Comparison
Can ChatGPT Make a Table of Authorities? What to Check Before You File
On this page4 sections
The short answer
yes, ChatGPT and similar chatbots can produce something that looks like a Table of Authorities, and they are genuinely useful for a first pass: pulling citations out of text, grouping them into cases, statutes and rules, and alphabetizing. What a general chatbot does not guarantee is the part a court cares about: page references that match the pages of the brief you actually file, a complete list with nothing added or dropped, and citations reproduced exactly as they appear in the brief. It also means sending the brief to the AI provider. If you use one, treat its output as a draft and check every entry against the final PDF. This page covers what a chatbot does well, where it can go wrong, a verification checklist that works for any AI-generated TOA, and how a deterministic tool such as BriefAuthority (in development) approaches the same job differently.
What a chatbot does well
It is fair to start with the strengths, because they are real. Given the text of a brief, a capable chatbot can usually:
- Spot most citations in running text, including case names, U.S.C. sections, C.F.R. provisions and court rules.
- Sort them into categories (cases, constitutional provisions, statutes, regulations, rules, other authorities) and alphabetize cases the way a TOA expects.
- Suggest formatting, such as italicized case names and a consistent citation style for the table.
- Explain conventions like passim, Id. and supra, or what your court's rule requires, which is helpful when you are learning.
For a short brief where you already know the authorities, that can save real time as a starting list.
What a chatbot does not guarantee
| Requirement | What can go wrong with a general chatbot | How to check |
|---|---|---|
| Page references match the filed brief | Pasted text has no page numbers at all; an uploaded PDF may be read as a text stream, so page breaks, cover pages and roman-numeral front matter can be miscounted or guessed | Spot-check every page reference against the final PDF's printed page numbers |
| Completeness | Long briefs can exceed what the model reads closely; short-form references by case name alone (e.g. “Rehaif”) and Id. cites are easy to miss | Search the PDF for “v.”, “U.S.C.”, “C.F.R.”, “Id.” and each case's short name |
| Nothing added | A model can insert an authority that is not in the brief, or one that does not exist | Confirm every table entry appears in the brief itself |
| Citations reproduced exactly | A model may “correct” a reporter volume, year, pin cite or case name, so the table no longer matches the text | Compare each entry character by character with the brief |
| Repeatable result | Running the same prompt twice can give different lists | Regenerate and diff, or verify fully once |
| Filing-ready format | Output is usually plain text or Markdown; tab-aligned dot leaders and your template's styles need rebuilding | Rebuild in Word or Pages and check alignment |
| Confidentiality | The brief's full text goes to the AI provider under that plan's terms | Check your office policy, client terms and ABA Formal Opinion 512 |
None of this means a chatbot always gets a TOA wrong. It means the tool is not designed to guarantee these properties, so the responsibility for them stays with whoever signs the brief. That is the lesson courts have drawn in the AI citation cases.
In Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), 678 F. Supp. 3d 443 (S.D.N.Y. June 22, 2023), the court sanctioned counsel under Rule 11 after they submitted non-existent judicial opinions, with fake quotes and citations, created by ChatGPT, and then continued to stand by them after the problem was raised.
Mata involved AI-drafted argument, not a TOA. But the same failure applies to a table: if a generated list contains an authority that is not in the brief, or a citation altered from the one in the text, the table misrepresents the brief to the court. A TOA is often the first page a clerk uses to pull your authorities.
Verification checklist for any AI-generated TOA
Before you file, confirm:
- The source is final. The table was generated from the version you will file, after the last edit and after the cover and front matter were added.
- Every entry exists in the brief. Search the PDF for each case and statute listed. Delete anything you cannot find.
- Every authority in the brief is in the table. Search for “v.”, “U.S.C.”, “C.F.R.”, “Fed. R.”, “Const.”, “Id.” and each case's short name. Short-form and Id. references count toward page lists.
- Citations match the text exactly. Reporter, volume, page, court and year in the table match the brief. If the brief has an error, fix the brief, then the table.
- Page references use printed page numbers, not the PDF viewer's page index. Check a sample on every page of the argument, and all entries for the cases cited most often.
- The TOA's own pages and the cover are excluded from the page references, unless your court says otherwise.
- Categories and order follow your court's rule: cases alphabetical, then constitutional provisions, statutes, regulations, rules and other authorities as your rule specifies.
- Passim is used only as your court allows. Some courts discourage it; some practitioners use it above a set number of pages.
- Formatting is final: italics, dot leaders aligned, headings consistent with the Table of Contents.
- Confidentiality was cleared before the brief was pasted or uploaded to any AI tool.
- The authorities themselves were checked in a citator. A TOA tool, AI or not, does not tell you whether a case is still good law.
How BriefAuthority differs from a chatbot
| Item | General AI chatbot | BriefAuthority (in development) |
|---|---|---|
| How citations are found | Generated by a language model from the text it reads | Extracted deterministically by a citation parser from the text of the PDF; it lists only what is in the brief |
| Page references | Not tied to printed pages unless the model infers them | Read page by page from the final PDF, including roman-numeral front matter offsets |
| Review | You check the output yourself, entry by entry | A review screen shows every authority and its pages; you merge short forms, link Id. and supra, re-categorize or exclude |
| Output | Text or Markdown you reformat | A formatted .docx with category headings, dot leaders, page references and optional passim |
| Where the brief goes | To the AI provider's servers | Nowhere: built to process the PDF in your browser, on your computer |
| Citation validity | Not a citator | Not a citator; formatting and indexing only |
| Availability | Available now, various plans | Pre-launch; early access, planned $39 one-time or $9 per brief |
Deterministic extraction has limits too. In an internal pre-launch test on two U.S. Supreme Court merits briefs (Greer v. United States, No. 19-8709; Bouarfa v. Mayorkas, No. 23-583), the BriefAuthority citation engine found 129 of the 131 cases listed in the briefs' own professionally prepared Tables of Authorities. Page lists matched exactly for 77–84% of cases before review; most misses were short references by case name alone, which is what the review screen is built to catch. Statutes and rules have not yet been benchmarked. The checklist above applies to BriefAuthority output as well.
Frequently asked
Can ChatGPT make a table of authorities for my brief?
It can produce a draft list of authorities grouped by type, and that can be a useful start. It does not guarantee that page references match your final PDF, that every authority is included, or that nothing is added or altered. Check every entry against the filed version before you rely on it.
Why are the page numbers ChatGPT gives me wrong?
If you pasted text, there were no page numbers to read, so any numbers are inferred. If you uploaded a PDF, the model may count PDF pages rather than printed page numbers, which differ once a cover, Table of Contents and roman-numeral front matter are included. A TOA must cite the printed page numbers of the brief.
Can ChatGPT invent cases in a table of authorities?
Language models can generate citations that do not exist or alter real ones, which is what led to sanctions in Mata v. Avianca (S.D.N.Y. 2023). In a TOA the risk is an entry that is not in your brief or a citation that no longer matches the text. Confirm every entry appears in the brief.
Is it safe to paste a confidential brief into ChatGPT?
That depends on your plan's data terms, your office policy and your client's terms. ABA Formal Opinion 512 asks lawyers to assess the disclosure risk and, for tools that learn from inputs, to get informed client consent first. This is not ethics advice; if you cannot upload, see our page on making a TOA without uploading.
What is the fastest way to check an AI-generated table of authorities?
Search the final PDF for each listed authority to confirm it exists and the citation matches, then search for “v.”, “U.S.C.”, “Id.” and short case names to catch anything missing. Finally, spot-check page references against the printed page numbers, starting with the most-cited cases.
Does BriefAuthority use AI?
BriefAuthority is designed as deterministic software: a JavaScript citation parser reads the text of your PDF and lists the citations it finds, so it does not generate or invent authorities. You then review everything before export. It is pre-launch; this describes the specification, confirmed at launch.
Sources
- Legal Information Institute — Federal Rules of Appellate Procedure, Rule 28 (Briefs)law.cornell.edu
- CourtListener — Mata v. Avianca, Inc., No. 1:22-cv-01461 (S.D.N.Y.) docketcourtlistener.com
- ABA Standing Committee on Ethics and Professional Responsibility — Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024)americanbar.org
- American Bar Association — Model Rule 1.6: Confidentiality of Informationamericanbar.org
- Microsoft Support — Create a table of authoritiessupport.microsoft.com
- Supreme Court of the United States — Rules and Guidancesupremecourt.gov
BriefAuthority · launch price $39
In development · early access