AIを活用したCHATGPTで銀行はどう変わるか

ChatGPT is a smartbot recently released to the public by OpenAI, which specializes in content generation and is heavily funded by Microsoft. GPT stands for generative pre-trained transformer, a large language model (LLM) using deep learning technology. According to OpenAI: “GPT is different from other AI in that it is specifically designed to generate human-like text. It uses a transformer neural network architecture, which allows it to process input data and make predictions more efficiently than some other types of language models. GPT is also trained on a very large dataset, which allows it to generate text that is more diverse and coherent than models trained on smaller datasets.”

Transformer models, developed in 2017, remain the state-of-the-art of deep learning models. Every time you use autofill, you’re benefiting from an LLM.

So why the buzz over a six-year-old technology? Representing a scaling breakthrough for machine learning, ChatGPT and other generative AI models are poised to dramatically change the way we generate content across all industries. The tech can answer questions, summarize topics, write short essays, songs, and poems, translate, and write and debug code.

First, it’s trained on 175 billion parameters, able to generate responses in under 30 seconds, and support multi-turn conversations. Second, it is simple to use. Third, it has been released to the public for free (for now) and gained one million users in its first two weeks. That’s just the beginning: OpenAI’s next model, GPT-4, to be released in 2023 could have one trillion or more parameters, making it more accurate and faster.

What does this mean for banking?

We took a look around the corner and extrapolating from current use cases of natural language processing (NLP) for how this new solution could help banking.

At its base, LLMs enable us to access information and perform tasks more efficiently, turbo-charging our productivity and intellectual potential.

Use cases can be placed in a pyramid of needs as shown below. Currently, ChatGPT and LLMs in general will power enhancements in the banking industry’s “Tell Me” and “Do it for Me” uses cases, which are at the base of the pyramid. LLMs will expedite real-time assistance to customers and employees and provide basic answers. They can not only assist in generating text but also code. Coders report that ChatGPT can help them generate better code faster as well as debug code.

Figure: Potential Banking Use Cases for GPT Models

While it is clearly early days in the application of GPT, it is not too early to begin anticipating how it and other deep learning models will transform engagement among customers, employees, and banking institutions as well as content generation by the financial services industry.

Celent anticipates:

• Interest in the power of LLM will bubble up from younger employees experimenting with ChatGPT, who then share their impressions with their managers, who then in turn share their thoughts with senior management. We’ve already heard this from customer-facing fintech employees.
• A few AI tech vendors, specializing in banking, will refine the GPT-4 model (once it’s released) by training it with banking data.
• A few banks will experiment with GPT-4 to expedite the generation of training materials for employees and then expand to customers.

Celent recommends that banks seek a point person to track trends, collect and distill ideas and experiences from colleagues, and map potential use cases.

What ChatGPT won’t do

ChatGPT has a social contract. It won’t answer questions with malicious, harmful intent. It cannot be used to determine when to approve use of essential services such as medical care or credit.

It also has limits. No one knows how ChatGPT arrives at its answers and the wrong answers are well articulated. It also provides caveats with cautionary warnings that often end with “It is important to carefully consider the potential risks and challenges of using these technologies, and to ensure that they are used in a way that is transparent and ethical.” As ChatGPT puts it, “[my] knowledge is limited to what was present in my training data, and my training was completed in 2021.” In other words, it doesn’t know about the world beyond 2022.