Below is an expanded version of the guidelines for using ChatGPT for legal research—with sample interactions for each section that illustrate how an attorney might set up guardrails against hallucinations and verify citations. When using ChatGPT’s public model you should never upload case related documents that are not public record. You should never use prompts which contain confidential information. Your uploaded data and prompts are likely used to further train the model.
1. Understand the Tool’s Limitations
Guideline Recap:
Recognize that GPT models can “hallucinate” details or citations and that their knowledge may be outdated or incomplete. Always verify any legal information or citation provided by the AI.
Sample Interaction:
Attorney:
“ChatGPT, please summarize the key legal principles from Roe v. Wade.”
ChatGPT:
“Roe v. Wade is a landmark decision that established a woman’s right to choose, setting guidelines for state regulation of abortion based on trimester frameworks. (Note: Please verify the details and citation in a reliable legal database.)”
Attorney (Internal Note):
“I need to check this summary and the details on Westlaw or LexisNexis, as ChatGPT’s information might be incomplete or outdated.”
2. Use ChatGPT as an Assistant, Not a Sole Source
Guideline Recap:
Use ChatGPT to generate ideas, summaries, or initial research drafts. Always confirm the results with trusted legal databases or texts.
Sample Interaction:
Attorney:
“ChatGPT, can you provide an overview of the common elements required to prove breach of contract?”
ChatGPT:
“To establish a breach of contract, a plaintiff generally must prove the existence of a valid contract, performance or tender of performance by the plaintiff, a breach by the defendant, and resulting damages. This is a basic overview—please consult a legal database for detailed elements specific to your jurisdiction.”
Attorney (Internal Note):
“I’ll use this overview as a starting point and then verify each element against an authoritative source like Westlaw.”
3. Set Up Guardrails Against Hallucinations
Guideline Recap:
Instruct ChatGPT to include sources or note that outputs need verification. Use follow-up queries to ensure the response is grounded in verifiable information.
Sample Interaction:
Attorney:
“ChatGPT, provide a case citation that supports the proposition that ‘implied consent can be inferred in cases of negligence.’ Please include any source details.”
ChatGPT:
“One commonly cited case on implied consent in negligence is ‘Doe v. Smith, 456 U.S. 789 (2001)’. However, I must note that this citation should be verified in an official legal database, as the information provided here might be based on patterns in the training data rather than direct sources.”
Attorney (Internal Note):
“I’ll look up Doe v. Smith on LexisNexis to ensure the citation is accurate and contextually appropriate.”
4. Check and Verify All Citations
Guideline Recap:
Always cross-check any citations or legal references provided by ChatGPT using reputable legal research tools.
Sample Interaction:
Attorney:
“ChatGPT, summarize the holding of Brown v. Board of Education and provide the full citation.”
ChatGPT:
“Brown v. Board of Education is a landmark decision that declared state laws establishing separate public schools for black and white students to be unconstitutional. The full citation is ‘Brown v. Board of Education, 347 U.S. 483 (1954)’. Please verify this citation in your legal research platform.”
Attorney (Internal Note):
“I will verify the citation and holding details on Westlaw to ensure everything aligns with the official records.”
5. Incorporate Human Oversight and Peer Review
Guideline Recap:
Use AI-generated research as a draft that is reviewed by colleagues or within your legal team to catch potential errors or hallucinations.
Sample Interaction (Narrative):
After receiving a draft legal memorandum generated with ChatGPT’s assistance, the attorney shares it with a colleague for review:
Attorney:
“Here’s the draft summary and citation list ChatGPT produced for our upcoming brief. Could you take a look and cross-check the citations with our internal legal database?”
Colleague:
“Sure, I’ll verify each citation. It looks like a good starting point, but I noticed one citation might be outdated. I’ll update it based on our latest research findings.”
Attorney (Internal Note):
“This peer review helps ensure that any inaccuracies from the AI are caught before we finalize our memorandum.”
6. Stay Updated and Provide Feedback
Guideline Recap:
Keep abreast of updates to the AI tool and provide feedback to vendors if recurring issues are observed.
Sample Interaction:
Attorney:
“ChatGPT, your summary of the case law regarding ‘implied warranty of fitness’ seems to conflict with recent developments. Can you re-check your data and provide an updated response?”
ChatGPT:
“Based on my current training, the standard summary for implied warranty of fitness is [summary]. However, I advise verifying this against the latest legal updates on your trusted databases. If discrepancies persist, please report them to your AI vendor for further review.”
Attorney (Internal Note):
“I’ll note this feedback and report it through our vendor’s feedback mechanism, ensuring that future updates improve accuracy.”
By following these sample interactions and guidelines, attorneys can effectively use ChatGPT as a valuable assistant while ensuring that every piece of information is rigorously verified and ethically sound.