By Ray Birch
MADISON, Wis.—GPT—or generative pre-trained transformer—tools may be the most useful forms of artificial for credit unions to use today, contends one expert, who has outlined several credit union use cases for the technology—but also several challenges.
“AI and is something that I found fascinating from the very beginning,” said Ignacio Luri, assistant professor of marketing at DePaul University, who partnered with Noël Manushi, finance and marketing student at DePaul on the Filene study titled Text-Generative AI and the Future for Credit Unions. “Everyone is focusing on how AI generates text. But it is also able to consume data. It understands it and it can read and process text for you.”
GPT is a language model tool used to decipher and generate human-like text.
Luri said the “normative” implication for credit unions is that ChatGPT should be seen as a powerful tool to aid copywriting, translating or rephrasing, or combing through text-like consumer reviews, harnessing its skill for language use.
But the popular tool with which many have experimented is not something else, according to Luri.
“ChatGPT and closely related GPT-based tools should not be seen as a replacement for human decision-making nor as a knowledge repository,” he said.
Luri offered the following considerations for AI:
TextGen AI Opportunities
- The most promising future of TextGen AI is in ChatGPT-like technologies as a component for existing IT systems, Luri said.
- GPT tools can be used in creative brainstorming tasks, such as for suggesting slogans for a marketing plan or suggesting SEO-friendly Google ads. “This is particularly the case when there is no one right answer, and the built-in randomness of the model can be harnessed as an advantage,” Luri explained..
- GPT translations could open the door for underrepresented communities. “Credit unions must keep in mind the nature of the data with which most GPT models have been trained and the biases introduced by the reliance on these models,” Luri said.
- Hybrid large language models (LLMs) can use GPT as a complement to tools.
Ignacio Luri
“For example, Bing AI is, in essence, a ChatGPT interface that reads and understands what you ask it, turns your question into an internal, invisible Bing search, and, instead of giving you links, reads out to you the combined results in natural language,” Luri told CUToday.info. “Likewise, a GPT user interface can be given access to financial algorithms so that credit union members can interact with them in simple words—like asking about a financial goal and being told the savings rate they need to achieve it. ‘Arming’ TextGen AIs like GPT by giving them access to tools—a simple financial calculator, a list of financial products, real life stock prices, etc.—represents the biggest opportunity in this field.
“The only requirement is to have GPT models communicate with other applications via application programming interfaces (APIs),” Luri added.
TextGen AI Challenges
Lusri said GPT and other LLMs create responses that are “probabilistic” in nature and as a result create some challenges.
“These technologies give a range of responses that roughly relate to their frequencies on the Internet data on which they were trained,: he said. The most likely percent or dollar amount given in online conversation around questions like this will thus appear most often, but the probabilistic nature of the model means that far less frequent responses will be given by GPT by chance,” he said.
According to Luri, additional challenges include:
GPT and other LLMs are Limited by their Training Data
“LLMs are only as good as their training data. While the data uses diverse AI algorithms, it is an important source of bias as well—it reproduces the opinions and voice of a white middle-class, pro-environmental, left-leaning human,” Luri said. “It overrepresents English sources and it is still trained with data that ends in January 2022—two years ago at this point.”
GPT is Not Capable of Computation
“ChatGPT and other TextGen AIs do not have a built-in calculator,” Luri said. “In fact, a simple hand calculator or web application will outperform them on financial computations such as interest rates, mortgage calculations, etc.”
GPT Models are Guided and Biased by Human Evaluation Criteria
“GPT is known to accept false premises from users too easily and to hallucinate, confidently stating made-up facts that helpfully respond to user prompts but lack any grounding,” Luri explained. “Early research on the topic of finance already suggests that hallucinations introduce a risk of overreliance on GPT as a financial advisor or in conversations about the social responsibility of financial institutions.”
GPT is Not Capable of Live Access to the Internet
“TextGen AI models learn ‘to speak’ using vast Internet databases, but once trained, they have access to statistical patterns only, not to the actual sources—i.e. they lack live access to the Internet,” Luri said. “The models do not update or change until the next iteration of their technology is released.”