Subject workflows · Updated 22 September 2026

How to Study Business Law With AI (Safely)

Business law is the one course in a business degree where the right answer has a source attached to it. That makes it the course where a chatbot is most useful for explaining a rule, and most dangerous for telling you where the rule came from.

Use a chatbot for this course, but let it explain rules and never let it supply authority. The measured pattern in legal work is that these models are most accurate on the famous, frequently discussed decisions and least accurate on lower courts, less prominent cases and very recent law — which is the exact opposite of the gradient you need, because nobody opens a chat window to ask about the landmark case the textbook already explained. The answer arrives in the same confident register either way.

The reliability gradient runs the wrong way

The foundational measurement is "Large Legal Fictions", published in the Journal of Legal Analysis in 2024, in which Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel E. Ho profiled how models answered direct questions about federal court cases. Treat its headline as history rather than a current rate — it tested the models of its moment, reporting hallucination "between 58% (ChatGPT 4) and 88% (Llama 2) of the time". What has aged better is the shape of the errors, because that shape follows which parts of the law get written about most, and that has not changed.

Three findings matter more than any percentage. "Hallucinations are lowest in the highest levels of the judiciary, and vice-versa" — best on the Supreme Court, worst on district courts. "Case prominence is negatively correlated with hallucination", with accuracy on Supreme Court cases improving sharply for the most famous decisions. And hallucinations were "most common among the Supreme Court's oldest and newest cases".

Two further findings explain why the model cannot police itself. These models "seem to suffer from contra-factual bias on these legal information tasks" — put a false premise in the question and it runs with it, so "why did the court hold the contract void for lack of consideration" gets an explanation whether or not the court held anything of the kind. And under their default configurations they are "imperfect predictors of their own tendency to confidently hallucinate legal falsehoods", so the confident tone carries no information about reliability.

Newer models did not quietly fix this

The obvious objection is that all of this describes older models. It has now been tested directly. In "Who Checks the Citations?", posted in June 2026, Patty Liu, Dominik Stammbach and Peter Henderson tracked fabricated citations in real court filings and note that this happened "despite predictions that newer models would hallucinate less or that court sanctions would deter negligent filers". They identify over 1,000 filings containing hallucinated citations, "with this number growing year-over-year".

The trend is not a straight line. Early GPT-4o models from mid-2024 hallucinated citations at 1.23%, down from "around 25% in earlier GPT-3.5 models" — but GPT-5.1 came in at 6.57%, higher than that 2024 low, a difference the authors report as statistically significant. The field improved enormously and then gave some of it back, so "use the newest model" is not a safety strategy.

You can check the scale yourself, because someone is counting. Damien Charlotin maintains a public AI Hallucination Cases database tracking "legal decisions where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by the court or tribunal". On 22 September 2026 it listed 2,046 cases, updated the day before. Those are qualified professionals with malpractice insurance and a duty to the court; your safeguards are thinner than theirs. None of this research is about undergraduate coursework — it measures lawyers and court filings, so take the error patterns from it rather than the rates.

The invented case is the easy problem

Here is the part almost every warning skips: "it might make up a case" is the failure students guard against, and it is the most detectable one. The Liu paper sorts legal citation errors into five kinds, and only the first is the one you are watching for.

• Non-existent citation — "the citation does not correspond to any real case".
• Case name mismatch — the case name and the reporter citation "refer to two different real cases".
• Incorrect pincite — the case is right, but "the cited page number does not support the quoted language or proposition".
• Verbatim misquote — "the exact quoted language does not appear in the cited case".
• Content misrepresentation — "a cited case exists but does not support the proposition for which it is cited".

Those last three survive the check you were going to run: you search the case name, it exists, you move on. And when the researchers tested automated detection, those same three were the categories "all models struggle most with" — the strongest configuration caught incorrect pincites at 18.2% recall. The error you would never think to look for is also the one a checker misses.

There is a structural reason, and it hands you an advantage. The paper notes that free repositories "do not include official reporter pagination, which originates in commercial publisher formatting, making it infeasible to verify the pincites using publicly available sources". The reason an automated checker cannot reliably verify a citation is that the authoritative version sits behind a subscription — and your library almost certainly buys one. Ask a librarian which legal database your institution subscribes to and you are holding the exact resource the research says those checkers lack.

The law moves, and the training data remembers every version

Chunks of this syllabus change while you are studying them, and the model has read every stage of the change with equal confidence. Non-compete agreements are the cleanest example: the position moved four times between 2024 and 2025, and the FTC's own page on the Noncompete Rule now states where it landed. "The Noncompete Rule is not in effect and it is not enforceable. On August 20, 2024, a district court issued an order stopping the FTC from enforcing the rule. The FTC appealed that decision on October 18, 2024. On September 5, 2025, the FTC took steps to dismiss its appeal in the Fifth Circuit."

Each of those four states of the world generated a mountain of confident writing while it was current. A model asked "are non-competes legal" draws on all of it at once, and whichever version surfaces arrives without a date. The same trap sits in employment law, consumer protection, data privacy and anything touching an agency rule.

Where the free ground truth is

Start with your casebook and syllabus, because your course's stated rule is the one you are marked against. For reading actual opinions, the Library of Congress maintains a guide to free case law online: CourtListener is "a legal database operated by the Free Law Project, a nonprofit organization", and the Caselaw Access Project, "maintained by the Harvard Law School Library Innovation Lab", includes "all official, book-published state and federal United States case law through 2020" — search it through CourtListener, since CAP's own search was switched off in September 2024. The guide is candid that CourtListener's citation-tracking features are "not considered as authoritative as subscription options available", which is an argument for the library card rather than against the free ones. For the doctrine itself, OpenStax publishes Business Law I Essentials free under a Creative Commons licence, written for courses on "Business Law or the Legal Environment of Business".

Split the job

The workflow, step by step

Four prompts for ChatGPT, Claude, Gemini or whatever you already have open. None of them asks the model for a citation.

1. Find out which of your sentences need authority. Most students cannot tell which of their own claims are safe generalities and which are assertions about a specific case or state.

Below is a passage from my business law coursework.

Go through it sentence by sentence. Label each sentence:
- GENERAL RULE (a doctrine stated in general terms)
- CASE CLAIM (asserts what a specific case held or said)
- JURISDICTION CLAIM (asserts what the law is in a named
  state, or under a specific statute or agency rule)
- FACT APPLICATION (my own reasoning about my fact pattern)

For every CASE CLAIM and JURISDICTION CLAIM, write one line
saying what a reader would have to open to confirm it.

Rules:
- Do NOT tell me whether any sentence is correct.
- Do NOT supply case names, citations or corrections.
- Do NOT rewrite my passage.

Passage: [paste]

The two middle labels are your checking list, and it is usually shorter than people fear — often three or four sentences carrying a whole assignment. Take those to the database.

2. Ask for the search, not the source. This is the habit that makes the rest of the course safe.

I need to find authority on this issue: [describe the issue]

Do NOT give me any case names, citations, statute numbers or
quotations. I am not asking for them and I will not use them.

Instead give me:
1. The legal terms of art I should be searching for.
2. The doctrinal area this sits in, and the neighbouring
   areas it is often confused with.
3. The elements a court would be working through, in order.
4. What kind of source would settle it - a state statute,
   a uniform code as adopted by my state, a federal agency
   rule, or case law - and roughly which court level.

If the answer genuinely varies between states, say so and
stop rather than picking one.

You get the vocabulary, which is what actually makes a database search work and what a beginner lacks, without putting an unverifiable citation into your notes where it can quietly migrate into your essay.

3. Make it read only what you paste. Run this one if you run nothing else: it converts the model from a source into a reader, which is the job it is good at.

Below is (a) the text of a court opinion I pulled from my
library database myself, and (b) a proposition someone has
claimed it supports.

Work ONLY from the pasted text. You must not use anything you
remember about this case, this area of law, or any other case.
If the pasted text does not settle a question, say so.

Tell me:
1. Does the exact quoted language appear in the pasted text?
   Quote what is actually there, or write NOT PRESENT.
2. Does the pasted text support the proposition, partly
   support it, or not support it? Point to the sentences.
3. Is the proposition about the holding, or about something
   the court said in passing?

Do not summarise the case. Do not tell me whether the
proposition is good law.

Opinion text: [paste]   Proposition: [paste]

This is how you catch the three error types a name search cannot. It also works on your own notes: run it against a case you wrote down from memory three weeks ago and you will learn quickly whether you remembered a holding or a vibe.

4. Date and locate every rule before you rely on it. The non-compete story above is not a curiosity; it is the default condition of half this syllabus.

I am about to rely on this rule in a business law assignment:
[state the rule in your own words]

Before I do, tell me:
1. What kind of law is this - state common law, a uniform
   code as adopted by individual states, a state statute,
   a federal statute, or a federal agency rule?
2. Does it vary from state to state? If yes, name the
   dimensions on which states differ, not the states.
3. What is the most recent change to this area that you are
   aware of, and when did it happen?
4. What would you NOT know about the current position, and
   why?

Write NOT SURE wherever that is the honest answer. Do not
estimate a date and do not guess at the current position.
Question 4 is the important one - answer it properly.

Question four does the real work. An honest answer tells you which parts of your assignment need a current source rather than a confident paragraph, and the cut-off is usually further back than students assume.

Where the line is

Everything above is studying: being taught a doctrine, having your own draft sorted into claim types, being handed search vocabulary, having a text you supplied read closely. The issue spotting, the application and the conclusion are the assessed work and they stay yours. If your syllabus is vague, our class AI policy checklist and homework help without cheating cover how to read it, and citing AI in APA, MLA and Chicago covers disclosure.

One warning is worth taking from the people already living it. Charlotin's reading of his own database is that "people are rarely sanctioned only because they erred in using an AI tool but, when caught out, refused to own up to it, double-downed, made up stories, or blamed the intern." The penalty attaches to the cover-up, and academic integrity offices work the same way. He also names a recurring category of "lawyers who are surprised that an existing tool suddenly has an AI component that hallucinates" — worth remembering the next time your research database or citation manager grows a summarise button.

Related reading

• The rest of the business core. Business case study analysis, managerial accounting and financial accounting.
• If you are heading to law school. Reading and briefing cases, the LSAT and the NextGen bar exam.
• Checking sources properly. The citation audit checklist and the source evaluation checklist — both built for exactly this problem.
• Checking anything a model tells you. The hallucination checklist and verifying AI answers before you study them.

FAQ

Can I ask a chatbot to find cases for my business law assignment?

Ask it for search terms instead. The research on legal citation errors sorts them into five kinds, and only one of them is the fabricated case most people watch for. The other four involve real cases attached to the wrong name, the wrong page, a quotation that is not in the opinion, or a proposition the case does not actually support. Search vocabulary carries no authority and cannot be wrong in that way, which is why the second prompt on this page asks for the search rather than the source.

The case it gave me is real, I checked. Does that make it safe?

Not on its own. Confirming that a case exists rules out one of the five error categories and leaves three of the harder ones untouched. When researchers benchmarked automated hallucination detection in June 2026, incorrect page cites were caught at 18.2% recall by the strongest system they tested, and verbatim misquotes and content misrepresentation were named alongside it as the categories all models struggled most with. The only reliable check is opening the opinion and reading the passage.

Does using the newest model fix this?

The measurements say no, and the trend is not even monotonic. The same June 2026 paper found early GPT-4o models from mid-2024 hallucinating citations at 1.23%, down from around 25% in earlier GPT-3.5 models, but GPT-5.1 at 6.57% — higher than the 2024 low, and reported as a statistically significant difference. The authors point out that the growth in fabricated citations reaching courts happened despite the widespread expectation that newer models would solve it.

My professor asks about a specific state. Can AI tell me my state rule?

Treat anything state-specific as unverified until you check it. Accuracy in the profiling study fell as you moved down the court hierarchy and away from prominent cases, and state law is where most business law problems actually live. There is also a currency problem: rules change, and the training data contains every previous version of a rule stated just as confidently as the current one. The fourth prompt on this page is built to surface that before you rely on it.

What happens if I hand in a fabricated citation?

That depends far more on what you do next than on the citation. Damien Charlotin, who maintains the public database of court decisions involving AI hallucinations, observes that people are rarely sanctioned merely for erring with an AI tool, but rather when they refused to own up to it, doubled down, made up stories or blamed someone else. Academic integrity processes behave similarly. If you find one in work you have already submitted, email your instructor and say so plainly before anyone raises it with you.

Bottom line

Business law splits cleanly into doctrine and authority, and a chatbot is good at one of them. Doctrine is the most repeated material in its training data, and it will explain consideration or respondeat superior patiently for as long as you want. Authority is the opposite: thinner the further you get from famous decisions, wrong in four ways that survive the check you were planning to run, and undated where the date is often the answer. So let it teach you the concept and hand you the search terms, take every case, citation and state-specific rule from your casebook or a real database, and keep the application to yourself — that was always the graded part.