AI and The Sorcerer’s Apprentice
An essay by Edward J. Kionka1
According to the legal news, all of us in the legal profession — lawyers, law firms, judges, law clerks, paralegals, law students, law professors, administrators — are rushing to hop on the AI2 bandwagon. Perhaps some of us are trying to use it too much, too soon.
Do you remember the story of the sorcerer’s apprentice? It originated in a 1797 poem by Goethe. As an old sorcerer leaves his workshop, he assigns his apprentice to fetch water. Tiring of making trips carrying a heavy pail, the apprentice enchants a broom to do the work for him, using magic in which he’s not fully trained. The floor soon floods, and the apprentice realizes he can’t stop the broom because he doesn’t know the magic required. He splits the broom in two with an ax, but each piece becomes a whole broom, and both brooms continue fetching water. The entire room floods. When all seems lost, the old sorcerer returns and breaks the spell, admonishing the apprentice that only a master should invoke powerful spirits. (Walt Disney’s animated short, Fantasia3, popularized Goethe’s story to a worldwide audience.)
Are we like the apprentice, with AI as the magic?
Learning AI is difficult and time-consuming. I’m not sure how solo or small-firm lawyers will find the time and patience to master it. AI is rapidly becoming part of law school curricula, so perhaps the next generation of lawyers will be up to speed early on. BigLaw is spending big bucks to develop proprietary AI platforms (see Appendix). The Illinois Supreme Court appointed an AI Task Force. The Illinois State Bar Association has a Standing Committee on Artificial Intelligence and the Practice of Law, and its steering committee has begun vetting AI apps. The entire Illinois court system, from the top down, is using AI.
The simple truth is that whoever you are — law student, lawyer, paralegal, judge, law teacher, legal administrator — you have no choice. You must learn to use the appropriate AI tools and apply them to your work as best you can.
Some experts predict we will soon have AGI (artificial general intelligence) that can understand, learn, and apply knowledge across any intellectual task, just like a human. After that, they say, will come ASI (artificial superintelligence), a theoretical AI that would surpass human intelligence and capabilities across all fields, including creativity, general wisdom, and problem-solving. If and when ASI arrives, we’re all in trouble. Robots will be the next species, and they won’t need humans.
At present, however, we have only narrow AI (ANI), which performs specific, predefined tasks, such as facial recognition, voice assistants, and playing chess. For our purposes, these tasks include retrieving information (legal research, data collection), drafting documents (pleadings, contracts, legal briefs and memoranda), and analyzing and summarizing documents.
As you surely know, you’re already using ANI every day — for example, when you do a Google search, use a maps app, or ask Siri or Alexa for information.
When I started using Westlaw for legal research, we had to use Boolean syntax for queries. Today, we can use full natural-language sentences. You can enter a narrow legal question (1-2 sentences) in the main search bar and get a direct answer with hyperlinked citations. You can also use Westlaw Deep Research.4 Lexis+ offers similar capabilities.
Thomson Reuters also offers Westlaw’s Co-Counsel Legal and Lexis offers Protégé, both of which they claim can create timelines, gather surveys across jurisdictions, analyze and update contract terms, outline arguments, identify affirmative defenses, draft, summarize, or compare documents, find potential claims, research legal questions, identify contradictions in witness statements, summarize filing deadlines, adjust the tone of a draft document, and suggest courses of action. vLex Fastcase integrates AI-driven legal tools, including AI drafting tools such as Clearbrief, allowing lawyers to access case law while drafting in Microsoft Word. vLex Fastcase is available free to ISBA members.
Popular standalone AI tools that are not legal-specific include ChatGPT, Claude, Gemini, Perplexity, Copilot, and others. Gemini is Google’s AI assistant, and Copilot is Microsoft’s. If you use Google or Microsoft apps, you already have access to Gemini and Copilot. Apple’s Siri is a basic AI, but Apple plans to make it a more capable generative AI through Apple Intelligence and partnerships with platforms like Gemini.
Legal-specific AI plugins and purpose-built applications — for example, CoCounsel (by Thomson Reuters), Claude for Legal (from Anthropic), and Spellbook (Microsoft), which runs inside Microsoft Word — are now available, along with many others. OpenAI will soon offer legal-specific tools.
NotebookLM is a special-purpose AI app that works only with the data you supply. This is hugely important because using open tools such as ChatGPT and including actual client or case information in your submission can lead to improper disclosure of confidential or privileged information and raise serious professional responsibility issues.
The Illinois Supreme Court adopted a Policy on Artificial Intelligence, effective January 1, 2025.5 Briefly, the use of AI is permitted and encouraged, provided the user complies with existing legal and ethical rules (see Appendix).
Although AI characteristics can be classified in multiple ways, for our purposes, we will focus on two: generative AI and traditional AI.
-
Generative AI creates new, original content — text, images, video, audio, or code — from user prompts.
-
Traditional or Classic AI relies on programmed rules and logic. Rather than creating new content, it performs specific tasks such as legal research, finding other information, making predictions, classifying data, or recognizing patterns using existing information. It’s like a research assistant that can also summarize, organize, and perform similar tasks.
So what’s not to like?
First and foremost, AI is too prone to errors and hallucinations. This is especially true of generative AI. In tests of generative AI in legal contexts, hallucination and error rates ranged from 17% to 33%.6
Here’s an example using a recent event and traditional AI.
In an April 2026 game, Joe Adell, the Los Angeles Angels’ right fielder, made three amazing over-the-wall catches that robbed the Seattle Mariners of three home runs.7 The Angels won 1-0. Did Joe have the greatest defensive game in MLB history? A sports writer asked one of the most prominent AI tools. The AI’s response was negative; the greatest was Ozzie Smith’s performance against the Dodgers in 1978, when he recorded 13 assists at shortstop, “a modern-era record.” The response went on to elaborate on the game and his performance. It sounded so credible.
Except it was totally false. No such game was played. And Ozzie never had 13 assists in any game that year. In fact, Ozzie never had 13 assists in any game he played in the big leagues.
The bot also described “other famous defensive masterpieces worth mentioning.” However, the details of another supposed defensive gem were significantly inaccurate.
To see examples of hallucinations and errors in generative AI, one need only read the legal news. Hardly a day goes by without a story about a lawyer or law firm that was sanctioned — and embarrassed — when the court discovered that a case cited and even discussed did not exist. A database recently identified more than 1,000 U.S. legal cases with fabricated citations, misquoted authority, or other AI-generated legal errors.
When I read these stories, my first thought is, “Where did you go to law school?” In the first semester of my first year of law school, my legal writing instructor taught: “Never cite a case you haven’t read.” Despite some lawyers’ protests to the contrary (which unfortunately lack credibility), there is no doubt that these fake citations and discussions were produced by generative AI.
Another problem with using generative AI to produce a draft of your brief, memo, or opinion letter is that, even if the authorities are accurate and authentic, and the draft reads well, it may not be the best you can do. It may even be defective. You won’t realize that because you didn’t marshal the facts and the law yourself and you didn’t use your brain to craft the writing.
Good legal writing — briefs or other documents — is part education, part training, part skill, but also an art. Words are the lawyer’s tools; you are a wordsmith. Words can be combined in countless ways. Robots aren’t the best legal writers, yet. You can do better.
Recently, I wrote a brief that required a line-by-line analysis of the record. Doing so, I noticed something that led me to a new way to frame an argument. It turned out to be a winner. If I had delegated that part of the brief to AI, I’m quite sure that argument would never have been made.
In another appeal, the circuit court’s ruling rested solely on estoppel. I wanted to make a laches argument. I found a brief reference to laches in the administrative hearing transcript. Browsing cases, I found an Illinois Supreme Court opinion that opened the door for me to argue laches in my brief. I don’t trust that my robot would have found that path.
That’s what it comes down to — trust. In several recent CLE courses, the instructor advised treating AI like a student law clerk or a first-year associate. Use it to gather information but verify the results. Do the writing yourself. Use AI research as a springboard for your own work. That seems like sound advice.
Using AI in brief writing
Rather than tell you what you should do (that’s up to you), I will tell you how I use it.
Jeff Su, the instructor in many YouTube videos (four are cited in the Appendix), uses an airplane cockpit analogy to illustrate three ways to use AI.
First is autopilot mode. You hand the task to AI with clear instructions and trust the output with minimal review. The AI handles everything on its own. Think level flight.
In handling an appeal, there is not much in this category. Possibly looking up rules to confirm dates or requirements, finding a single case by its citation, finding something in the record where a basic search won’t do, simple tasks like that. For example: “How many plays did Shakespeare write?”8 I use it all the time to check words, phrases, and facts.
Some use AI to summarize testimony or documents, but I wouldn’t totally trust it for that job.
Second is collaboration mode. You and AI iterate together through multiple rounds until the output meets your standard. Neither you nor AI could have produced the result alone. Think takeoff and landing.
Legal research is the best example, but this mode can be used for anything short of writing as long as the result can be verified. In their output, some AI tools include a link or reference to the source, a very useful way to verify.
Third is manual mode. You do the work yourself because AI either can’t do it well, or the risk of getting it wrong is too high. Think in-flight emergency.
Manual is my writing mode. I believe the brief I craft myself will be superior to one written by a robot, no matter how good the robot is — at least for now. In other words, I don’t trust a robot to write my brief, not even the first draft.
There is one AI app I use when editing my writing — Grammarly. In my experience, it’s quite good. It has made my writing better and more concise. Other similar apps include Copilot (integrated into Microsoft 365), ProWriting Aid, Hemingway, Linguix, Readable, WordTune, QuillBot, Ginger, and others. But I always work interactively with Grammarly, and I don’t accept all of its suggestions.
A recent CLE course identified skills that AI cannot replace:
Bryan Garner, widely regarded as one of America’s top legal writing gurus, argues that high-level writing tasks — including legal drafting, literary style, and persuasive argument — remain a uniquely human domain where style, tone, and ethical considerations matter. Critical thinking, originality, and ethical communication are integral to writing, and these elements can never be outsourced to AI. Good writers, he says, will always have the edge, especially over those who allow their skills to wither by overrelying on AI.
Aristotle taught that the three principles of persuasive rhetoric are logos (logic), ethos (credibility), and pathos (in our context, your reader wants you to win). As far as I know, ANI will not do ethos and pathos as well as you and I, if at all.
Writing — especially legal writing — is hard work. We are all lazy by nature. It’s our genetic heritage. AI is a tempting shortcut. We may think it can save our client money, make an appeal more affordable, or save our own precious time. But our objectives are to (1) win and (2) provide the maximum possible assistance to the reader, in both content and form. In legal writing, as in all other work, we should use the best tool available at each step. More often than not, that tool is the human brain.
Appendix
BigLaw
-
Kirkland & Ellis is developing a proprietary AI platform and is prepared to spend about $500 million on its AI build over the next three to four years.
-
Other firms are using commercial AI platforms, such as Harvey or Legora, or using AI tools for a specific part of a client matter rather than for the entire matter.
-
Sullivan & Cromwell partnered with a firm to develop AI discovery and deposition assistants.
-
Cleary Gottlieb Steen & Hamilton acquired the LLM developer Springbook AI.
-
White and Case created its own tool.
-
Troutman Pepper Locke developed its own AI chatbot adding agentic AI capabilities, and incorporated Thomson Reuters’ Deep Research.
Illinois Supreme Court Policy and Guidelines
-
Permitted Use: Litigants, attorneys, judges, and court staff are permitted and encouraged to use AI.
-
Ethical Compliance: All use of technology must strictly comply with the Rules of Professional Conduct and the Code of Judicial Conduct.
-
Total Accountability: Attorneys, judges, and self-represented individuals bear ultimate responsibility for their final work product.
-
Mandatory Verification: Users must thoroughly review all AI-generated content for accuracy before submitting it to the court.
-
Required Competence: Before using any tool, users must understand both the general capabilities of AI and the specific functions of the software.
Some AI videos
Footnotes
1 The author is a retired law professor and a not-yet-retired appellate lawyer. See www.KionkaLaw.com. Disclosure: I used AI in researching and writing this essay. All info is believed to be current as of June 2026. © 2026 Edward Kionka
2 If you don’t know the first thing about Artificial Intelligence, my friend, you are at the bottom of a long, steep hill.
3 https://youtu.be/UEEMzrKAeNc
4www.thomsonreuters.com/en-us/help/cocounsel/legal/skills/skills-prompts-workflows/westlaw-deep-research
5ilcourtsaudio.blob.core.windows.net/antilles-resources/resources/e43964ab-8874-4b7a-be4e-63af019cb6f7/Illinois%20Supreme%20Court%20AI%20Policy.pdf.See also www.illinoiscourts.gov/News/1485/Illinois-Supreme-Court-Announces-Policy-on-Artificial-Intelligence/news-detail.
6 https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
7 https://youtu.be/Ru3xR5_IFtQ
8 The correct answer is, we don’t know exactly. Scholars mostly agree that he wrote 37-39, a few in collaboration with other playwrights. We know of two lost plays, Cardenio and Love’s Labour’s Won. Probably there were others. The First Folio contains 36 plays. In this case, AI got it right.
9 https://lawprose.org/bryan-garner
.png)