Imagine yourself as a judge reading a brief. The writing and argument are capable, but you notice citations to cases you have never seen before. Curious, you check for the cases in Fastcase, only to discover there are no cases with those names. Confused, you check other sources with the same result. The cases do not exist. They have been hallucinated, or made up, by a generative AI (GenAI) research tool. This is not an imaginary situation; it has happened numerous times in courts across the country, as most famously reported in the New York Times in May 2023.^[1]

Since then, the GenAI tools have improved, but concerns remain. GenAI tools can be a powerful method for unearthing legal resources that will make your professional life easier, but only when they are used competently and in line with well-established research guidelines. As GenAI tools continue to evolve and more legal professionals adopt them into their daily workflows, it will become even more incumbent on attorneys to understand both the potential and the limitations of these tools and to know what to consider when using them for research.

Legal Research and AI: Ethics and Integration

Competent legal research is a fundamental skill expected of all practicing attorneys. Legal research is an undercurrent flowing through the Rules of Professional Conduct for Attorneys, touching on subjects ranging from meritorious claims to diligence to confidentiality to competence. GenAI tools can make compliance with these rules more complicated by introducing new features that, at times, can seem equally magical and frustrating.

Just as with all technology that lawyers use to do their jobs, GenAI must first be understood and used on a fundamentally competent level. Comment 8 to SCR 20:1.1 explicitly states that lawyers must keep current on changes in the law and practice, “including the benefits and risks associated with relevant technology.” Whether we like it or not, GenAI is a relevant technology that has the capability to improve lawyers’ research skills and cut down on inefficient practices.

In a bid to streamline the time that lawyers and other legal professionals spend on legal research, numerous companies have introduced GenAI tools into their platforms. Thomson Reuters (CoCounsel), Lexis (Protégé), Bloomberg Law (AI Answers and Summarization), and VLex/Fastcase (Vincent) are just the most well-known legal-specific research tools.^[2] Each of these tools can be purchased as an add on (with the exception of Bloomberg Law, which has one single price for all its tools and features) to the already familiar search strategies and content found in each company’s database. Additionally, many of the most popular GenAI tools, such as ChatGPT, Claude, and Gemini, can do basic research, though the returns diminish with tools that are not focused on the complicated area of legal-specific content. As time goes on, these tools will become further ingrained in the platforms as more attorneys familiarize themselves with GenAI.

Researching with GenAI: Start with Effective Prompting

Many attorneys begin digital legal research with a search in their favorite legal research tool. That search could be a terms-and-connectors search, in which you use symbols and tech shorthand to craft searches that provide targeted results, such as “negligent homicide” /p run!, which is a search for resources that have the phrase “negligent homicide” in the same paragraph as any word that starts with the letters R-U-N. Others use natural-language searches, which are simply Google-esque searches in which you include certain keywords or questions that you want the resulting resources to either answer or include. Both legal research strategies remain vital to proper legal research because they can help validate AI results and because some areas of law require more specific human-created searches that remain superior to searches generated by AI.

Legal research with GenAI tools is quite different from these more established methods. Instead of creating a search that brings back a list of results, the goal is to draft a thoughtful and detailed prompt, the result of which is output that is an effective starting point to build on using one’s own expertise. There is an entire cottage industry surrounding the effective creation of these prompts called “prompt engineering.” Clio and Lexis, for example, have suggestions for molding proper prompting to legal work.^[3]

Seven Tips for Prompting

As you begin your work, keep in mind these prompting tips (the seven Ps) to get the most out of your GenAI tool of choice:

1. Persona: Determine the perspective and tone you want the AI to take. Do you want cases that help the defense or the prosecution? Do you want the resulting output to be legalistic or something a layperson could read? This will help you and others more quickly build off the result down the line.

2. Product: Tell the AI what the final product should be. Do you want a list of cases, statutes, or secondary sources? Do you want them listed in a research memo or just a bulleted list with key passages excerpted? Again, this gives you a firmer foundation on which to build and will save you from further refining later.

3. Prompt: Clearly ask the AI to perform a specific task. This can be as simple as one action verb, such as “research” or “analyze” or “draft.”

4. Purpose: Give the AI a reason for the prompt. Explaining the why, such as “the cases will be used to provide a defense against the castle doctrine,” will give more nuance to the AI and make it more likely to highlight cases that are relevant to that aspect of the research.

5. Prime: This is all about context. Give the AI an appropriate amount of details, such as where the case will be argued and basic facts of the case that could help the AI find on-point and similar cases.

6. Privacy: Do not overshare. Do not include a client’s name or other personally identifiable information. Only share relevant information that you would be comfortable sharing in a public filing.

7. Polish: Continue the conversation with the AI to refine the results or the output. Anything that AI generates must be vetted and reviewed for accuracy. While each of these seven tips is important, this might be the most important one.

Anyone can feed a question into an AI tool and get an answer. However, if you keep these seven Ps in mind, you will not only get a better answer but also get it faster without constant and continued refinements.

GenAI Legal Research Red Flags

With continued practice in creating effective AI prompts, benefits of using AI will soon appear. AI tools can, in seconds, summarize cases relevant to a fact pattern in a particular jurisdiction, answer questions with cited and sourced case law, or point toward a secondary source that examines a situation in further depth. The time saved is an immediate and obvious benefit.

However, the job is not yet done. Lawyers are still responsible for their work (and for the work of people they supervise) – the AI is not signing off on the research. That means you must check that the cases AI provided and the assertions AI made are all legitimate. Some AI experts have flipped the old saying “trust but verify” on its head when dealing with generated content. You must verify, then trust.

Warning Signs of Inaccurate Results

Some warning signs that the information generated is not accurate are the following:

Uncited statements or unlinked case names:The major legal database AI tools will provide links to the cases they locate via AI searching. However, there are times when a case will not be linked. This is a very clear sign that the case does not exist. Or there might be a statement, such as “In fair use cases, the Seventh Circuit uses a five-factor test,” with no cite. This assertion must be checked for validity because the absence of a cite suggests that it is not true.

Cites to headnotes or bad law: Even if a case is on point and is linked, the AI tool could be inappropriately reaching beyond reasonable sources to answer the user’s question. The AI tools may read headnotes or overturned cases as legitimate authority for an argument. The person doing the research must check each case to make sure that the “winning” point that the AI has made and that seems to be perfect for the situation is from an opinion, not merely a headnote, and that the opinion still carries precedential value.

Mild hallucinations or logistical confusion:According to a 2024 study, one in six cases that Westlaw and Lexis provided with their GenAI tools was misleading.^[4] What this means is that the cited case did exist, but it did not provide support for the argument that the AI tool was making or was about an entirely different subject, rendering the research a waste. Other times, the AI relied on persuasive authority exclusively. Once again, checking the cases to make sure they are on point and relevant (that is, polishing) is a step that must not be skipped.

The more general the tool, the more likely a problem: Legal research platforms such as Lexis, Westlaw, and Bloomberg have trained their AI tools on their resources and data. This means that those AIs can better understand legal questions and research and respond in kind with case law or statutes (though as we have seen, results can still vary). General tools, such as ChatGPT, are trained on other sets of data, which can potentially provide better overviews, but also dilute results: the AIs become confused about how to best answer a complicated research question. If you research with a general tool (that likely costs less), be aware that even more due diligence on your part will be required, concerning both the accuracy of the results and the privacy of the information entered into the prompt.

Answers that seem too good to be true: One recurring problem that has persisted through the various iterations of GenAI tools is a tendency to be overly agreeable with the user.^[5] This eager-to-please behavior could result in the human user missing important counterarguments or believing that one has a winning case when the case is filled with holes. The AI-competent legal researcher can overcome this shortcoming by approaching results with skepticism and using the AI-generated output as an efficient method for quickly gathering a first round of results, which can then serve as scaffolding on which to set confirmed research findings.

Researching Smarter with GenAI: Critically Assess Results and Enhance Skills

When used smartly, AI can make you a better legal researcher. Researchers who take the time to draft a thoughtful initial prompt using the seven Ps will find that the results will be more focused and produced more quickly than with many traditional digital research strategies. However, these tools are not currently capable of replacing a lawyer’s eye for detail or nuance. The results will likely save time in locating relevant sources, but lawyers still must check the results for accuracy and relevancy. Viewing the generated output critically will guarantee that you do not get caught off guard with a case that does not support your argument or, worse, does not exist. Continue to hone and practice your skills with natural-language and terms-and-connectors searches.

As these tools enter further into the legal research mainstream, take the time to try it out on an area of law that you know well. Compare the cases and resources a GenAI tool provides with what you already know. You will likely find that the content is a good starting point, and one that a competent legal researcher could use to build a winning case.

Ignoring these new tools is not the answer. In doing so, you risk becoming a less efficient and more expensive lawyer. Instead, take full advantage of the powerful tools that can, when properly used and vetted, boost you to a higher level of efficient legal research.

Also of Interest
Keep Online Business on the Move

The internet has revolutionized commerce, making it easier than ever for businesses to reach customers, sign contracts, and conduct transactions across state and national borders. But with convenience comes complexity. Businesses operating online must contend with an intricate legal landscape to avoid liability, protect intellectual property, and stay compliant with evolving regulations.

The Law of Business on the Internet, newly published by State Bar of Wisconsin PINNACLE, is a vital resource for attorneys advising clients on internet-based transactions. This comprehensive guide distills crucial principles from state and federal laws, providing the tools to counsel businesses confidently.

This book covers key legal issues that arise in online business operations, including enforceability of online contracts, jurisdictional challenges, intellectual property considerations, data privacy and cybersecurity, website content and disclaimers, and AI and commercial contracting. You’ll also find practical insights into online advertising regulations, tax considerations, and best practices for commercial email communications.

The Law of Business on the Internet includes sample website disclaimer language, a detailed appendix with links to internet-law references, and in-depth discussions on data breach lawsuits and AI-driven commercial contracting.

Ensure you have the latest knowledge to proactively guide clients doing business in the digital age.

wisbar.org/AK0447

Endnotes

[1] Here’s What Happens When Your Lawyer Uses ChatGPT, N.Y. Times (May 27, 2023), https://www.nytimes.com/2023/05/27/nyregion/avianca-airline-lawsuit-chatgpt.html (behind paywall for some readers).

[2] More information about these tools is available at the following websites: CoCounsel, https://www.thomsonreuters.com/en/cocounsel; Protégé, https://www.lexisnexis.com/en-us/products/protege.page; Bloomberg Law Platform, https://pro.bloomberglaw.com/products/ai-and-bloomberg-law/#overview; Vincent AI, https://vlex.com/vincent-ai.

[3] Clio, Chatgpt for Lawyers: How to Write Better Prompts, https://www.clio.com/resources/ai-for-lawyers/resources-ai-for-lawyers-write-better-prompts/; Lexis, Prompting Best Practices: Lexis+ AI™, https://www.lexisnexis.ca/pdf/2024/5p-prompting-methodology.pdf.

[4] Faiz Surani & Daniel E. Ho, AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries, Stanford HAI (May 23, 2024), https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries.

[5] Open Tools, OpenAI’s Charm Offensive: Why ChatGPT Became Too Agreeable (last updated May 1, 2025), https://opentools.ai/news/openais-charm-offensive-why-chatgpt-became-too-agreeable.