CBA Record July-August 2023

YOUNG LAWYERS SECTION

The Advantages and Dangers of AI Chatbots By Nelson Rosario

I n March 2023, over 1,000 people involved in the research and devel opment of artificial intelligence (AI) signed a letter urging a moratorium on the development of certain AI systems. The signers warned that “AI systems with human-competitive intelligence can pose profound risks to society and humanity.” (“Pause Giant AI Experiments: An Open Letter,” Future of Life Institute, https:// futureoflife.org/open-letter/pause-giant ai-experiments/). They advocated for a moratorium on development of AI sys tems that are more powerful than GPT-4 until they develop a framework to address their associated risks. You may be asking, “What is GPT-4?” To some, GPT-4 may sound familiar due to its use in ChatGPT. But what is Chat GPT? As explained below, it’s a chatbot, and lawyers—like it or not—will need to become familiar with this technological development. This article will define what an AI chatbot is, why lawyers should care about them, and some of the advantages and disadvantages of AI chatbots lawyers need to be aware of.

What is an AI Chatbot? An AI chatbot is an AI-powered computer program or application that is designed to simulate human conversation and pro vide automated responses to user inqui ries or prompts. Chatbots utilize natural language processing (NLP) techniques to understand and interpret user input and generate appropriate responses. In fact, the previous two sentences were gener ated from a prompt to ChatGPT asking, “What is an AI chatbot?” One of the most popular new kinds of chatbots is ChatGPT, developed by an organization called OpenAI. Newer chat bots like ChatGPT are an interactive way to conduct research, generate content, and conduct business. The “GPT” stands for “generative pre-trained transformer.” It is the most common approach used as part of building large language model (LLM) artificial intelligence. These LLMs have been “compared to supercharged autocorrect engines” in part because they operate largely by “predicting” what word to produce in response to a prompt in the context of the prompt. (“Pause Giant AI Experiments: An Open Letter.”) Put

another way, these chatbots can contex tualize information in relation to other information to produce query results that can be quite useful. But how are these chatbots able to contextualize information? One way is based on how the chatbots are trained. In essence, these chatbots are fed tremen dous amounts of text-based information, usually from the internet, that helps them “know” certain things that allows them to respond to queries and prompts. More specifically, the generative pre-trained transformers are a type of neural network that, when fed information, can iden tify patterns and give the appearance of “knowing” things and being able to inter act with a human user. In addition, these chatbots can “learn” from their users by adding past interactions with users and past queries to their repertoire of knowl edge. The chatbots are also able to take on certain personas and respond accord ingly. For example, feeding the prompt “Respond to the question, ‘what is an AI chatbot?’ as a computer scientist would” to ChatGPT produced the following response:

38 May/June 2023