4 minute read
Biting the ChatGPT Bullet
by Henrietta Verma
ChatGPT. If it seems like it’s all you hear about, all the time, it’s no wonder: the artificial intelligence (AI) platform surpassed 1 billion users in March 2023, after launch the previous November. The platform—which isn’t the only AI tool available but is getting the most attention—had a million users after five days, a milestone that took Netflix more than five years. Lots of the attention is negative; most damning is that users report that the system provides some wrong answers and fake citations. The pitfalls are real, but in my view, they are part of what makes the tool fascinating and imperative for librarians to master.
What is ChatGPT?
ChatGPT is a “generative” AI tool, with “GPT standing for “Generative Pre-trained Transformer.” Created by the OpenAI company, it is offered as a freemium resource¹ While users will find there a Google-like interface, it isn’t a search engine. Rather it generates text in response to “prompts,” with the text based on OpenAI’s ingest of web data such as Wikipedia. Answers are formulated by the system predicting the next likely word based on word occurrences in the ingested text. These answers, then, are grammatically correct, but they may not be factual, a problem stated up front on the ChatGPT homepage.
There are Problems…
As well as unreliable output, there’s of course the issue of plagiarism. Plagiarism policies will now have to note that text must be by the writer rather than stating that the writing must be original, because ChatGPT output is original.
Bias is also an issue. As noted in Safiya Noble’s brilliant Algorithms of Oppression, “Institutional relations predicated on gender and race situate women and people of color outside the power systems from which technology arises.” Simply put, society’s biases are reproduced on the internet and are therefore inherent in ChatGPT. And OpenAI introduces bias of its own in preferring politely worded prompts. Politeness standards aren’t universal, and the question is, whose version is preferred by the platform, and to what end is OpenAI’s pushing of this anthropomorphism?
Then there’s the problem of copyrighted material being used to generate answers. Books are part of the material used to respond to prompts, so that some authors’ intellectual property is being exploited without compensation. ChatGPT can also be used in competition with copyrighted material—when users ask the platform to generate art, for example. Jobs are also at risk, since some employers will surely choose to have AI-generated content on their websites and other output rather than paying content creators.
1 at https://chat.openai.com
2 https://www.stevehargadon.com/2023/03/library-20s-chatgpt-bootcamp-for.html
3 https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/
4 https://prompts.chat
5 https://www.youtube.com/watch?v=N9XaLNfExgM
Making Lemonade
What does this mean for librarians? An opportunity.
First we will need to drop any knee-jerk rejection of the platform as a whole (except if it continues to be able to rework itself into something devastating by writing its own code, a problem that has caused AI pioneer Geoffrey Hinton to quite Google’s AI initiative). We already went through this with Wikipedia and Google. They’re not perfect and we’d prefer library users to use library resources. But since patrons will use what’s easiest, and our job is to help them to be better information consumers, our task now is to help patrons to use ChatGPT effectively and ethically. Like Wikipedia and Google, the key is to use the resource as a starting point for work, something to emphasize both to those who think it shouldn’t be used at all and those who want to use it for everything.
Start with becoming familiar with using ChatGPT for your own work. If you can, watch the Library2.0 webinar series “ChatGPT Bootcamp for Libraries and Librarians” 2 and read Stephen Wolfram’s What is ChatGPT Doing…And Why Does it Work? 3 Set up a free ChatGPT account and play with the “Awesome ChatGPT Prompts”4 and come up with some of your own based on your needs. Some great ways to use ChatGPT are for creating lesson plans, paper outlines, and “boilerplate”-type text.
Playing with the system, you’ll notice that once you get your initial answers, it’s possible to ask the system to tweak them. You can ask for more (type “continue” in the prompt box) and tell the system that something it said is wrong and ask it to try again. You can also paste in the prompt box some citations you’ve already found on a given subject and ask the platform if there are any relevant sources missing. Use ChatGPT in a recursive way by asking for the best ways to use ChatGPT to research your topic. Or make other comments of your choosing—ChatGPT can receive all kinds of natural-language input, and it responds in an amazingly (albeit sometimes frighteningly) natural way. You can even specify the audience an answer is for and the system will answer appropriately.
Once you’re comfortable, you can move on to patrons. Just as we show patrons catalog and database searches during reference interactions and information literacy instruction, show them ChatGPT. Its problems are teaching opportunities—discussions of bias, academic integrity, fact-checking, and privacy (ChatGPT saves searches) are the bread and butter of information literacy.
Introduce your students to—or in a public library, use concepts from—Mike Caulfield’s Web Literacy for Student Fact Checkers to help users evaluate what they find. Lean on the fake-citation problem to lead users to your databases. No fake citations there! Teachers and professors can also be taught to use the citation issue as a quick way to check student work for ChatGPT use. The problem that plagiarism detection software can’t catch AI-generated text can be used to encourage educators to use alternative kinds of assignments and/or to gather examples of student writing produced in class for comparison. To make users aware of the bias issue, use Courtney Thomas, Jr.’s TedX talk “Tech Bias and Algorithmic discrimination” 5 in addition to Noble’s book mentioned above.
To educate users on copyright issues, there is much on the internet about whether users can copyright ChatGPT answers (no, as the user provides only an idea in the form of a prompt, and ideas aren’t copyrightable). On what to do with the problem of copyrighted materials being used to generate answers, practice with this prompt: “is copyrighted material used in Chatgpt.”
Is there more to it? For sure, but you don’t need more from me— just ask ChatGPT!