Courts are starting to ask about AI use, clients are asking tougher questions, and The Bluebook hasn’t fully caught up with tools like ChatGPT. If you used an AI tool while drafting, how do you show that properly?

This guide walks you through what to cite, what to disclose, and how to format it so a judge, partner, or editor nods and moves on. Nothing fancy—just clear steps you can use today.

Here’s what we’ll cover:

• When to cite AI output and when to simply disclose that you used AI
• How to adapt Bluebook Rule 18 for ChatGPT and similar tools
• Citation formats and examples you can copy as-is
• What judges are asking for in disclosures and certifications
• Practical templates, record-keeping, and risk tips
• How LegalSoul saves time on citations, archives, and disclosures

By the end, you’ll know how to cite ChatGPT in 2025 without overthinking it—and how to keep your filings safe from avoidable headaches.

Why AI citation and disclosure matter in 2025

In 2025, judges mostly want one thing: you verified what you filed. After Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023), courts look closely at anything that smells like unreviewed AI text or made-up citations.

Some judges now require certifications about AI use. Judge Brantley Starr (N.D. Tex.), for example, asks lawyers to certify that either they didn’t use generative AI or, if they did, a human checked every quote and citation. If your court expects an AI disclosure statement for court filings 2025, plan that language before you draft.

Also, The Bluebook doesn’t treat AI text as authority. If you’re making a legal point, cite the cases, statutes, or rules. Save AI citations for moments when you’re analyzing or quoting the AI itself.

Last thing: add timestamps and a time zone. Models change, outputs vary, and you want a stable record. Archive the transcript early so if the court asks “What exactly did it say?”, you can produce it immediately.

When to cite AI vs. when to disclose AI use

Quick rule of thumb: cite AI when the AI text is the thing you’re talking about. Don’t cite AI to support a legal argument—cite the primary authority instead.

Disclosure is separate. If a local rule, standing order, or client policy requires it, say whether and how you used AI. Keep it brief and focused on verification.

Examples:

• Quoting a ChatGPT paragraph in a law review piece? Cite it and add an archival link.
• Asked AI for a Rule 12(b)(6) outline and replaced everything with verified cases? Probably no AI citation. Disclosure only if the court requires it.
• Used AI for style suggestions? No citation needed. Disclose only if ordered.

Mata v. Avianca is the warning sign: the lawyers trusted AI citations without checking and got sanctioned. Judge Starr’s order says the quiet part out loud—courts want a human in charge.

Keep a quick internal note: “Used AI for brainstorming; replaced all cites with primary sources; no AI text quoted.” That tiny log helps if anyone asks later.

The Bluebook rules you’ll rely on (and how to adapt them)

No single Bluebook rule covers generative AI yet, so use Rule 18 (internet/online sources) and treat the provider as the institutional author. Give a descriptive title, include the date and time, and add a URL or note “on file with author.”

If the content can change, use an archival link like Perma.cc. That’s squarely within Bluebook Rule 18 online sources for generative AI, and it preserves what you relied on.

• Institutional author: “OpenAI, ChatGPT …”
• Title: “response to prompt ‘[your prompt]’”
• Timing: include time and time zone
• Stability: share link if possible; otherwise “(transcript on file with author)”
• Archiving: add a Perma.cc link or similar

Journals usually accept this approach, with a note that outputs can vary. If you edited the AI text, be clear about what’s quoted verbatim versus your edits. That clarity protects you and keeps editors happy.

Core elements of a Bluebook-style AI citation

Here’s the basic structure for a Bluebook citation for ChatGPT 2025:

• Author: OpenAI (or the relevant provider)
• Product/model: ChatGPT (include model/version, e.g., GPT-4o)
• Descriptive title: “response to prompt ‘[exact or paraphrased prompt]’”
• Date and time, with time zone
• URL to a transcript, or “(transcript on file with author)”
• Archival link (Perma.cc or similar), if available
• Parenthetical: “AI-generated text; outputs may vary”

Use cases:

• Quoting a short explanation of Rule 11? Include the exact prompt, timestamp, and link.
• No public URL? Export to PDF, store it, and note “on file with author.”

When citing AI prompts and transcripts “on file with author,” consider adding an internal ID or hash. If someone questions authenticity, you can prove the file hasn’t changed. Also, standardize a time zone (e.g., ET) so your timestamps match docket activity.

Citation formats and examples you can use now

Here are sample AI citation formats and examples for lawyers that line up with Rule 18:

1) With public share link and archive:
OpenAI, ChatGPT (GPT-4o) response to prompt “Summarize Fed. R. Civ. P. 11,” June 10, 2025, 2:13 PM ET, https://chat.openai.com/share/abc123, archived at https://perma.cc/XXXX-XXXX (AI-generated text; outputs may vary).

2) No public link (“on file with author”):
OpenAI, ChatGPT (GPT-4o) response to prompt “List elements of common-law fraud under New York law,” May 2, 2025, 9:05 AM PT (transcript on file with author) (AI-generated text; outputs may vary).

3) Referencing AI for workflow only (authority is primary law):
See OpenAI, ChatGPT (GPT-4o) response to prompt “Draft a memo on Rule 12(b)(6) standards,” Apr. 7, 2025, 4:42 PM ET (on file with author). Cf. Ashcroft v. Iqbal, 556 U.S. 662, 678–79 (2009).

A Perma.cc archival link for AI output citation is useful when a judge or editor wants to see exactly what you saw. After the Avianca sanctions, courts expect verification and preservation. Add a short parenthetical explaining why you quoted the AI (e.g., discussing methodology) to show you’re not treating it as authority.

Context-specific guidance: filings, memos, and publications

Court filings. Follow any local rule or standing order that asks for an AI disclosure statement for court filings 2025. If you discuss the AI output, attach the transcript as an exhibit, say “on file with author,” and include an archive link. Redact sensitive information, and move to seal if appropriate.

Internal memos and client work product. Keep documentation light but complete. A short note—prompt, model, date/time, and verification steps—keeps your file clean without cluttering your analysis.

Law review and publications. Many journals accept AI citations under Rule 18 with Perma.cc links. Label AI-generated text clearly and match the journal’s house style.

Practice-area notes. In litigation, be cautious with exhibits and privilege. In deals, let AI help with a first pass, then swap in verified authorities before sharing drafts externally.

One more thing: when attaching AI transcripts, scrub names and matter details from prompts. Maintain a private cross-reference in your file so your archive stays useful without exposing client information.

Preserving the record: prompts, transcripts, and confidentiality

Courts care about your process as much as your footnotes. Save the exact prompt, the output, the model/version, the date and time (with time zone), and where you stored it. If there’s no public link, mark it “(transcript on file with author)” and export to a locked PDF.

For extra assurance, generate a SHA-256 hash and drop it into your internal log. If someone challenges authenticity, you can match the file to the hash, just like other digital evidence under Fed. R. Evid. 901.

Keep a clean, discoverable copy separate from privileged notes. If the AI output might become evidence, track who exported it, when, and from which account. Treat it like any other exhibit.

If a court wants exhibits, file a redacted public version and an unredacted sealed version with your motion. A surprisingly useful habit is to build standard redaction language for prompts so client names never hit a public-facing tool.

Disclosure and certification language (templates)

When a judge requires disclosure, be clear and brief. Many standing orders (like Judge Brantley Starr’s in N.D. Tex.) focus on human verification of quotes and citations.

Short-form filing disclosure (when ordered):
“Counsel used generative artificial-intelligence tools to assist with drafting. A licensed attorney verified the accuracy of all legal and factual citations, and the brief relies solely on primary legal authorities for propositions of law.”

Certification (when required):
“I certify that no portion of this filing was drafted by generative artificial intelligence tools, or, to the extent such tools were used, I have reviewed and verified all generated text and citations for accuracy and completeness, and this filing cites only authentic, accurate sources.”

Client engagement addendum:
“The firm may use vetted AI tools to enhance efficiency. All legal analysis and citations will be verified by licensed attorneys. We do not input client-confidential information into public AI systems without consent.”

One line that helps: “All authorities were Shepardized/KeyCited post-generation.” Courts want to know a human checked the cites. Say it plainly.

Common mistakes and how to avoid them

• Using AI as authority. Don’t. The Bluebook treats AI like an online source, not law. Cite primary sources. Avianca is the cautionary tale.
• Skipping timestamps. Outputs change. Include date, time, time zone, and model/version.
• Flimsy links. Share links can break. Add an archive (Perma.cc) or use “on file with author.”
• Over- or under-disclosing. Follow the specific rule or order. Don’t volunteer strategy. Don’t ignore a requirement.
• Quoting edited AI text without saying so. If you tweaked wording, bracket edits or clarify what’s verbatim.

Two habits pay off: run a reverse-citation check—every claim traces to a primary authority you pulled yourself. And if you include AI text at all, keep it to background or methodology, not the core legal point. Ethics and Rule 11 expectations haven’t changed: your signature means you vouch for all of it.

Ethics and risk management

Your duties are familiar: competence (Model Rule 1.1), confidentiality (1.6), candor (3.3), and accuracy (Rule 11). Post-Avianca, judges are faster to ask, “Did you verify?” So build verification into the workflow and keep proof.

Good guardrails:

• Use enterprise or on-prem tools when possible; avoid feeding sensitive facts into public systems.
• Strip names and identifiers from prompts; keep a private crosswalk.
• Have a lawyer check every citation an AI suggests.
• Archive transcripts in a controlled repository “on file with author.”

Publishing? Note how AI was used, matching common law review Bluebook citation of AI-generated text practices (Rule 18 with archives). Watch for inadvertent waiver—don’t over-share internal drafting steps unless a court asks for them. Keep public disclosures narrow and precise.

Workflow: from AI output to compliant citation

Here’s a simple path you can repeat:

1. Write safer prompts. Avoid client names; ask general legal questions.
2. Skim, then research. Let AI help with structure, but find your own sources. Never paste in citations sight unseen.
3. Verify everything. Shepardize/KeyCite and replace AI-suggested sources with official reporters.
4. Decide what to include. If you quote AI, build a Bluebook citation for ChatGPT 2025 with model and timestamp.
5. Save the record. Export the transcript, note a file hash, store it, and create a Perma.cc archival link if possible.
6. Handle disclosures. If required, add a short certification and keep your verification log “on file with author.”

Make yourself a one-page style sheet with your go-to parentheticals, time zone, and archiving steps. Hashing the transcript sounds nerdy, but it’s gold if anyone questions the document later.

How LegalSoul helps with compliant AI citations

LegalSoul is built for lawyers who want speed without cutting corners. It captures the institutional author, model/version (like GPT-4o), your prompt, and the exact date and time—then generates a Bluebook-ready citation in one click.

What you get:

• Footnotes and short-form cross-references ready to paste.
• An archiving workflow that produces a snapshot suitable for Perma.cc and stores a copy with a file hash.
• Authority checks that compare AI-suggested cites to official reporters and flag anything suspicious before it hits your draft.
• A disclosure builder with judge- and jurisdiction-specific language.
• Privacy controls, including prompt redaction templates and workspace policies to protect client data.

A midsize litigation team can turn an AI brainstorm into a clean, court-ready record—citations, archive, and disclosure—without burning a weekend, and with an audit trail that stands up to questions.

FAQ

Do I need to include the model/version?
Yes, it helps a lot. Models change, so noting “GPT-4o” (or similar) makes your reference reproducible.

What if there’s no public URL?
Use “(transcript on file with author)” and save a PDF export. Add an archival link later if your platform allows.

Can I cite screenshots?
Sure. Pair them with a text export for accessibility. Label them clearly and include date/time.

Should I say outputs vary?
Yes. A short parenthetical—“AI-generated text; outputs may vary”—avoids confusion.

How do I handle changes over time?
Archive promptly (e.g., Perma.cc) and keep an immutable internal copy.

Are journals okay with this?
Most follow Bluebook Rule 18 for online sources with archival links. Check the journal’s policy.

Can AI ever be authority?
No. Use it to draft or brainstorm, but cite the primary law for your propositions.

Key Points

• Cite AI only when you’re talking about the AI text itself. For legal points, cite cases, statutes, and rules. Disclose AI use only when a rule, order, or client policy requires it—and attach the transcript if you reference it.
• Adapt Bluebook Rule 18: institutional author, descriptive title with the prompt, model/version, exact date/time with time zone, stable URL or “(transcript on file with author),” an archival link (Perma.cc), and a note that outputs vary.
• Preserve your record: export transcripts, keep an internal log, consider hashing files, and redact sensitive details. Post-Avianca, courts expect human verification and stable references.
• LegalSoul helps by generating Bluebook-ready citations, storing verifiable transcripts, creating archival snapshots, checking authorities, and drafting judge-specific disclosures.

Conclusion

In short: treat AI like an online source, not authority. Cite it only when you’re quoting or analyzing the output. For legal arguments, cite primary sources. Follow Rule 18 basics—author, title with the prompt, model/version, date/time, link or “on file with author,” and an archive.

Keep transcripts, redact sensitive bits, and disclose only when required, with a clear note that a lawyer verified everything. Want to make this part easy? LegalSoul can produce the citation, store the transcript, confirm the authorities, and draft the disclosure in a few clicks. Grab a 15-minute demo or start a trial and put this on autopilot for your next filing.