There has been considerable interest, discussion, and even concern about large language models over the last six months. Large language models? Some readers are likely thinking, “I am not familiar with this term. What are large language models?”
If large language model doesn’t resonate with you, maybe ChatGPT and Med-PaLM are familiar terms. If we scan the digital health landscape, these terms are unequivocally the hottest topic going.
The world of artificial intelligence (AI) encompasses tools, techniques, and models for using data with the end goal of allowing machines to think like humans. Similar to the goal of using AI, the field of AI is complex in that it includes subfields, terms that are used interchangeably, hierarchical technologies with varying levels of sophistication and capabilities, and constant evolution.
These aspects of AI can make it challenging to make sense of advancements in the field. Considering the recent significance of ChatGPT and Med-PaLM, this column is devoted to those terms and their implications.
We will begin with a term that you may be more familiar with than large language models: chatbots. Apple’s Siri is the most widely used virtual assistant, with Amazon’s Alexa and Microsoft’s Cortana as other common examples. But I thought we were talking about chatbots?
The AI Field
Chatbots and virtual assistants are similar, with some suggesting they are essentially the same technology. Virtual assistants are sometimes referred to as “conversational AI” because of their ability to respond to spoken comments or requests and perform tasks asked for by the user.
Chatbots are similar in that they respond to human questions (usually in written form) via messaging apps. The chatbots most of us are likely to encounter are commonly deployed by companies to interface with their customers, whereas virtual assistants are consumer oriented.
This is where the challenges of making sense of the AI field enter the picture. ChatGPT and Med-PaLM (more later) are at the core, chatbots. They respond to questions posed by the user. They are much different, however, from the typical chatbot you interact with when you click the “Chat” help window on a website, or talk to virtual assistants like Siri.
Those technologies are limited to a finite number of questions, requests, and answers. Alternatively, ChatGPT and Med-PaLM are sometimes called “AI chatbots” because they enable human-like dialogue between the technology and the user.
Summary of Chatbot Resources
ChatGPT and privacy concerns in Italy
For ChatGPT’s self-identified and
Which brings us back to large language models. Large language models (LLMs) are a type of AI that creates natural language text in response to a user’s question or request. Large language models are the engine, if you will, for AI chatbots like ChatGPT and Med-PaLM. How does a large language model actually answer a question?
As a type of artificial intelligence, LLMs scan, analyze, and summarize large amounts of text. By large, we are referring to 570GB and 300 billion words (from the internet) in the case of ChatGPT-3 (). By learning from such a large amount of text, LLMs allow prediction and generation of text to create a conversational experience for the user.
Simply stated, ChatGPT and Med-PaLM allow users to ask a question or provide a textual prompt, and the tool answers the question or prompt with a textual response. Well, that seems great, right? Who wouldn’t want to ask a question and receive an extensively researched answer?
Apparently, a lot of people like that idea. Two months after its launch, ChatGPT reached 100 million users, compared to nine months for TikTok and two and a half years for Instagram to reach that same number of users ().
Chat Limitations
If ChatGPT is so popular, why was its use temporarily banned in Italy? And why are educators concerned about it? In Italy, the primary concerns were privacy and access (https://www.bbc.com/news/technology-65139406). For privacy, the Italian data protection authority expressed concern over how ChatGPT uses personal data to develop the training algorithms for the software.
In terms of access, the concern centers on the potential for minors to be exposed to inappropriate content. Broadly, legislation is being developed in the European Union to regulate artificial intelligence due to perceptions that they may “deceive” people.
Closer to home, in educational settings of all levels, educators have expressed concern that ChatGPT (and comparable tools) can be used to write essays or similar types of papers. However, a few relevant limitations are worth noting.
First, at this time, ChatGPT’s training data is current through 2021. So it cannot answer questions seeking factual answers based on information in 2022 or beyond. Second, questions or prompts need to be framed in a certain manner for ChatGPT to answer them. If the question isn’t framed correctly, the tool guesses the intent. This can lead to inaccurate answers.
And last, according to the folks at ZDNet, it is more likely that the sources ChatGPT uses for an answer are wrong rather than right (). This highlights an important limitation of AI chatbots in general — they do not truly understand questions. Instead, they use statistical procedures to construct responses.
Considering these limitations, why has ChatGPT’s user base grown so quickly? Can you envision a high school or college student who procrastinated in writing a term paper? ChatGPT to the rescue! Or have you ever wanted to go deeper in the answer to a question than a simple Google search provides? For those who write software code, ChatGPT is a viable source of support.
ChatGPT can also be used for common documents like résumés or cover letters. But maybe the novelty of having a conversation with AI is partially driving interest. Along those lines, I took a few minutes to test ChatGPT. I asked, “What are the benefits of ChatGPT?” and I also asked, “What are the limitations of ChatGPT?”. For ChatGPT’s self-identified benefits, go to , and for limitations go to .
Maybe the novelty aspect is a significant factor influencing the number of users. But what if a large portion of users are relying on the information ChatGPT provides? In many cases, like trivia night, incorrect information from ChatGPT will not have significant impact.
And maybe a high school student’s grade isn’t what they hoped due to ChatGPT providing incorrect, or partially correct, information. This is where Med-PaLM enters the discussion. Med PaLM is an AI chatbot specifically for healthcare-related questions. I will dig into Med PaLM in a future column. Please send your questions and comments to me. CT
Brent I. Fox, Pharm.D., Ph.D., is a professor in the Department of Health Outcomes Research and Policy, Harrison School of Pharmacy, Auburn University. He can be reached at foxbren@auburn.edu.