New Bluebook Rule On Citing to AI Generates Criticism from Legal Scholars and Practitioners

Has there ever been a time since the advent of legal reporting systems when citations have been under greater attack? Driven by their unwitting reliance on AI to generate legal briefs, lawyers seem to have forgotten everything they ever learned in law school about how to research and cite the law.

Standing as a bulwark against this attack, one would think, is The Bluebook, the uniform system of citation that is among the first things taught to a first-year law student, and to which virtually all lawyers are expected to abide, except where excused by local rules of court.

Yet now that very bulwark is itself under attack, thanks to the release last May of its 22nd edition, which introduced Rule 18.3, The Bluebook’s first standardized format for citing to generative artificial intelligence content.

While the addition of AI citation guidance would seem to reflect The Bluebook’s expected role of evolving to address new types and formats of sources, the new rule has sparked criticism from legal scholars and practitioners who argue it is fundamentally flawed in both conception and execution.

As Cullen O’Keefe, director of research at the Institute for Law & AI, says in his analysis, “Citing AI in the New Bluebook,” “I’m afraid The Bluebook editors have fallen a fair bit short in the substance and structure of Rule 18.3.”

Susan Tanner, associate professor of law at the University of Louisville, put it more bluntly: “This is bonkers.”

What the New Rule Requires

So what’s the reason for all the brouhaha? Let’s start with what the rule says.

The new Rule 18.3 is divided into three subsections, each designed to address different types of AI-generated content:

• Rule 18.3(a), “large language models,” covers text output from AI tools.
• Rule 18.3(b), “search results,” addresses AI-powered search engine results.
• Rule 18.3(c), “AI-generated content,” focuses on non-textual AI output such as images.

For each of these, the citation rules differ slightly. Notably, the rule requires the author who is citing AI to save and store a screenshot of the AI output as a PDF.

According to a Bluebook citation guide published by the law library at the University of Cincinnati, citations to a large language model (LLM) under 18.3(a) must include:

• The author of the prompt.
• The name of the model used (include the version number if one exists).
• The exact text of the prompt submission in quotation marks.
• The date the prompt was submitted.
• A parenthetical stating where the PDF is stored.

For AI search results under 18.3(b), the guide says, the citation must include:

• The name of the search engine.
• The exact text of the query in quotation marks.
• The number of results (if available).
• The date the search was conducted.
• A parenthetical stating where the PDF is stored.

For AI-generated content under 18.3(c), the citation should conform to the relevant subrule for that type of content. In addition:

• If the subrule requires an author, then use the name of the person who submitted the prompt or omit it if the prompt author is unavailable.
• Include a parenthetical indicating the content was AI generated and the name of the AI model used.

Some critics say that even just this structural logic of the new rule is confusing.

“What is supposed to be the difference between ‘large-language models’ (Rule 18.3(a)) and ‘AI-generated content’ (Rule 18.3(c))?” asks O’Keefe. “Content generated by LLMs is a type of AI-generated content!”

When Should AI Be Cited?

Perhaps the most significant criticism of Rule 18.3 is its failure to address the fundamental question of when AI citations are appropriate.

As Jessica R. Gunder, assistant professor of law at the University of Idaho College of Law, argues in her comprehensive critique, “Yikes! The Bluebook’s Generative AI Rule is Flawed,” the rule provides “detailed guidance regarding how a writer should cite their use of generative AI technology but does not include any discussion or deliberation regarding when a citation is appropriate.”

This absence of guidance, Gunder argues, makes it unclear whether Rule 18.3’s citation form is to be used where a legal author relies upon AI as a source of authority or evidence, or also where AI is employed as a research or writing tool.

In the latter case, when AI is used as a tool, then the traditional rationales for legal citation — allowing readers to locate sources, communicating authority weight, demonstrating credibility and avoiding plagiarism — do not support citation to the tool, she argues.

“Ultimately, generative AI is a tool, not an authority in and of itself,” Gunder writes. “Much like a professor would not cite to their research assistant, a partner at a law firm would not cite to the legal assistant who put together the first draft of a motion, or a judge would not cite a law clerk who crafted the first draft of an order, one would not expect a legal writer to cite to a generative AI tool.”

Internal Inconsistencies

Critics also say that Rule 18.3 suffers from numerous internal inconsistencies. O’Keefe points out that The Bluebook’s own examples fail to follow the stated requirements. While Rule 18.3(a) specifies that citations should include “the exact text of the prompt submission in quotation marks,” The Bluebook provides an illustrative example that contains no quoted prompt text whatsoever.

Similarly, the rule is inconsistent about whether to include company names alongside model names. O’Keefe points to two examples given by The Bluebook, one of which gives the company name of OpenAI while the other gives no company name.

Gunder, pointing to these same inconsistencies with the new rule, writes, “These inconsistencies make it impossible for legal writers to discern whether they are complying with the parameters set by Rule 18.3(a),” undermining The Bluebook’s fundamental goal of creating uniform citation standards.

Unreasonable Technical Burden

The rule’s technical requirements also present practical challenges, critics say. Gunder cites empirical studies showing that legal professionals lack basic technological competency, with only about one-third of law students able to perform fundamental Microsoft Word tasks like accepting track changes or inserting hyperlinks.

Even more concerning, while over 85% of lawyers can convert files to PDF, only 4% can successfully create PDFs that maintain active hyperlinks and convert headings to bookmarks.

“The Bluebook’s citation requirements reflect an optimistic view of the technological skills possessed by attorneys,” she writes. “However, the reality is significantly bleaker.”

Jayne Woods, professor at the University of Missouri School of Law, writing in “The New Bluebook Rules for Generative AI,” notes that many attorneys “do not know how to take a scrolling screenshot (one that covers more than just the content visible on a single page) or convert files to PDF, the process can quickly become cumbersome, given that conversations with a generative AI model likely span more than one visible page.”

[Note that, thanks to the demise of blogging platform TypePad, Woods’ article appears to be no longer accessible.]

Creating citation formats that are too complex, Gunder warns, “can encourage practitioners and scholars to deviate from the standard norms,” undermining The Bluebook’s primary goal of uniformity.

Incompatibility with Actual AI Usage

Another criticism of the rule is that it fundamentally misunderstands how gen AI is actually used in legal practice.

As both O’Keefe and Gunder point out, effective AI prompting typically involves complex, iterative conversations that may span multiple rounds of prompting and refinement and include uploaded documents.

Gunder provides a realistic example of an attorney responding to a motion for extension of time:

“A skilled prompter would upload the motion that was filed. They may upload sample responses they have employed in other cases so that the generative AI tool could emulate the style and tone of those documents. They would review the initial output and see that it has errors or omissions and ask the tool to craft a revised response that addresses those problems.”

The final prompt in such a sequence might be something mundane like “please rewrite that to capitalize the word Court,” which clearly doesn’t capture the substance of the AI assistance provided. Conversely, Gunder notes, including the entire iterative conversation in a citation would be “unworkable.”

Potential Ethical Violations

Perhaps the most serious concern with the rule, Gunder argues, is that it could force attorneys to violate their ethical obligations.

The duty of confidentiality under Model Rule 1.6(a) broadly prohibits lawyers from revealing “information relating to the representation of a client.” This duty extends beyond attorney-client privilege to encompass information obtained from any source during representation.

When attorneys use AI effectively by providing factual details about cases — such as “how a client described their pain, information about the client’s mental state, details regarding profits” — they create prompts containing confidential information.

If Rule 18.3 were interpreted to require citing AI tool usage in all circumstances, Gunder argues, attorneys would be forced to disclose this confidential information in their court filings.

Similarly, the work product doctrine protects attorneys’ “mental impressions, conclusions, opinions, or legal theories.” The conversational nature of AI tools naturally leads to sharing such protected thoughts, as when “an attorney might respond to generative AI output, indicating that they were concerned that a particular witness was not credible, and asking it to rewrite the statement of facts with less focus on the statements from that witness.”

Why Cite AI At All?

Susan Tanner, the University of Louisville professor, in an article published on Medium, “Disclose, Don’t Cite AI: Why You Should Ignore Bluebook Rule 18.3,” argues for a different approach than the one taken by The Bluebook.

“In 99% of cases, we shouldn’t be citing AI at all,” she writes. “We should cite the verified sources AI helped us find.”

For the rare cases where the AI outputs themselves are relevant, she says, citations should clearly signal their purpose, such as in this example:

OpenAI, ChatGPT-4, “Explain the hearsay rule in Kentucky” (Oct. 30, 2024) (conversational artifact on file with author) (not cited for accuracy of content).

“The key is absolute clarity that such citations document what was said, not the truth of what was said,” Tanner argues.

Implications for Legal Practice

The flaws in Rule 18.3 could have significant implications across the legal profession, critics say. As Gunder notes, law schools must now “teach students how to use Rule 18.3 – including drawing a firm line around when it is appropriate to cite to generative AI output” while ensuring they don’t inadvertently teach students “to violate their duty of confidentiality or expose work product.”

Law review editors will have to develop policies determining when their journals will require AI citations, while practitioners will face particular challenges in jurisdictions with standing orders requiring AI disclosure.

The problem is compounded in jurisdictions, such as Florida, that have adopted The Bluebook as their official citation authority.

‘Deeply Flawed’

I reached out to the editors of The Bluebook for comment but received no response as of yet.

No doubt, these changes reflect the editors’ well-intentioned attempt to respond to technological change, but critics of the changes make a strong case that their execution fundamentally misunderstands both how AI is used and what legal citation is meant to accomplish.

As O’Keefe concludes in his analysis, “This Rule lacks the typical precision for which The Bluebook is (in)famous.”

Gunder calls the rule “deeply flawed” in multiple ways, ranging from its internal errors, to the unreasonable burdens it imposes, to its incompatibility with how AI is actually used, to its potential conflicts with ethical requirements.

Tanner says the rule “fundamentally misunderstands how large language models work, conflates documentation with citation, and creates burdensome requirements that miss the point of why we cite sources in the first place.”

Based on these criticisms, it seems that The Bluebook’s editors would be well-served to revisit and reconsider Rule 18.3.

Or maybe, as O’Keefe suggests, they won’t have to.

“The good news,” he writes, “is that the next issue will likely be entirely AI-written and lack any of the lacunae that mere mortal drafters are doomed to leave.”