Citations for AI assistants like ChatGPT ... are they enough?

The governing bodies of many of the major citation styles (like APA, MLA, and Chicago) have published guidelines for citing generative AI assistants like ChatGPT, Google Gemini, Microsoft Copilot, and Anthropic Claude. 

But if you look at the posts listing the way to cite them as a source, you'll notice something. They're clear on how to write the citation.

They don't give much guidance on whether you should -- or how you should use what AI provides you.

I've done lots of workshops with educators -- and participated in discussions online -- about AI in the student writing process. And every time, someone brings up: "MLA has a way to cite AI." 

It's up to the writer to determine whether AI is appropriate in that situation. That can be a tough decision to make -- and one our students (and all of us) will wrestle with for the rest of our lives.

Should you cite AI? Well, would you cite Google?

 If you've taught students writing and researching long enough, you've probably seen it. 

On a bibliography page or works cited -- or in a footnote of some sort -- a student lists "Google" as a source.

(Funny story: One Sunday at my church, the pastor cited the source of a statistic as "reddit.com". I wanted to say, "Have you ever seen some of the stuff they post at Reddit?" Poor guy. He had no idea.)

When you see it, you know ... it's time to have that discussion about what is and isn't an actual source. Here's how I would approach that discussion:

• Google isn't a source any more than saying "I found it at the library" counts as a source. ("OK, can you name the source at the library?")
• Google is a place to find sources. It directs you to all sorts of places online where you can find the original information in context.
• Google might direct you to an inaccurate, biased, or outdated source. Check your facts out. (It's like the old journalism maxim ... "Even if your mother says it, check it out.")
• Cite the most direct sources that you can.

AI assistants like ChatGPT summarize text and create text based on what they have learned from their datasets (their libraries of information to draw upon) and from their interactions. 

Citing an AI assistant is like passing along second-hand or third-hand information. It might be true ... but then again, it might not be.

AI and its relationship with sources: An analogy

I've tried to wrap my brain around the relationship that AI assistants like ChatGPT have with the facts, with their sources, with reality.

Here's the best analogy I can come up with. (Note that I'm from Indiana, a very agriculture-based state, so if this analogy seems crude, my apologies ... I'm just speaking the language of my people. 😂) 

AI assistants produce information kind of like a meat grinder produces sausage.

At the top of a meat grinder, you can add all sorts of meat ... all different parts of the animal ... and it all goes into the meat grinder.

What comes out? Hotdogs. Sausages. Ground beef and sausage and ... other things.

What would happen if someone questioned the butcher later and asked, "What is this little part I see here?"

Likely answer: the butcher would shrug and say, "I'm not exactly sure. It's just sausage now."

The AI models that power AI assistants like ChatGPT learn from a variety of sources. For example, some more reputable (published books and academic journals) ... some less reputable (personal blogs and popular Reddit posts).

AI models don't learn in a linear fashion like we'd prefer -- with direct links back to everything it produces. 

Instead, they're more like our own brains. (Because their architecture is inspired by the human brain.) We file away lots of things in our memory and often say, "I can't remember exactly where I heard this, but ...".

We wouldn't cite "I can't remember exactly where I heard this" from a friends, family or colleagues. Unless we can see the information in context, there will be doubt.

When might you cite AI as a source

So why in the world would reputable citation sources like MLA, APA, Chicago and others give us citations for something we might not even want to include?

Because there might be times when using AI as a source is, indeed, the most direct source for our information.

Example: When we're talking about AI and want to show what an AI model produced.

Let's say you're writing a research paper or an article or even an essay about generative artificial intelligence. To provide an example of an output by the AI model, you might need to include the text it created so you can comment on it.

Because, in that case, it's the source of the original text -- the exact example that you want to include -- you would quote it directly and cite it in your bibliography.

There are lots of times, of course, when you would NOT use AI as a direct source.

Another analogy: Let's imagine that a politician claims something as fact in a speech. A writer could report that the politician said it, and that would be fact. But would it be a reliable source? Just because someone said it doesn't make it fact. (It might make it fact that the person said it, but again ... that's only helpful if we are speaking to the way the politician was communicating and not trying to gather facts.)

Just because AI said it doesn't make it fact ... and it doesn't give us permission to state it as such.

Can you "plagiarize" AI assistants?

Let's say that an AI assistant creates a response to a prompt. Sometimes, writers (not just students, but adult writers, too!) will use the whole response verbatim in a paper, an essay, an article, or some other form of writing.

This underlines a concern that many of us educators have had since AI assistants like ChatGPT have been available -- we don't want students "plagiarizing" with AI.

(Did you notice my quotation marks around the word "plagiarizing"?)

Everyone throws the P word around with reckless disregard. I understand why, but there's a problem.

You can't plagiarize AI.

(Note that I'm not saying copy/paste/submit from AI is OK.)

Here's why: the concept of "plagiarism," as it's defined today, has to do with passing someone else's work off as one's own. The idea of "someone" being another human.

Black's Law Dictionary defines "plagiarism" as: "The act of appropriating the literary composition of another, or parts or passages of his writings, or the ideas or language of the same, and passing them off as the product of one’s own mind."

It's based on intellectual property -- the creative work of a human being.

Prompt an AI assistant to create some text for you. Then ask it who owns the response. It'll tell you that it doesn't own it. Why? Because it's not human.

What does this demonstrate? That our definitions aren't keeping up with the rapid pace of technology.

So ... what do we do? Let's be precise about the words that we use. What don't we want students to do? We don't want them to pass off AI generated text as their own. And then, let's be clear and specific about how they can (should) (might) use AI in their work and how not.

AI responses are impermanent and fleeting

If all of the previous reasons weren't enough, there's also this issue: when AI creates a response, it's hard to refer others to it.

When AI responds to a prompt, it's likely that the response lives in your account for that AI assistant. But it's likely not published to the web and made to be a finished product.

Sure, lots of the AI assistants do allow you to create a shareable link to a response. But those are intended for temporary, informal access to the response ... not as publication of a solid source.

The idea behind cited sources is to promote trust -- to allow the reader to review the claims, to check them out, to see the facts in context so they can make their own judgments.

Citing an AI response that can't be accessed and verified doesn't do that.

Which AI assistants cite their sources well?

Today, I prompted several major AI assistants -- ChatGPT, Google Gemini, Microsoft Copilot, Anthropic Claude, and Perplexity -- with a question: "What are the most populated cities in the United States?"

A caveat: I asked this question knowing that it was directly verifiable with data online. But here's the problem: large language models weren't created to be a replacement for search engines. They're not really made to look up facts and real-time information. AI models take static data and train on it, looking for patterns, commonalities, and consistencies. Then they make their best guess at what they think the answer should be. So, by admission, I'm asking it to do something it wasn't made to do. But it does shed light on the differences between the assistants and their practices on citing sources.

Here's what I found, ranking from most forthcoming with sources to least:

• Perplexity (free version) started with a source -- linked and summarized and right at the top -- before it even presented its response to my question.
• Microsoft Copilot (free version) linked to sources in footnotes with its initial response to my prompt -- without my having to ask.
• ChatGPT (free version based on GPT-4 model) gave a detailed response with population statistics -- but without sources cited. I asked for sources and it asked me if I wanted it to look it up using real-time information. It then provided the same stats as before but with a link to World Population Review -- whose information contradicted the ChatGPT data slightly (swapping a different city into fifth place by a very narrow margin and adding two New York City boroughs as cities after listing NYC as first overall).
• Claude (free version based on Claude 3.5 Sonnet) listed 10 cities and said it couldn't cite sources, explained that the data came from its original training data, and suggested sites that would have that type of data.
• Gemini (paid advanced version) listed five cities and said it couldn't cite sources. Its brief explanation said that the data came from its pre-training and that population rankings are "widely known" and can be easily verified.

So, what did we learn from this exercise?

• Some AI assistants -- Perplexity and Copilot -- lead with sources even if we don't ask.
• ChatGPT would provide sources if prompted but didn't start with them.
• Claude and Gemini wouldn't provide direct sources to verify their information.

The standard we expect in classrooms would not match the effort provided by some of these AI assistants!

Four practices we can implement with AI assistants and sources

If we want to get this right -- and help students to improve their thinking and their skills -- what can we do?

Here are four ideas I like related to that topic ...

1. Use AI when we have a clear reason to do so -- and when it helps us reach our goals. Provide students with guidance about what AI can help them do so that it helps them to develop as human beings. And discourage their use of it when it doesn't help them to think or develop skills. Most students aren't getting any guidance whatsoever. Be that trusted adult who has their best interest at heart.
2. Encourage transparency with use of AI. If students are going to use AI in their work, encourage them to be transparent with it. Disclaimers are a good starting point. Encourage them to make a note at the end of their work disclosing how AI was used in their work. If AI is used directly -- copied text or AI-generated media -- encourage them to label it.
3. Use AI as a way to get back to the original source. Again, remember that AI wasn't really made to do this in the first place. But if an AI assistant can help you get back to an original source for information or data, students can then cite that more direct source. Citing the most direct source is better than citing the means of getting to it (i.e. an AI assistant or search engine).
4. Skip AI all together and search for direct sources yourself. If AI is best at creating "original" text and not best at providing data that's verified by a source, consider not using it at all. Use the searching tools that do provide the sources you're looking for.

Generative artificial intelligence is still new technology to many of us. We're trying to figure out how it fits in our lives -- and in our academic work.

Be the guiding force for students to help them understand the implications of it -- and when it should and shouldn't be used.

We get the opportunity to cast a vision and set the culture for how this technology is used responsibly. Let's not miss out on it!