(Note: this post was drafted in April 2023, but posted October 17, 2024. It has been backdated to reflect the time of writing, not posting.)
Both the Modern Language Association and the American Psychological Association recently published guidelines for citing the output generated by large language models. The MLA called their blogpost “How do I cite generative AI in MLA style?”, while the APA called theirs “How to Cite GhatGPT”. Both are obviously making an underlying assumption that is at odds with mine in “Why you can’t cite ChatGPT” and “Why you shouldn’t cite, acknowledge, or credit an AI with authorship”. Actually, we agree on a number of important points, but they think that, at the end of the day, you can cite an AI if you want, while I think the very idea is incoherent. In this post, I want to explain why I think this is important.
Consider my last post. I showed there that it’s possible to generate a passable undergraduate essay about The Great Gatsby in 10 minutes using ChatGPT. I was directly inspired to do the experiment by the MLA’s proposed guidelines for paraphrasing and quoting ChatGPT’s output, which surprised me. After all, they rightly “do not recommend treating the AI tool as an author”; what authority does citing an AI then actually invoke? How does sourcing a claim to ChatGPT contribute to the reader’s understanding of the writer’s argument? How could I have acknowledged ChatGPT’s contribution to my essay about the Great Gatsby without simply undermining my credibility as an author?
In their examples, they work through what looks like a standard essay prompt. The student has presumably been assigned an interpretative essay on The Great Gatsby and immediately went home and gave the prompt to ChatGPT:
Describe the symbolism of the green light in the book The Great Gatsby by F. Scott Fitzgerald.
ChatGPT obliged with a numbered list of “various symbolic meanings” that MLA Style then proposed to paraphrase as follows:
While the green light in The Great Gatsby might be said to chiefly symbolize four main things: optimism, the unattainability of the American dream, greed, and covetousness (“Describe the symbolism”), arguably the most important—the one that ties all four themes together—is greed.
In the “Works Cited” list, they propose the following entry:
“Describe the symbolism of the green light in the book The Great Gatsby by F. Scott Fitzgerald” prompt. ChatGPT, 13 Feb. version, OpenAI, 8 Mar. 2023, chat.openai.com/chat.
My somewhat simple-minded question at this point is: Other than that the writer has apparently done zero reading or thinking to determine what the green light symbolizes, what has this citation actually told us? What have we learned about the basis of the idea that it symbolizes optimism, the unattainability of the American dream, greed, and covetousness? Do we know anything more about whether this is a good reading of the novel than we would if the writer had not told us where they got it? The obvious answer is no. Since we know what language models do, we know only that this is a plausible string of words about the topic.
It gets worse when we get to the issue of quotation. This time, they suggest we prompt GPT to produce some prose rather than a list. “In 200 words, describe the symbolism of the green light in The Great Gatsby.” ChatGPT obliges with some plausible prose and the MLA now suggest we ask it for sources. “What scholarly sources were used to generate this description?”
… wait, what?!?!
This question simply AI-illiterate, MLA! We know that language models don’t “use scholarly sources” to generate their output. And that is indeed what ChatGPT patiently explains: “As an AI language model, I do not have the ability to conduct research or cite sources independently.” But then it does what we know it also often does: hallucinate sources; it generates a list of plausible-looking scholarly sources that, as you do discover if you try to actually find them in a library database, simply don’t exist. Unperturbed by this discovery (I suspect because they didn’t check the sources), MLA’s citation experts happily suggest we just uncritically report what the AI told us:
When asked to describe the symbolism of the green light in The Great Gatsby, ChatGPT provided a summary about optimism, the unattainability of the American dream, greed, and covetousness. However, when further prompted to cite the source on which that summary was based, it noted that it lacked “the ability to conduct research or cite sources independently” but that it could “provide a list of scholarly sources related to the symbolism of the green light in The Great Gatsby” (“In 200 words”).
Here’s their suggested entry for the works-cited list:
“In 200 words, describe the symbolism of the green light in The Great Gatsby” follow-up prompt to list sources. ChatGPT, 13 Feb. version, OpenAI, 9 Mar. 2023, chat.openai.com/chat.
Notice how weird this reference is. We’re being told exactly what the original prompt was, but we’re being given a paraphrase of the “follow-up prompt”. It is completely unclear to me what this quotation accomplishes in the text and why it is being recommended by the Modern Language Association.
Next we’ll be given guidelines for “how to cite your mom”!
Much of this can apparently be traced back to a somewhat superficial understanding of what language models actually do. The MLA cites the New York Times for its definition of “generative AI”: “a tool that ‘can analyze or summarize content from a huge set of information, including web pages, books and other writing available on the internet, and use that data to create original new content'”. This is actually not what ChatGPT does. It doesn’t “summarize content”, nor can it summarize the content of “huge sets of information”. Rather, you can prompt it with a short passage from a book or article and it can summarize that.
The MLA agrees with the APA that an AI, like ChatGPT, can’t be an “author”. But neither style guide seems to understand the implications, namely, that you then can’t cite it. APA explains that
the results of a ChatGPT “chat” are not retrievable by other readers, and although nonretrievable data or quotations in APA Style papers are usually cited as personal communications, with ChatGPT-generated text there is no person communicating. Quoting ChatGPT’s text from a chat session is therefore more like sharing an algorithm’s output; thus, credit the author of the algorithm with a reference list entry and the corresponding in-text citation.
It’s great that they don’t suggest citing ChatGPT as personal communication, but what does it mean to cite OpenAI as the “author of the algorithm” when what we’re actually quoting is the algorithm’s output? I have search the blog for pre-GPT guidelines on citing algorithmic output but came up empty. Here’s the example they give us:
When prompted with “Is the left brain right brain divide real or a metaphor?” the ChatGPT-generated text indicated that although the two brain hemispheres are somewhat specialized, “the notation that people can be characterized as ‘left-brained’ or ‘right-brained’ is considered to be an oversimplification and a popular myth” (OpenAI, 2023).
Reference
OpenAI. (2023). ChatGPT (Mar 14 version) [Large language model]. https://chat.openai.com/chat
Is there any context in which this is a useful reference? Do we learn anything about the brain by this means? And if the only thing we’re learning something about is ChatGPT, why do we need to cite it? We’ve already said where the quote came from and what prompted it. What more did the citation provide?
My gut tells me these guidelines will have to be revised in light of how writing practices develop and AI improves.