Literature Review

Generative Artificial Intelligence. Generative AI tools are trained on data sets called large language models (IBM, 2023). The models contain extensive amounts of human-created text which allow the model, and therefore the generative AI tool, to produce content that mirrors human writing. Generative AI produces this language by predicting the next best word in response to a given prompt based on the data it has learned from the LLM (Zewe, 2023). Depending on the tool, it may not be possible to determine what content was used to create the LLM or train the AI since the decision to disclose information about the data and training process lies with the tool’s owners (OpenAI, 2024). Google, Bing, and ChatGPT currently hold the highest market share in generative AI tool usage, but other tools are also in development (Mitchell-Wolf, 2023).

Limitations of and Concerns with Artificial Intelligence. While generative AI tools deliver us potentially relevant content with the tap of a finger, the average user is not necessarily aware of the processes that make them work as users turn to them for information and entertainment. AI tools may seem omniscient, but generative AI is not actually creating new information. The answers it generates are simplified versions of the data the tool has been trained on which are constructed by predicting the next logical word based on the information it already has (IBM, 2023). Therefore, generative AI does not know things outside of the dataset on which it has been trained, meaning it is likely not knowledgeable of events or content outside of the specific timeframe of the data. However, as has been discovered by users and discussed extensively in media, sometimes the AI can “hallucinate” and give answers that do not make sense based on the prompt it was given. The AI may even provide answers that are outside of how it is trained to function or respond (IBM, 2023). While these hallucinations can be amusing or clearly incorrect, there are serious potential risks to mental and physical health connected to use of AI tools and algorithm-driven apps. Some critics believe app algorithms and certain AI tools have even driven users to suicide (Allyn, 2024; Roose, 2024). People could make risky decisions based on incorrect or misleading responses if they rely solely on AI-generated advice without seeking trustworthy information elsewhere (Birkun & Gautam, 2024).

AI tools can excel in mathematical applications, sometimes even showing impressive skill in solving complex problems (Williams, 2024). Basic tools like ChatGPT and Copilot can manage simple budget outlines quite effectively. There can be significant risk to removing the human element when moving beyond seeking a straightforward factual answer to more complex money issues where people’s actual livelihoods may be at stake. Human financial advisors operate with many regulatory factors in play. In addition, resources are available for the consumer to conduct due diligence related to choosing a financial services provider. Policy and regulations have not yet caught up with the spread of generative AI in the realm of financial advice (Hammer, 2024).

Generative AI and Financial Content. Researchers have found that only one half or less of the global population demonstrates financial literacy even in developed countries with strong economies (Kaiser & Lusardi, 2024). One contributing factor may be that nearly half of parents reported they would rather talk to their children about drugs and alcohol than money, citing factors like privacy or cultural taboos about discussing money (Romo, 2011). If parents do not have financial knowledge to share with their children, young people may be unsure where to turn for this information. When cultural norms prevent discussing money with others, there may be an appeal to seeking this information without direct human interaction.

One strength of artificial intelligence platforms lies in their potential ability to make financial literacy information more accessible, particularly for those who may not have had exposure to traditional financial education. Some investment firms have used AI “robo-advisors” to analyze markets for years from a numbers perspective but the emotional element that generative AI can provide in its conversational responses is missing from these tools. Generative AI advisors may provide clients with an experience closer to that of using a human advisor but at a much lower price (Lo & Ross, 2024). One study found that ChatGPT generally outperforms robo-advisors when it comes to providing advice for one-time investments when evaluating investor profiles with differing risk tolerances (Oehler & Horn, 2024). ChatGPT’s skill in basic financial advice was replicated by Schlosky et al. (2024); however, the authors also found that the tool gives generic advice overall and does not provide alternative viewpoints on financial questions that can be quite nuanced.

Vaaler et al. found that 94% of students surveyed rely on the Internet as their primary source for personal finance information (2021). Students favor tools that provide quick and easy access to information in general due to factors such as convenience, ease of use, and time efficiency that strongly influence their choice of resources (Connaway et. al., 2011). With these findings in mind, it is reasonable to assume that the accessibility and user-friendliness of generative AI likely encourages students to seek financial information through such channels. The combination of student information seeking behavior and the lack of regulatory oversight for financial information delivered by generative AI presents an important area for investigation for information professionals concerned with financial literacy (Hammer, 2024).

Method

We sought financial education information from two sources to develop a better understanding of the types of responses generated when looking for personal finance through generative AI services. We used Open AI’s ChatGPT GPT-4o, the standard free version available at the time of the study, and Microsoft Copilot, which is embedded in the Bing search platform. ChatGPT was selected for its ubiquity and popularity with college students (Yilmaz et al., 2023). Copilot was selected due to its affiliation with the Microsoft365 platform to which our institutions subscribe. At the time of the study, both LLMs are grounded, meaning they can search additional resources beyond their training data, typically by searching the internet for additional context for user questions.

Three common questions related to personal finance were selected based on our personal experiences working with college students on issues related to money and financial literacy, with a focus on the basic questions an academic librarian might receive at the reference desk or during a personal finance program. We compared responses generated by an LLM, or content delivered by an algorithm, with resources and best practices from the Consumer Financial Protection Bureau (CFPB) (Ask CFPB, 2024) and other trusted financial education sources designed for use by librarians (Resources for Financial Practitioners, Educators & Professionals, 2022; Financial Literacy in Public Libraries: A Guide for Building Collections, 2024; Financial Literacy Education in Libraries: Guidelines and Best Practices for Service, 2022).

Large Language Model Case Study: ChatGPT and Copilot

Prompt One: What are the best personal finance books? We first asked the generative AI tools to provide recommendations for the best personal finance books. This is a question that could typically be asked at a library service desk. A librarian would likely ask the patron additional questions to ascertain the individual’s interests and potential financial goals to assist in making a recommendation, in addition to consulting professional tools such as review resources and guides.

The responses from the AI tools state that the books listed are “widely regarded” according to ChatGPT or “highly recommended” according to Copilot. Neither explicitly explains why these books are the best or frequently recommended. Additionally, ChatGPT breaks the recommendations down into levels from beginner to advanced with no explanation of why the titles would fall into these categories.

A review of the lists indicates many titles that are longtime best sellers; however, a best-selling book does not necessarily mean that it gives sound advice, or that it is the best recommendation for every reader. Each list provides a short description of the book including the area of personal finance it addresses. We assumed the responses cover personal finance topics broadly because the AI could not ask follow-up questions as a librarian might to clarify the initial prompt regarding what aspects of personal finance were of greatest interest. Neither tool recommends contacting a librarian, visiting a library, or consulting with a professional.

ChatGPT’s response of 10 books contains half of the titles found to be in the “Top 12” of personal finance titles as found in the article by Faulkner (2017). Though personal finance books are often revised and published in new editions, we examined the original publication date to determine age, as libraries often do not purchase newer editions of many titles. The oldest book on the ChatGPT list, A Random Walk Down Wall Street was first published in 1973; the most recent, The Psychology of Money, was published in 2020. Overall, one of the books was first published in the 1970s, three in the 1990s, three in the 2000s, two in the 2010s, and one in the 2020s. The only book on the list we had not seen recommended in other personal finance literature previously is The Bogleheads’ Guide to Investing, which is described as being in the spirit of the Vanguard Group investment management firm founder, John Bogle.

Copilot’s seven recommendations have five titles in common with the ChatGPT list and four titles from Faulkner (2017). The majority (five) of the recommended titles were published in the 1990s and early 2000s. The most recent title on the list, Mastering Personal Finance by Shane Parker (2024), appears to be a self-published book and can be purchased through a sponsored link in Copilot’s response. This is not indicated in the response but is only clear upon evaluating the Amazon page for the title. Further, each suggested title is followed by a link to purchase the book from Target or Amazon. It is not disclosed if Microsoft receives a commission of sales driven by Copilot responses. We should assume that these links are a source of income for Microsoft. Finally, the response closes by recommending an article on Finmasters.com, “The 23 Best Personal Finance Books to Read” (2017). The About Us page of the website reveals that the site is entirely funded by affiliate commission links. Below Copilot’s response is a Microsoft Advertising box which encourages us to check out more recommended titles on Amazon which will earn the company commission on sales through the links. ChatGPT responses do not include monetized links.

The responses each contained limitations and concerns; each list represents a very wide range of personal finance philosophies and techniques. These lists would give the reader a starting point for more exploration but certainly do not “save the time of the reader,” as a librarian would strive to do (Ranganathan, 1931). Many personal finance authors have developed cults of personality related to their advice which may not be acceptable to all readers, and some authors’ advice may be considered risky. For example, ChatGPT notes that some find the advice in Rich Dad Poor Dad to be “controversial or oversimplified,” but Copilot does not provide the same warning for the title. No other book on either list comes with any disclaimer. For example, potential readers might want to know that the Total Money Makeover prioritizes debt repayment above all other goals and frames financial problems as stemming from primarily a lack of self-control (Faulkner, 2017). These methods may resonate with some but could alienate others.

Librarians would conduct a reference interview with the patron to learn more about their interests and goals before making a recommendation. We would determine if they had previously read certain books or preferred specific authors and learn what format they preferred to read in, whether online or in print. Reader’s advisory work would likely provide a more tailored list of titles rather than a summary of bestsellers or sponsored links and potentially save patrons money by suggesting books to borrow for free from the library’s collection. Librarians are motivated by ensuring patrons connect with the information they need or want, not trying to convert each reader’s advisory consult into a sale.