Everyone is talking about it!
Like many other school librarians, I have several questions about the newest artificial intelligence, OpenAI ChatGPT. So I thought, what better way to get answers than to ask my questions using the software itself? I must tell you that the interview was fun and even a bit addictive. A few questions about using the tool became a much deeper search. 
Here are some of the initial questions I asked the chatbot:
- Is ChatGPT a tool that librarians and teachers can use, or is it a threat?
- How will the integration into the Bing search engine help one’s ability to search?
- How can one determine the legitimacy of answers given in research-based queries?
The interview continues.
Once I received answers, I began to ask even more about the tool. For instance, can a ChatGPT Bing search for a research-based query provide source information, and identify bias, provide corroboration, evidence, and expert opinions? Further, how should school librarians teach research differently with AI-aided searching? And are there sources already available?
When I found that both Google and Microsoft have plans for AI in their search engines, I wanted to know how they compared to one another. I asked, is there information about how different Bing’s ChatGPT and Google’s AI Bard will be in the search process? When does Bing ChatGPT launch, and will there be instructions on search techniques? And when does Google’s Bard launch, and will there be instructions on search techniques? Then based on the precise information given, I made even more technical requests like “please provide source information for “LaMDA” (Language Model for Dialogue Applications).”
And continues!
I found myself asking “one last question” again and again. So here is a four-part question. It required four questions because I was not getting the exact response I was looking for in the results.
- One last question, you have covered the basics of evaluating responses and sources given in AI-assisted search; how can you help to provide both vertical and lateral reading to help give a fuller understanding of a topic?
- Give me an example of a vertical and lateral reading of AI-assisted search and its use as a tool in research.
- What are sources recommended for vertical and lateral reading on the ChatGPT and Bing search engines?
- Can you provide specific resources for vertical and lateral reading that one might apply to the AI search? Can they be accessed online?
In conclusion, these new tools may be key to engaging students and faculty in the never-ending search for information that leads to lifelong learning. And I found that sometimes it takes several tries to ask the most constructive question.
Important to note
You will notice in the linked transcript that many of the hyperlinks given as resources are dead links. This is because ChatGPT draws from data published in 2021 or earlier; I believe this is a risk you encounter using ChatGPT right now for research. Here are several articles comparing ChatGPT with its counterpart BARD from Google.
Bard vs. ChatGPT: What’s the difference?
Top 5 Differences Between ChatGPT and Google Bard AI
For a full transcript with permission [from the chatbot] – http://www.librarianlittle.com/2023/02/an-interview-with-chatgpt.html
Nice article Hannah, but … is not one of the reasons why “many of the hyperlinks given as resources are dead links” is not because ChatGPT’s data is 2021 and the pages have been moved or closed as you imply, but because ChatGPT does not draw upon source material at all, and when asked for sources, it tends to make them up?
John
Hi John,
In this case most of the resources were LibGuides and other Library specific resources that we often change and move over time.
Certainly the AI is in experimental stages. It is making guesses and using templates in its responses. What I am trying to ascertain is how this plays out once it is part of the search engines. At that point librarians need to understand and make use of the AI enabled search so that we can teach research skills.
We are planning a webinar in the AASL Independent School Section and I am hoping some tech experts can shed some light on our questions.
Thanks so much for your comments,
Hannah
Thank you for exploring this–I’ve been wondering all of these things and I know we have to be realistic about the use of this new tool. No matter what our opinions are, students will be using it so we need to become knowledgeable about its benefits and drawbacks. I’m hoping it won’t erode creative writing among teens any more than other technology has!
Hi Hannah,
I think you’re misunderstanding how ChatGPT works. John is basically correct in stating that ChatGPT “makes up” sources. ChatGPT is not searching the internet in a conventional sense. Instead, out of a large corpus of information, which includes some information gleaned from the internet, ChatGPT figures out what to say by trying to predict the next likely words (or phrases or characters) that will make sense in context. To anthropomorphize it, it’s basically thinking “Ok, the user asked for sources, what information do I have in my corpus about what an appropriate response would be? Looks like there’s often an ‘https://’ in the answer. Looks like the portion following ‘https://’ is often ‘xyz’. Looks like the portion following ‘xyz’ is often ‘abc'” and so on. That means that it will sometimes generate legitimate urls because the parts of the url appear in sequence often enough for ChatGPT to consider that a likely sequence. But it will often generate made up urls because it’s just “guessing” what a likely url that’s a response to the question might be.
To check whether the links were ever real, you can plug them in to archive.org. Of course, archive.org isn’t perfect, but you’ll see that most of the links generated by ChatGPT probably never existed.
Greetings – I just wanted to follow up on the URLs and resources. I checked all of the links I could find in the transcript against the Internet Archive’s wayback machine, as well as live, and only three of the fifteen appeared to be functional. Two of these were in the first six resources (in the libguides section) and one of these was an ArXiv.org DOI that was actually a different title and author than what was listed.
Generative AI is somewhat of a black box, but the way it functions is not(1). It *computes* the type of word (noun, verb, etc.) that you are looking for and then computes what kind of word from the database of text it pulls from for what content word is most likely to appear. The job it has is to generate words, and so sources/urls are rather beside the point. Any generation of an actual working url (the Berkeley libguide, for example) is likely to just be an exception as it sticks the pieces of Berkeley and libguide and the topic together. It kind of worked for the ArXiv DOI (in that the url works, not in terms of citing anything). And this formula does not work for all of the other 12 dead links that were never archived by the Wayback Machine, a fictional Harvard LibGuide among them.
(1) I recommend Stephen Wolfram’s excellent article on this topic:
https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/