Article by Joel Blechinger, Faculty LibrarianCover Image: This work, "The Pope's Puffer", is adapted from "The Pope Drip" by Midjourney, used under CC BY-NC4.0.
In late March 2023, an image that appeared to show Pope Francis in a white Balenciaga brand puffer jacket flooded social media. Twitter users repeatedly shared the image making jokes about the pontiff’s style, with one early Tweet that contained the posted image amassing close to 28,000,000 views and 207,000 likes as of mid-April 2023.
The problem: this image was created with generative AI tool Midjourney by a 31-year-old construction worker in Chicago as a lark, and it was pure fabrication. Actress Chrissy Teigen spoke on behalf of the many who were duped when she tweeted, “I thought the pope's puffer jacket was real and didnt [sic] give it a second thought. [N]o way am I surviving the future of technology,” while web culture writer Ryan Broderick declared this to be the “first real mass-level AI misinformation case.”
Though generative AI tools like image generators (Stable Diffusion, Midjourney, DALL·E 2) and large language models (OpenAI’s ChatGPT) are increasingly used across sectors, their use also encourages the widespread sharing of misinformation and disinformation.
Writing in Scientific American in December 2022, leading AI thinker Gary Marcus presciently warned that “chatbots [like ChatGPT] could be used to mass-produce misinformation.” ChatGPT has been shown to fabricate nonexistent sources and references, and AI-powered chatbots like Google’s Bard and Microsoft’s Bing have erroneously cited one another in what technology writer James Vincent has called “a massive game of AI misinformation telephone.” OpenAI’s own documentation for GPT-4 states that “[t]he profusion of false information from [large language models] … has the potential to cast doubt on the whole information environment, threatening our ability to distinguish fact from fiction.”
AI’s role in the spread of false information reinforces the importance of information literacy education. Information literacy is defined by the Association of College and Research Libraries as “the set of integrated abilities encompassing the reflective discovery of information, the understanding of how information is produced and valued, and the use of information in creating new knowledge and participating ethically in communities of learning.” MRU Library supports the MRU community through the delivery of a robust information literacy curriculum across academic programs.
In light of generative AI’s troubling ability to mislead, what role can information literacy play in this brave new reality?
Amy Scheelke, Instruction and Liaison Librarian at Salt Lake Community College, notes that there are two information literacy techniques that can be used to vet the outputs of large language models: lateral reading and citation verification.
Lateral Reading
Lateral reading, a powerful technique developed by the Stanford History Education Group as part of its Civic Online Reading curriculum, involves reading other online sources to gain more information about the particular source that you are assessing. You could choose to laterally read a large language model’s output in two ways.
Firstly, you could use other credible online sources to read more about the specific large language model that you are assessing. What have been reported as its potential biases, and how might those biases be influencing its output on your particular subject? For example, ChatGPT has revealed common gender and racial biases when prompted to write performance reviews in the past. This shows that, despite the size of the textual corpus that ChatGPT was trained on, that corpus’ prejudices still pervade it as a model and shape its outputs, though OpenAI has attempted to mitigate against bias.
Secondly, you could read across other credible online sources on your subject to vet the large language model’s output. As Scheelke states, “[d]on't take what ChatGPT tells you at face value. Look to see if other reliable sources contain the same information and can confirm what ChatGPT says. This could be as simple as searching for a Wikipedia entry on the topic or doing a Google search to see if a person ChatGPT mentions exists.”
When laterally reading ChatGPT’s output, keep in mind the way that large language models work. ChatGPT has been trained on an enormous corpus of text to predict the next word that follows in a sequence of outputted text. At times, this prediction is inaccurate or wrong.
Citation Verification
Citation verification is a simpler process which entails searching for a source used by ChatGPT to (1) confirm that it actually exists, and (2) verify that it actually contains the information that ChatGPT claims it does. Scheelke advises searching for the source in Google Scholar or LibrarySearch to put sources through that two-step test.
The techniques detailed above presume, of course, that you are aware that text has been generated by a large language model, and that you must vet its output. Many of the concerns about large language models, however, involve their ability to output undetectable, passable erroneous text that can then be used in other contexts to questionable ends.
In the case where you may not know (yet) that text is AI-generated, basic tenets of information literacy can help you scrutinize the source. Information professionals have developed many different acronym heuristics to use in assessing information sources. Mike Caulfield’s SIFT method and RADAR are two particularly effective ones that can be used on any text, AI-generated or not, to determine its quality.
Applying information literacy skills to AI-generated images may be different than text, but there are emerging practices that can help users assess these outputs.
Germany’s international broadcaster Deutsche Welle suggests six techniques that users can employ to detect AI-generated images. These include examining the body proportions of humans in the images, looking for common errors or anomalies (like humans having an improper number of digits), and paying attention to the image’s background to see if it has been reused or artificially blurred.
Another technique that was popularized by fact-checking journalists, but that still applies when examining AI-generated images, is the use of a reverse image search engine (such as TinEye or Google’s Image Search) to trace an image back to where it has been posted elsewhere on the internet. A TinEye search of the Pope image mentioned above, for example, reveals that it has been posted to the website Know Your Meme, which could be a helpful source in assessing its trustworthiness.
InVID-WeVerify’s verification plugin for Chrome is another free tool used by journalists that provides a suite of video and image examination capabilities. The plugin’s toolbox includes a magnifying lens to thoroughly examine an image for any irregularities, along with reverse image searching options for Bing, Yandex, Baidu, Reddit, TinEye, and Google.
AI-generated image detectors have also been developed, such as this one hosted by the organization Hugging Face. Detection systems still have their limitations, and as with AI-generated text detectors, they can quickly become outdated.
If there is a lesson to be taken from generative AI’s powerful ability to misinform, it is that practicing information literacy is of greater importance than ever before. Whether you are prompting or assessing the output of a generative AI tool, it is best to proceed with an attitude of informed skepticism which you can cultivate through the practice of information literacy techniques.
Australian Library and Information Association. (2023, February 21). AI, libraries and the changing face of information literacy: Navigating the complexities of a digital world [Video]. Vimeo. https://vimeo.com/801056924
Badke, W. (2023, April). AI challenges to information literacy. Computers in Libraries, 43(3). https://www.infotoday.com/cilmag/apr23/Badke--AI-Challenges-to-Information-Literacy.shtml
Papini, A. (2023, January 27). ChatGPT: A library perspective. Bryant University - Krupp Library. https://library.bryant.edu/chatgpt-library-perspective
Woods, L. (2023, February 21). AI chat tools and information literacy. CILIP InfoLit Group Blog. https://infolit.org.uk/ai-chat-tools-and-information-literacy-laura-woods/
0 Comments.