N/A | Free |
Prompt Precision: Getting Jurisdiction and Issue Right
Vague prompts produce vague legal summaries. The single most effective upgrade to any legal research prompt is adding three things: the jurisdiction, the specific legal issue, and the intended output format. A prompt that says 'tell me about non-compete agreements' will return a generic overview. A prompt that specifies California, the enforceability standard under Business and Professions Code §16600, and asks for a summary memo returns something immediately useful.
Framing the output format matters too. Asking for a 'bullet-point summary of the key elements a plaintiff must prove' produces a checklist you can use in client intake. Asking for a 'comparison of majority and minority rule jurisdictions in a two-column table' produces a reference tool. The AI follows your format instructions reliably, use them every time.
- State the jurisdiction explicitly: 'Under New York law...' or 'In the Ninth Circuit...'
- Name the specific statute, rule, or doctrine if you know it.
- Specify the output format: memo, bullet list, table, timeline, checklist.
- State your role context: 'I am a paralegal preparing a client summary...'
- Set a length constraint: 'in 300 words or fewer' keeps output scannable.
- Ask for limitations explicitly: 'Note any areas where the law is unsettled or jurisdiction-specific.'
| Weak Prompt | Strong Prompt | Why It's Better |
|---|---|---|
| What is negligence? | Summarize the four elements of negligence under Texas tort law in a numbered list suitable for a client explainer. | Jurisdiction-specific, formatted, audience-aware |
| Find cases about wrongful termination. | List the key legal standards for wrongful termination in Illinois, focusing on public policy exceptions. Note any recent developments after 2020. | Scoped issue, jurisdiction, recency filter |
| Explain GDPR for our privacy policy. | Summarize the GDPR requirements most relevant to a U.S.-based SaaS company with EU customers, in plain language, under 250 words. | Audience, scope, length, all specified |
Free Tool for Citation Verification
Prompt
I am a paralegal at a small employment law firm in California. A client was fired one week after reporting workplace safety violations to OSHA. Draft a concise research memo summarizing the legal protections against retaliation for OSHA complaints under California law, including relevant statutes and the standard a plaintiff must meet to establish a prima facie case. Format it as a short memo with headers. Note any areas where the law is unsettled.
AI Response
MEMORANDUM TO: File FROM: AI Research Assistant RE: Retaliation Protections for OSHA Complaints. California **Applicable Law** California employees who report workplace safety violations are protected under two primary frameworks: (1) Section 11(c) of the federal Occupational Safety and Health Act (29 U.S.C. § 660(c)), enforced by OSHA, and (2) California Labor Code § 6310, which provides broader state-level protection and allows for private civil actions. **Prima Facie Case. Elements** To establish a retaliation claim under Cal. Labor Code § 6310, a plaintiff generally must show: 1. The employee engaged in protected activity (e.g., reporting a safety violation to OSHA or internally). 2. The employer knew of the protected activity. 3. The employer took an adverse employment action. 4. A causal connection exists between the protected activity and the adverse action. Temporal proximity, termination one week after the complaint, is strong circumstantial evidence of causation. **Burden-Shifting** California courts apply a burden-shifting framework. Once the employee establishes a prima facie case, the burden shifts to the employer to articulate a legitimate, non-retaliatory reason for the termination. **Unsettled Areas** The interaction between federal OSHA remedies and California state court claims involves procedural complexity, specifically, whether filing a federal OSHA complaint affects the timeline for state civil action. This area benefits from case-specific analyzis. **Recommended Next Step** Verify current case law interpreting § 6310 in Westlaw, particularly post-2020 decisions addressing the causal connection standard.
Confidentiality and Ethical Guardrails
Bar association ethics opinions are increasingly clear on one point: attorney-client privilege and confidentiality obligations do not pause because you are using a technology tool. Inputting a client's name, identifying facts, or privileged communications into a consumer AI platform, one that may use inputs to train future models, creates a real professional responsibility risk. Several state bars, including New York, California, and Florida, have issued formal guidance on AI use that emphasizes confidentiality compliance.
The practical solution is not to avoid AI, it is to anonymize inputs before they reach the tool. Replace client names with 'Client A,' substitute generic industry descriptors for identifying business details, and use hypothetical framing: 'In a scenario where an employee was terminated one week after...' This preserves the research value while keeping protected information off third-party servers. Enterprise versions of AI tools (Microsoft Copilot for Microsoft 365, Westlaw AI) typically offer data-isolation agreements, check with your firm's IT or compliance team.
Never Paste Privileged Information Into a Consumer AI Tool
Goal: Use a free AI tool to research a real legal question, verify the output, and produce a one-page reference summary, without using any client information.
1. Choose a legal topic relevant to your current work, pick something general, not client-specific (e.g., 'enforceability of non-solicitation agreements in Texas'). 2. Open Claude.ai (free) or ChatGPT (free tier) in your browser. 3. Write a structured prompt: state your role, the jurisdiction, the specific issue, and ask for a numbered summary of key legal standards. Request that the AI flag any unsettled areas. 4. Read the full response and highlight any case citations or statute numbers it mentions. 5. Open Google Scholar (scholar.google.com) and search for each citation. Confirm the case exists, the court is correct, and the holding matches what the AI described. 6. Copy the verified, accurate portions of the AI summary into a Word document. Delete or flag anything you could not verify. Add a header noting the date and tool used.
AI Legal Research Cheat Sheet
- Always specify jurisdiction, legal issue, and output format in every prompt.
- Verify every AI-generated citation in Google Scholar, Westlaw, or Lexis before using it.
- Use Shepard's or KeyCite to confirm cited cases have not been overruled.
- Treat AI output as a first draft, not a final product.
- Never input client names or privileged facts into a consumer AI tool, anonymize everything.
- Legal-specific tools (Westlaw AI, Lexis+ AI) carry lower hallucination risk than general AI.
- Ask AI to flag unsettled areas of law, it often knows where doctrine is contested.
- Use free Google Scholar to verify citations at no cost.
- Check your state bar's ethics opinion on AI before using it for client matters.
- Date-stamp all AI-assisted research documents and note which tool was used.
Key Takeaways
- AI hallucination is a documented, serious risk in legal research, citation verification is mandatory, not optional.
- Prompt precision (jurisdiction, issue, format) dramatically improves the quality of AI legal research output.
- General AI tools are useful for summarizing, drafting, and explaining, not for authoritative cite-checking.
- Legal-specific platforms like Westlaw AI and Lexis+ AI are grounded in verified databases and are safer for primary research.
- Confidentiality obligations apply fully to AI use, anonymize all inputs on consumer platforms.
- AI legal research saves significant time on first-draft work; human judgment remains essential for strategy, advice, and filings.
This lesson requires Pro+
Upgrade your plan to unlock this lesson and all other Pro+ content on the platform.
You're currently on the Free plan.