ChatGPT for tax research: how to check every citation it gives you

What AI is good for in tax research

Tax staff who use AI well describe it the same way: it points the way, and they never take it at face value. Used like that, it's good at:

• suggesting the Code sections, regulations and IRS guidance likely to apply, so you know where to start reading;
• explaining a provision in plain English once you have the text in front of you;
• listing the facts that would change the answer, which makes the questions for the client clearer;
• drafting an outline for a research memo that you then source and write yourself.

How AI gets tax research wrong

• It invents sources. In 2023 a federal court sanctioned lawyers who filed a brief citing cases that ChatGPT had made up, with fake quotations (Reed Smith). A made-up revenue ruling looks exactly like a real one.
• It cites a real source for a point it doesn't make. A 2024 Stanford study of AI research tools built for lawyers found they still gave incorrect information more than 17% of the time. It counted an answer as wrong when it cited a real source that doesn't support the claim (Stanford HAI). This kind of error is harder to catch than a fake source, because the citation checks out.
• It describes the law as it used to be. A model's training stops at a cutoff date, and tax law doesn't. Dollar limits the IRS adjusts for inflation each year are an easy place for an old figure to slip in.
• It answers a different question. It may apply the rules for an individual when your client is a corporation, give a federal answer to a state question, or fill gaps in the facts with assumptions it doesn't mention.
• It sounds just as sure on unsettled points. Where the law is genuinely unclear, a fluent answer can hide that there's a judgment to make.

A prompt that makes the answer checkable

Ask for citations you can look up, a summary of what each one says, and a flag on anything uncertain. Leave out names and anything else that identifies the client: see what can safely go into an AI tool.

I'm researching a US federal tax question. The facts: [the facts, with no names or identifying details].

Question: [your question]

1. List the Code sections, Treasury regulations and IRS guidance (revenue rulings, revenue procedures, notices, publications) most likely to apply, with full citations.
2. For each one, say in one sentence what it says that matters here.
3. List the facts that would change the answer.
4. Mark anything you're unsure of, and anything that may have changed since your training data, as "verify".

Don't cite anything you aren't sure exists. I will check every source myself.

Check every citation in three steps

1. Does it exist?

Look it up in the primary text, not in another AI answer. The Internal Revenue Code and Treasury regulations are free to read at Cornell's Legal Information Institute. Revenue rulings, revenue procedures and notices are published in the Internal Revenue Bulletin, and IRS publications are on irs.gov. Your firm's research service works too. If you can't find it, treat it as invented.

2. Does it say what the AI says it says?

Read the passage, not just the heading. Check the definitions it relies on, the exceptions that follow, and who it applies to. This is where a real citation for the wrong point gets caught.

3. Is it still current, and for the right year?

Check whether the provision has been amended, and whether later guidance modified or superseded it. For any dollar amount or threshold, confirm the figure for the tax year you're working on.

Write it up so a reviewer can follow it

The research isn't finished until someone else can check it. In the memo or the file note:

• cite the primary source with a pinpoint (the section and paragraph), never the AI;
• state the facts you relied on, including any you assumed and still need to confirm with the client;
• note the date you checked each source, since guidance can change after you've read it;
• say where the answer involves judgment, and what would change it.

If AI helped draft the memo, read it against the sources line by line. A draft can overstate a conclusion even when every citation is real.

Can ChatGPT do my taxes, or my clients' taxes?

It can help you understand a form or a rule. It can't take responsibility for a return, and for client work the responsibility is yours. If you practice before the IRS, Circular 230 requires you to exercise due diligence in determining the correctness of what you tell clients and the IRS (31 CFR §10.22). "The AI said so" isn't due diligence.

Client tax information also brings its own rules on what can go into an AI tool at all. The client data guide covers IRC §7216.

Which AI is best for tax research?

People often ask whether Claude, ChatGPT, Copilot or a research tool built for tax professionals is best. The tools change too quickly for a ranking to last, and purpose-built tools still make mistakes, as the Stanford study found. Whichever you use, the three checks above stay the same.