The New Bluebook Rules for Generative AI
The Bluebook’s Twenty-Second edition is now out, and it includes new rules for citing emojis, electronic storage media, and, of course, generative AI.
But these rules attempt to cram a very large square peg into an increasingly shrinking round hole. What I mean is this: rather than recognize important distinctions between born-digital materials (i.e., those created and residing in a digital environment) and those that were born-tangible (i.e., those created in a tangible form, such as something handwritten on paper or painted on a canvas), the new rules attempt to force existing technological outputs into something that resembles a born-tangible to better fit the traditional citation framework. For example, Rule 18.3 (the new rule for citing large language models and generative AI output), requires a user to create a document via screenshot and conversion to PDF that can and must be stored somewhere for later viewing before an author may cite it. The rule also requires a citation to contain “the exact text of the prompt submission” used in the interaction.
These rules are unworkable on so many levels that it’s hard to know where to begin. Let’s start with logistical impracticality. Compliance with these new rules requires technological savvy that many likely do not possess. For example, if one does 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 most conversations with a generative AI model likely span more than one visible page. Second, the process of screenshotting and converting these files is time-consuming, which defeats one of the primary benefits of using generative AI in the first place.[i] Third, users must determine which generative AI interactions they might wish to cite at some point down the road and then find a place to store all these newly created digital documents.
Beyond impracticality from a logistics standpoint, there are other problems inherent in born-digital materials, such as the facts that they are easily manipulated, not static, and incredibly individualized. In the most common generative AI platforms, users’ past conversations with the model can be preserved by choice or automatically. For users, this is a great way to revisit an issue at a later point in time, like hitting pause on a discussion and coming back to it after further reflection. But this also means that, if a user screenshotted a conversation and converted it into a PDF to store somewhere, that conversation may change, and where does that leave the previously stored file? Must it be re-saved with the new discussion added, or may the user simply save the additional discussion separately, knowing the prior interaction is already stored? And, if saved separately, must there be a cross-reference to the prior conversation? Furthermore, conversations with a generative AI platform are iterative, with both questions and responses building on prior discussions, so it is difficult to determine which of many prompts should be quoted exactly in the citation, not to mention that some prompts can be multiple paragraphs long or involve the user uploading a document or image to discuss.
The new rules for citation feel a lot more like rules for evidence preservation than methods for citing authority. According to the Bluebook’s General Principles of Citation, “[t]he central function of a legal citation is to allow the reader to efficiently locate the cited source.”[ii] In both the Bluepages and the Whitepages of the manual, citations are discussed in terms of “authority.” But generative AI platforms are not authoritative on anything, not even themselves. By way of example, I was looking at past conversations I had with ChatGPT to discern ways in which the platform has changed over time, and when I had trouble locating the date of certain conversations, I asked the AI to help me. It then provided dates I knew to be inaccurate. By requiring practitioners and academics to cite generative AI outputs, the Bluebook is cloaking these tools with an authority they do not (and frankly should not) possess.
But there may still be times where a human-AI interaction might require some sort of legal reference, so what are practitioners and academics to do?
One alternative to citation is disclosure. If some acknowledgement of generative AI usage is called for by the situation, a writer can include language disclosing the usage, such as, “Disclosure: Portions of this content were generated with the assistance of artificial intelligence (AI) tools and reviewed by the author. While AI was used to support drafting, all information has been verified by the author to ensure accuracy and compliance with legal standards.” Disclosure ensures transparency without giving a tool the appearance of legal authority.
Another alternative for practitioners is simply using the rules of evidence. If an AI-generated artifact must be referenced in a legal setting, it should be preserved as any other digital evidence and provided within the framework of evidentiary rules as an exhibit. Once it is an exhibit, it can be cited under the traditional legal citation rules.
For academics, any AI-generated output can be included in appendices where reference to the output itself is required (e.g., if the author is making a point about generative AI capabilities or functionality). And then citations may be made to the appendix where the specific material appears.
Whenever we are faced with something new, it is tempting to create new rules to cover the new situation. But it is important to recognize that many of our existing rules can still apply; we need not jump to over-legislation and risk overcomplication.
[i] Additionally, any time a born-digital material is converted into a new format, it changes and is no longer an exact replica of the original.
[ii] The Bluebook: A Uniform System of Citation (21st ed.), pg. 1.