How to Cite AI

Welcome to AutomatED: the newsletter on how to teach better with tech.

Each week, I share what I have learned — and am learning — about AI and tech in the university classroom. What works, what doesn't, and why.

In this week’s piece, I discuss how to cite AI. While citing AI isn’t without its challenges, there are solutions that go a good distance towards addressing them. Before that, I share some links of interest to higher educators, including some intriguing news from OpenAI and Canvas.

📢 Quick Hits:
AI News and Links

1. Canvas offerings will now include AI-powered “Intelligent Insights” from Instructure and “Teacher Tools” from Khan Academy.

2. Phillipa Hardmon tries to answer the question “How Close is AI to Replacing Instructional Designers?” Short story: [ID + AI] > [non-ID + AI] > [ID - AI]. A tentative result, given the methodology, but worth examining.

3. How to “easily tell if students copy and paste [your assignment prompt] into ChatGPT” by deceiving them with hidden text. Issues with this method abound, but it is instructive nonetheless. (For more comprehensive solutions, see our ✨Premium Guide on the topic.)

4. “In this article, we design and report on the first preregistered empirical evaluation of AI-driven legal research tools. We demonstrate that the providers' claims are overstated.”

5. OpenAI released SearchGPT, a competitor to Perplexity and Google, to a limited user group. They tout it as “a temporary prototype of new AI search features that give you fast and timely answers with clear and relevant sources.“

6. A recent study found that “ChatGPT outperformed the average student in seven classes [including Technology, TESOL, Introduction to Teaching, English, and Mathematics], securing A grades or higher in 10 out of 11 instances” and that “graduate teaching assistants accurately identified 50% of AI-generated assessments.”

7. GPT-4o Mini is now out (a much smaller, cost-effective model that is comparable to GPT-4 or better), and so is the open-source Llama 3.1 from Meta, which will be integrated deeply in many Meta products, is freely available to all, and is comparable in its 405B form to the closed frontier models (GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro).

8. 18% of US faculty agreed or strongly agreed that generative AI will have a positive effect on the instructional practices of their field over the next five years. Skepticism of generative AI’s value was highest among humanities faculty, where 45% disagreed or strongly disagreed that AI’s impact on instructional practices in their field would be positive. Yet, ~30-40% find it very or extremely valuable for instructional uses from personalizing lesson plans to creating study guides… Is that cognitive dissonance or did I misunderstand the survey results?

💡 Idea of the Week: 
How to Cite AI

Style guides like the MLA and the APA have been grappling with how to cite AI for more than a year, but it’s safe to say that no consensus has emerged — in part because there isn’t a simple or obvious way to cite AI in many cases.

Still, we can make some progress by reflecting on the goals of citation and how AI works.

Indeed, many of us must make progress of this sort if we are to teach our students how to cite AI, especially as it plays a greater and greater role in our fields and courses.

My Recommendations

To cut to the chase, what comes of my own reflection process — which I explain and justify below — are the following recommendations:

• As AI gets more and more intelligent, we should cite it to shirk more and more credit for the work we use it to produce. We need to admit what we are not responsible for, and we need to explain the degree to which we aren’t responsible for it by explaining the role AI played in its creation.

  • This requires us to document what of our final products we used AI to create (e.g. a draft that evolved into one section of a paper), what the AI was at the time that we used it (e.g. a specific LLM version), and how we used it (e.g. a prompt sequence; more on this below).

• (Side note: If AI achieves personhood, we should cite it to attribute credit to it for the work it produces, at least if we would do the same with a human in parallel circumstances. You may be confident this will never happen, but some experts aren’t so certain, as I note.)

• In order to enable others to verify what we establish or produce with AI, we should cite the AI in a way that either (a) provides evidence for it directly or (b) explains how we established/produced it so they can retrace our steps.

  • This will typically require us to discuss our AI methodologies and to provide appendices with all their gory details (e.g. full prompts), just like a scientist who uses a specialized tool or statistical technique to achieve a result and must explain how exactly they did so.

• The previous recommendation also helps others with their research. As they explore our citations of AI, they should be empowered to see how AI can be used to establish similar claims and results in the vicinity. This is parallel to how citations of journal articles reveal to readers who relevant domain experts might be and what relevant work in the domain exists already.

The Purposes of Citation

I arrived at these recommendations by first reflecting on the purposes of citation.

By getting a grip on them, we can get a better grip on how we and our students might cite AI to achieve them.

In most fields, the primary purposes of citation are:

1. Credit Attribution: citations give credit to the people who are partially or fully responsible for the cited work

2. Credit Shirking: citations make clear that the author of the citation is not fully responsible for what they cite — or the cited work that supports it

3. Verification: citations enable the verification of claims that are established or corroborated elsewhere

4. Research Facilitation: citations make it easier for readers, viewers, and listeners to discover relevant experts and relevant works

There are other purposes, of course, but I take it that these are the main ones.

In order to justify the preceding recommendations, I will now very briefly explore how AI interacts with each of these purposes, and then I will discuss the recommendations of the MLA and APA style guides.

Purpose 1: Credit Attribution

The question of giving credit to AI hinges on whether AI can be responsible for what it produces, which, in turn, depends on the degree to which it is a person or agent. A paintbrush can get no credit, an elephant “painter” can get some, and Monet can get a whole lot more.

Some philosophers argue that certain AI systems may already be conscious, which is a small but crucial step towards personhood. If AI achieves person status, responsibility would likely follow shortly thereafter, and all of the standard considerations about human credit attribution would become relevant.

However, as this remains highly uncertain at best, we must consider the other purposes of citation when it comes to current AI.

(It’s because of similar reasoning — linking authorship to responsibility, and assuming that AI cannot currently be responsible — that COPE argues against AI authorship.)

Purpose 2: Credit Shirking

Even if credit cannot be attributed to AI, that doesn’t mean that it belongs to us! Sometimes, we should shirk credit and point to a tool as the proximate cause of what we produced.

For example, even if ChatGPT is a mere non-person tool, this doesn’t mean that a student shouldn’t flag that they used it. Even if the AI tool doesn’t deserve any credit for helping a student brainstorm, say, it doesn’t follow that the student deserves all the credit for the resultant ideas.

Purpose 3: Claim Verification

Verification involves tracing claims back to their source and assessing their evidential support, validity, plausibility, etc.

With AI-generated claims, verification becomes problematic unless we have:

1. A stable, citable output that contains the AI-generated claims’ evidential support, validity, plausibility, etc. (The output itself is enough to verify the claim.)

or

2. A reliable method to reproduce the relevant output via the same AI tool or otherwise. (The output itself isn’t enough.)

The nature of AI complicates both aspects.

Citing stable AI outputs that support claims is challenging when chat logs aren't publicly accessible in perpetuity (or at all), when screenshots of AI interactions would be excessive or hard to provide, and when the method by which the AI arrived at the claim isn’t transparent (as in the case of ChatGPT’s Advanced Data Analysis).

And the obvious solution — to provide a method or process by which one arrived at the AI-produced claim — is compromised by changing LLM versions, inherent randomness in prompt-output relationships, and personalized contexts (e.g., ChatGPT's custom instructions). These factors make it difficult to replicate exact outputs consistently, even when the prompts are provided word-for-word and file-by-file.

Still, in both cases, we can go a good distance towards a solution by providing appendices with AI tools’ outputs, our full prompts, and as much information as would be useful for someone else to at least see how we arrived at the point we did at the time that we used the tool.

Purpose 4: Research Facilitation

Citations traditionally lead readers to domain experts, sources, and works relevant to further research in the same vicinity as the cited claim.

For AI to fulfill the role of domain expert, it needs to be a person or an agent, but it isn’t (yet) — this returns us to the issues related to credit attribution.

AI can be a reliable source — a way to explore or substantiate new claims — but, again, its reliability is dependent on our ability to reproduce the same sort of outputs produced by its prior users (revisiting the issue from claim verification). We need access to specific model versions that rapidly become outdated, the ability to prompt in the same way as in the past, etc.

At best, we can facilitate further research by documenting how we used AI in a given case with great detail, thereby showing others how they might complete similar tasks with similar tools, even if they can’t retrace our precise steps.

The Recommendations of the MLA and the APA

Last year, the MLA released some recommendations on citing (generative) AI tools.

Here they are:

Practically, the MLA recommended using their template of core elements, which translates to the following:

I agree that we nearly always need to check AI-generated sources (although maybe SearchGPT marks a turning point on that front) and that we should acknowledge many of our uses of AI tools (“all” seems too strong, given minor uses and ubiquitous AI-software integration).

I also agree that we shouldn’t (yet) treat AI as an author, although it is notable that the folks behind the Chicago style disagree on the latter point, at least with regards to citation formatting.

However, these recommendations suffer from a range of serious issues that you can surmise from my preceding discussion.

While naming the tool, its publisher, the date the content was generated, and the general URL of the tool is somewhat useful for shirking credit — and perhaps attributing credit to the developer of the tool — they don’t enable verification and enable research facilitation to only the slightest degree.

If you tell me you used OpenAI’s ChatGPT-3.5 on May 18, 2023 to help write a key section of your paper or establish some claim, that doesn’t help me check behind you. (I already know about ChatGPT in general!)

Moreover, it’s hard to see how a prompt could be included in a “Title of Source” field, at least in the case of real-world academic use cases. These involve prompts that are lengthy, complex, and include a range of file types.

A recent prompt I used to generate a draft of a grant application with Gemini 1.5 Pro was 450,000 tokens!

But I shouldn’t be too harsh; this was a first stab by the MLA more than a year ago.

The APA recently updated their guidance on citing AI (which matches the APA journal’s requirements), and it gets closer to the mark in recognizing some of the above issues. They recommend:

• Describing how you used the relevant AI tools in the “Methods” section of your paper.

• Providing full prompts used — and when they are too long, placing them in an appendix.

• Noting AI tools’ version names, dates of use, etc. as much as is feasible.

• Treating AI tools’ outputs like the outputs of algorithms, and citing their creators’ as authors (of the tools, not their outputs per se).

• Not treating unshareable conversations with AI tools as belonging to the “personal communications” category, despite the similarities, because AI is not a person.

I think these recommendations are close to the mark because they enable credit shirking, verification, and research facilitation, while also giving credit to the developers of the AI tools. (I am not confident that treating AI tools’ outputs like the outputs of algorithms makes sense, given the role of the prompter/user, but this is a minor point.)

So, to sum up and apply the preceding to the teaching context, my view is that professors should…

• Require students to describe and document what they use AI tools to produce, how they used the relevant AI tools, and what the AI tools were at that time (creator, version, date used, etc).

• Use methodology sections and appendices to show how they used the relevant AI tools. These sections should either (a) provide evidence for the result or product of the tool directly or (b) explain how they established/produced it so we can retrace their steps and conduct our own research in the vicinity.

For instance, in the short-term context of a single semester, a lower effort route with ChatGPT is to require students to provide their chat logs via the “…” button on each chat (it appears after mousing over the chat on the left side menu; click “Share” next), along with citations flagging where ChatGPT was used, a corresponding bibliographic entry, and a discussion of their methods.

🤖 Our Course Design Wizard GPT

Have you tried out our course design GPT? Give it a try if you are working on your courses for the next semester/term!

It can produce assignments, assignment sequences, rubrics, and AI course policies. I have designed it to be especially effective when it comes to pedagogical issues related to AI.

I am constantly working to improve it, so please give it a rating or email me to let me know what you think. I plan to release an update to it in the next few weeks and I am trying to get as much feedback from professors as I can!

✉️ What You, Our Subscribers, Are Saying

Unfortunately, the folks who answered “Yes, other” — in the poll from my last newsletter on the topic of requiring students to purchase AI tools — did not tell me what AI tools they have in mind…

✨Upcoming and Recent Premium Posts

Next Wednesday - Tutorial: 3 Ways Profs Should Use Gemini (1.5 Pro)

This Wednesday - Tutorial: 3 Ways Profs Should Use ChatGPT (GPT-4o)

July 3 - Guide: How to Train Students to Use AI

June 19 - Tutorial: Easy Student Consent Management in Microsoft 365