AI for case-law research: what Accursius's Gloss teaches today's lawyers

An afternoon, a brief to file, a line of case law to find. Asking a chatbot "find me three Supreme Court rulings on this point" is lightning fast. The problem is that the chatbot always finds them, even when they don't exist. This guide explains how to use AI for case-law research so that it saves time without putting the brief, the client and your reputation at risk. The examples come from Italian practice, but the method applies anywhere.

Accursius's Gloss: when a summary becomes "the" source

Back to Bologna, in the first half of the thirteenth century. Justinian's Roman law is again at the center of legal study, and for over a century Bolognese masters, in the tradition traced back to Irnerius, have been commenting on it with glosses: notes written in the margins and between the lines. The glosses pile up, contradict each other, cite each other. Order is needed.

Accursius, a Florentine jurist teaching in Bologna, provides it: around 1230 he completes a systematic collection of glosses on the whole Corpus iuris civilis. His work becomes the Glossa ordinaria, the standard one, and for centuries it accompanies editions of Roman law. Many jurists, and many courts, end up reading the text through the gloss.

You may say: a fine story of efficiency. Indeed. Except that when everyone consults the summary, the summary takes the place of the source. If a gloss misread a passage, the error traveled with it. The next generation of jurists, the commentators, started from the gloss but went back to reasoning on the text and on the concrete problems of their time.

Generative AI is the fastest and most persuasive gloss ever written. It hands you a ready summary, well written and self-assured. The lawyer's job, now as then, is not to mistake it for the source.

The Florence case and its lessons

In March 2025 the Court of Florence (business section) examined a brief citing Supreme Court rulings that did not exist. Counsel later acknowledged that the references came from research done with ChatGPT and never verified. The court did not award damages for vexatious litigation, but criticized the failure to check and described AI "hallucinations": invented results the system confirms even when asked again.

Three practical lessons:

1. A general language model is not a database. It generates plausible text; it doesn't retrieve documents.
2. Responsibility can't be delegated. Not to a junior, not to a tool. Italy's Law 132/2025, Article 13 says it too: AI is allowed for instrumental and support activities, with the professional's intellectual work prevailing.
3. Verification must be organized. Saying "check it" is not enough: the check must be a written step in the workflow.

Three kinds of tools, three levels of reliability

For Italian civil decisions of the lower courts there is also the Ministry of Justice's public case-law database, available since December 2023: a free checkpoint to confirm that a decision exists and what it says.

The six-step method

A typical scenario: a firm with three lawyers and two trainees that wants to use AI for research without taking risks.

1. Frame the question with a general AI, with no client data. Describe the issue in the abstract and ask it to break it into sub-questions, statutes to check, keywords and synonyms.
2. Search the sources, not the chatbot. Use the keywords in the firm's professional database, in public ones, or in the database's AI tool, which shows the document behind each answer.
3. Open every decision. Number, date, section, and above all the passage you need. If you get a headnote, read the reasoning.
4. Summarize only verified texts. Once you've downloaded the real rulings, AI can summarize and compare them. It works on texts you gave it, and you check the result.
5. Log the check. One line in the file: who verified which citations, when, on which source.
6. Write the argument yourself. The choice, order and weight of arguments are the heart of the work.

Useful prompts (and ones to avoid)

Useful:

• "Break this legal issue into sub-questions and list the statutory provisions to check for each."
• "Suggest 15 keywords and synonyms to search case law on this topic in a database."
• "Here are three rulings [pasted, anonymized]: compare the principles and show where they diverge."

Avoid:

• "Give me the Supreme Court rulings that say X." That's the request that produces invented citations.
• Any prompt containing identifying client data in a tool without a data processing agreement and training safeguards.

Confidentiality and ethics

Professional secrecy applies in front of a chat window too. Before adopting a tool, check where data is processed, whether conversations are used for training and whether there is a GDPR Art. 28 contract. The Milan Bar's HOROS charter lists transparency toward clients and the centrality of human decision-making among its principles; Law 132/2025 makes informing the client about AI use mandatory. Add it to your engagement letter.

Mistakes to avoid

• Copying a citation from the output without opening it. That was the Florence mistake.
• Letting the trainee research with a chatbot "because it's faster". Speed is paid for later, in court.
• Treating the headnote as the ruling. A risk before AI; amplified now.
• Saying nothing to the client. In Italy, since 10 October 2025, disclosure is required.

Going further

Research is just one front. For the full picture of rules, tools and method, read our dossier on AI for professional firms; for the firm's other activities there's AI for law firms. If you work closely with accountants, see how they organize in AI software for accountants.

Accursius saved generations of jurists a lot of time; the best of them still opened the text. If you want an AI research workflow that holds up in front of a judge, see how we work with lawyers and book a free intro call.