There have been numerous groundbreaking Artificial Intelligence (AI) advancements over the last year. Generative AI, which has exploded in popularity since last November with the release of ChatGPT, introduces major opportunities to boost productivity and advance NLM's capabilities. These tools also present some limitations that all users must be aware of.
Episode 8, “The Future of AI in Medicine” was recorded last summer before Generative AI exploded onto the public consciousness. So we felt compelled to follow up with a bonus episode guest produced by Dianne Babski, Associate Director for Library Operations at the National Library of Medicine.
In this bonus episode, Dianne interviews NLM scientists Dr. Sameer Antani and Dr.
Zhiyong Lu to explain how generative AI is transforming the scientific landscape. They discuss ChatGPT, large Language models (LLMs), AI hallucinations, and more.
As mentioned in this episode, all of the artwork for NNLM Discovery has been created using a generative AI text-to-image tool. We have added the text prompts for each image to the shows’ descriptions. Here is a summary of the entire season’s imagery:
NNLM Discovery show art: “A woman makes a fascinating discovery”
Episode 1 - Virtual Health: “Telemedicine at the library in Texas impressionism”
Episode 2 - Healthy Garden, Healthy You: “a Van Gogh style painting of garden boxes”
Episode 3 - Moby Bookmobile: “Landscape painting of plains with Native American books in Wyoming”
Episode 4 - Collection Equity: “oil painting of the diversity inside a library”
Episode 5 - Amateur Radio Club: “Ham radio operators on the islands in the style of starry night”
Episode 6 - Medical Librarians: “Medical librarians working hard in style of Gustav Klimt”
Episode 7 – Period Poverty: “A woman silhouette standing in front of a red ocean in the style of Van Gogh”
Episode 8 - Future of AI in Medicine: “an expressive oil painting of a robot and doctor working together looking at an x-ray”
Episode 9 - Speaking of Science - Lauren Porter: "Researchers talking about science in the style of Picasso"
Episode 10 – AI in the Scientific Landscape: “In the style of Georgia O'Keefe, paint me a landscape that shows a future with library users embracing AI technology”
00:00:03:01 - 00:00:28:16
I’m librarian Yamila El-Khayat and this is NNLM Discovery a podcast from the Network of the National Library of Medicine. This podcast series explores how NNLM is conducting research, development, and training for the purpose of improving the publics health. Today’s bonus episode is “A.I. in the Scientific Landscape.” This story is a continuation of our episode, “The Future of A.I. in Medicine,” from Season 1.
00:00:28:18 - 00:00:36:20
Dianne Babski, Associate Director for Library Operations at the National Library of Medicine, will be telling our story today. Hi Dianne!
00:00:36:22 - 00:00:37:21
00:00:38:01 - 00:00:42:00
What brings us back together so quickly after wrapping up Season one?
00:00:42:07 - 00:01:00:23
Well, “The Future of A.I. in Medicine,” podcast episode was actually recorded last summer but released this May. Generative AI, also known as Gen AI, a subset of traditional AI, has exploded in popularity since last November with the release of ChatGPT.
00:01:01:01 - 00:01:13:04
Yes! I hear mixed reports about AI in the media. Some say it will have a positively impact in our work and life, and some say the future of AI is all doom and gloom. What do you think?
00:01:13:05 - 00:01:47:23
Well, I think there are some major opportunities to boost productivity and advance our capabilities, but there are some limitations that we need to be cautious of. At this stage, I'm excited about how we are imagining new possibilities for the scientific landscape and striking a balance between leveraging the benefits of AI and mitigating its risk. During this episode, we'll be explaining the basics of this new technology and how it fits into the workflow of future information professionals, researchers and practitioners.
00:01:48:01 - 00:01:58:17
You know what? I really hate to admit this, but I was totally in the dark about GenAI until I attended your presentation at MLA in Detroit. Why is explaining this technology something that you're excited about?
00:01:58:20 - 00:02:25:05
Well, I'm really excited about the potential that A.I. offers and how it will enable us to achieve our mission. I think what we're seeing is as big and life changing as the advent of the Internet or the release of the first smartphone. What makes this technology so revolutionary is it's easy to use. You don't need to know any programing languages or have special software or need to know any syntax.
00:02:25:14 - 00:02:32:15
Now we're all capable of using this conversational language to interact with highly intelligent machinery.
00:02:33:01 - 00:02:40:08
Our listeners aren't aware, but they've already been exposed to Gen AI during the first season of this NNLM Discovery podcast.
00:02:40:10 - 00:02:44:16
So are you real or artificial Yamila?
00:02:44:18 - 00:03:01:07
I'm real, Dianne, but we did use Gen AI a lot this season. Every episode of NNLM Discovery has original artwork. All of this artwork has been created and sourced from prompts written in the Gen AI tool called DALL-E.
00:03:01:09 - 00:03:12:14
That is so cool. I had no idea until we were doing this podcast. I need to go back and check out the art. What sorts of prompts do you use to create artwork in DALL-E?
00:03:12:18 - 00:03:38:16
I'll give you an example. The previous A.I. episode called “The Future of and Medicine” had artwork that was created by typing in the prompt, an expressive oil painting of a robot and a doctor working together, looking at an x ray. The main show artwork for the NNLM Discovery Podcast series was generated from a pretty vague search, which was a woman makes a fascinating discovery.
00:03:38:18 - 00:03:45:21
Check out each show's description where we've added what prompts were used to create this artwork. Dianne kick off our story.
00:03:45:23 - 00:04:17:18
Okay, so let's start with, what is ChatGPT? ChatGPT is a product developed by OpenAI and instead of coming up with my own explanation of what it is, I asked the expert, ChatGPT. And here's how it answered. And I quote, “ChatGPT is a conversational AI model developed by OpenAI. It is a state of the art language model that has been trained on a vast amount of text data from the internet.
00:04:17:20 - 00:04:54:07
ChatGPT is designed to generate human like responses to user inputs, making it suitable for a variety of conversational tasks such as answering questions, engaging in dialog, providing recommendations and assisting with information retrieval. It can understand and generate text in multiple languages and has the ability to generate coherent and contextually relevant responses.” End quote. What's fun about ChatGPT is that you can build on responses with new prompts without retyping your questions.
00:04:54:13 - 00:05:20:23
And for someone like me that is typing challenged, that is a great opportunity. As another example I asked ChatGPT to recreate its previous response on “what is ChatGPT?” To give me a response at a third grade level. And here's the response. Quote: “Sure. ChatGPT is like a really smart computer program that can talk to you and have conversations.
00:05:21:01 - 00:05:45:22
It knows a lot of words and sentences because it has read and learned from many books and articles on the internet. You can ask it questions, tell it things, or just have a friendly chat with it. It will try its best to understand what you're saying and give you helpful answers. It's like having a really smart friend who can talk to you about almost anything you want to know.” End quote.
00:05:46:00 - 00:05:51:03
Wow, that's pretty amazing. It summarized the previous version so well.
00:05:51:18 - 00:06:23:20
As the Director of Library Operations, I think it's critical to understand how we're using generative AI within our own walls at NLM. So I interviewed a few researchers here to help share and explain these new advancements. We'll be talking to two NLM staff today who specialize in artificial intelligence research. The first is Dr. Sameer Antani. Dr. Antani is a principal investigator in the Computational Health Research branch of the Lister Hill National Center for Biomedical Communications.
00:06:23:22 - 00:06:35:23
Doing research in medical imaging, machine learning and artificial intelligence. I started our conversation by asking Dr. Antani what has changed since we recorded the previous episode.
00:06:36:01 - 00:07:06:20
So thank you for having me back. Indeed, a lot has changed the perception of and the power of AI has become more commonplace. Average citizens know about AI now and the access to the tools or the potential. It has raised the awareness that AI could take away their jobs in some senses, but it could also be a powerful ally for improving your day to day activities.
00:07:06:22 - 00:07:34:04
A common type of generative AI are large language models or LLMs that can recognize, summarize, translate, predict and generate text and other content based on knowledge gained from massive datasets. What I love about Gen AI is that you can interact with the chat bot in an iterative way and ask it to adapt its position, sentiment, emphasis or other requirement and it will generate a new output.
00:07:34:06 - 00:07:38:00
That's right. Just like how we asked it to re-answer our question for a third grader.
00:07:38:02 - 00:08:03:03
Exactly. So, I also brought in another researcher at NLM, Dr. Zhiyong Lu, who leads a group studying artificial intelligence and machine learning in the computational biology branch of the National Center for Biotechnology Information, more commonly referred to as NCBI., and asked him, What are these large language models that are powering ChatGPT?
00:08:03:04 - 00:08:37:07
They are called large because they are really large in two ways. One is they are trained with vast amounts of text data. Billions of text words. And they are also large because the models contain billions of parameters. Not only is trained by words, but also has a step that uses reinforcement learning that takes human feedback to improve its performance.
00:08:37:09 - 00:09:05:23
That's sort of one of the secret sources of what ChatGPT is compared to the previous version of GPT. Even though with everything nice about ChatGPT, but its output can be completely false because in the end it does not consult the truth. And that is the problem we know as a hallucination in ChatGPT’s output.
00:09:06:01 - 00:09:12:01
Hallucinations? Dianne, are we really talking about hallucinating artificial intelligence?
00:09:12:02 - 00:09:29:23
Hmm. Yes. A hallucination is when AI gives a response with high confidence, but the response is not supported by its own training data. Another example of AI hallucinating is when technology forgets its software but behaves as a human.
00:09:30:01 - 00:09:41:13
You know what? I've experienced that before. My daughter had me use Snapchat's “My AI” feature within the app. The chat bot definitely thought it was human. It kind of freaked me out.
00:09:41:19 - 00:10:14:15
Yeah, we're all freaking out a little bit. “My AI” is another great example of generative AI. Now, one caution for everyone to understand is that if you're using the open version of these tools, any information you provide, these models, including personal information, become integrated in the AI system. Because, like Dr. Lu mentioned, it's learning from your inputs. So don't enter private information like names, grant or pre published information.
00:10:14:17 - 00:10:17:16
Got it. I'll definitely have to make sure my kids know that.
00:10:17:22 - 00:10:42:00
We've been talking a lot about ChatGPT. But this is just one product utilizing generative AI. Embracing these new and emergent technologies is essential for NLM to better achieve its mission. Here's Sameer Antani again. So Sameer, it's not just ChatGPT though. This is a new wave of AI that has reached the public's awareness.
00:10:42:01 - 00:11:16:14
Exactly, exactly. It's not just ChatGPT, which of course, is a product. And generative A.I. is the underlying idea where the A.I. is able to synthesize facts, synthesize knowledge, synthesize context, synthesize information. That is derived from things it has seen and learned that could be on the web, in literature, in art, and then presented before you in the style that you would like to get that information in.
00:11:16:15 - 00:11:54:00
A nice example is where, if somebody is planning a vacation and wants to know what are the highlights to hit in a three day visit to a city, perhaps. It could generate an itinerary for you where you could be able to hit those highlights in the amount of time. What it might not be able to do, however, is give you current information about what is open or what is closed or what's in high demand, and you should have purchased the tickets six months ago. So that is where it fails. At the same time, it's extremely powerful in generating information.
00:11:54:15 - 00:12:18:17
Sameer makes another good point here about a limitation when using Gen AI. ChatGPT is limited because it does not possess any real world knowledge beyond what its been trained on. Its responses are based on patterns and examples it has seen in training responses to the data and its model. It's also constrained by that date stamp on the data used in its model.
00:12:18:18 - 00:12:39:03
Currently, ChatGPT’s knowledge cutoff date is September of 2021, so it does not include information since that date. Consider how the results would be limited if you were asking about COVID 19 vaccines. So, Yamila, let's talk about how GenAI fits in at NLM.
00:12:39:05 - 00:12:40:05
Let's do this.
00:12:40:16 - 00:12:55:23
I asked ChatGPT to generate a list of some of the limitations using generative AI in the scientific landscape. It generated a spot on list and pretty much all of the risk we're thinking about at NIH. Why don't you share a few?
00:12:56:04 - 00:13:17:10
Okay, so there's accuracy and reliability. I get that already. We must be cautious of hallucinations and bias. Reproducibility. Mm hmm. Ethical considerations, intellectual property, cost, authorship and attributions. Wait a minute. So is A.I. already being used in scientific papers?
00:13:17:18 - 00:13:24:01
Yes. We already have nearly 700 articles in PubMed that reference ChatGPT.
00:13:24:03 - 00:13:33:00
I see what you mean when you compare this to the invention of the internet. This is a very big change. This must raise all kinds of new issues for NLM.
00:13:33:05 - 00:14:01:12
This technology has advanced faster than we've had time to develop policy and procedures. That said, the International Committee of Medical Journal Editors, also known as ICMJE released an update to their authorship recommendations concerning AI as an author. It specifies that authors should not list AI and AI assisted technologies as an author or coauthor, nor cite A.I. as an author.
00:14:01:14 - 00:14:30:10
And over the last month we have seen instances of citations in PubMed with ChatGPT listed as an author corrected and removed—showing that the publishing community is being responsive. I would encourage disclosing whatever and whenever you are using A.I. generated content in your work, especially as it relates to scientific research and publications. For this, transparency is key.
00:14:30:12 - 00:14:39:14
It's hard to keep up with how fast this is being implemented. So how do you see this technology being used by NLM or scientists or even by librarians?
00:14:39:16 - 00:15:15:03
A.I. could unlock new pathways of biomedical discovery, operational efficiencies, and a better user experience. Gen AI can aid and summarizing articles, synthesizing data, drafting responses to queries and even basic coding in languages like Python and Sequel. That's why I've been sharing so much with staff on using this technology. That said, we still need humans in the loop to ensure integrity and shape this developing technology in a way that creates usable outcomes in a safe and secure environment.
00:15:15:05 - 00:15:20:12
I asked Sameer how he sees Gen AI advancing the field of medicine.
00:15:20:14 - 00:16:06:20
So generative A.I. could be actually a very powerful ally to medicine. The entire spectrum from looking up literature that is meaningful, minimizing the amount of time that the physician has to spend or doing mundane tasks and allowing them to spend more time with the patient. Where the power lies in a large language model transformer technology is that you can actually ask it to deconstruct the many, many, many pages that you would be forced to read and summarize them into a response that you want, in the kind of context that you have described yourself to be in.
00:16:06:22 - 00:16:23:20
So now it is actually generating novel language, new language, that didn't exist before, within which are the facts, presumably, that are drawn from these sources and placed together in the right order.
00:16:23:22 - 00:16:53:22
What is the value of doing, of us doing research any more? It's more like you just know how to write a prompt to ChatGPT and ask ChatGPT to answer. That's true. But in reality, what you'll find is even with ChatGPT, it does not solve real world problems in medicine. It’s just way more complicated and more complex than you would realize.
00:16:54:00 - 00:17:16:16
So there's always need for scientists like us for students to learn how to adapt a new technology and put that and modify it and improve it for our own domain applications. And that's where I think where the future is for a lot of our research.
00:17:16:18 - 00:17:21:00
And how do you see this A.I. being rolled out and used within libraries?
00:17:21:17 - 00:17:56:19
A library is a resource of information that's trusted. And with all we have described about technologies such as ChatGPT, which are only going to get more powerful, it will be a an important decision to consider how they are woven into the products and services that the library offers. While maintaining the issue of trust. The biggest risk to users of a library or authors is it will synthesize citations of which don't exist.
00:17:56:21 - 00:18:25:13
It might give you real references. It knows what the style of the references are. There’s authors first, title in the middle, published journal after that and the volume and page numbers. And if it does not know, it could synthesize something as made up names, a made up but believable title, a made up but believable journal name, and a made up volume number.
00:18:25:15 - 00:18:51:07
But then the point of being a trustworthy tool is defeated. So the problem of trust is definitely improving, but it also means that while you may have hundreds of thousands of factually correct responses, you could have two wrong ones which might end up actually hurting somebody than helping them.
00:18:52:08 - 00:19:20:04
So my long term vision is that we'll be able to leverage trusted health information at the National Library of Medicine and beyond to create specialized large language models to enhance easy access to health information using everyday language prompts. Can you imagine the impact that would have on reducing myths and disinformation? Sameer, how do you see these generative A.I. tools being utilized within the NIH and NLM?
00:19:20:09 - 00:19:59:20
For people such as us at the National Library of Medicine who are doing research in computer science, largely. Generative A.I. can synthesize code actually, reducing the time between which an idea is imprinted in one's mind. So as long as the scientist is involved in this process and does not treat OpenAI output as gospel, then what you have is a very powerful tool that can speed up the process of learning, and empirical science will definitely benefit.
00:19:59:22 - 00:20:28:12
Another example we are doing right now is to ask ChatGPT to serve as a evaluator to evaluate machine generated radiology reports. This is also related because this is a multimodal AI where we actually feed the AI algorithm with both the patient image and their previous information. And this is something in active research that we are doing right now.
00:20:28:14 - 00:20:42:06
Another concern with A.I. is introducing unintentional bias when dealing with large datasets. How do you make sure that there's not bias toward one segment of the world's population, especially when dealing with medicine and health?
00:20:42:16 - 00:21:30:07
We trying to develop A.I. algorithm that can limit the bias in the data. And that's active research area, where different kinds of methods called debiasing methods are developed. To counter those issues in the data sets. Another aspect of dealing with bias in A.I. algorithms is to develop the algorithm that is more transparent, instead of a black box system. Because, the better you can explain your results to the users, regardless, it’s the end users or general public or physicians, they will better understand how the system works, so that’s also very important.
00:21:30:09 - 00:21:53:23
I agree with you. And that's why we're using the NIST AI risk management framework in our work to ensure that our use of AI is done in a way that is safe, transparent and trustworthy. And for those that don't know, NIST is the National Institute of Standards and Technology. So most important question, Dr. Lu, how will AI transform the way we work?
00:21:54:01 - 00:22:18:04
Uh, I think in the future, anybody who knows about AI and uses AI, is it going to replace the person, the same person not using the AI. Okay, so physicians who know how use AI is going to replace the physician don't know how to use AI. So this is something that people have been talking about in the literature and talking about in our community.
00:22:18:06 - 00:22:39:13
I agree that A.I. will force us to continue learning and evolve in our work. The potential for productivity gains and growth opportunities are far too great to ignore. I encourage everyone listening to play around with an experiment a little with these chat bot tools. Well Yamila, now that you know what Gen AI is, what do you think?
00:22:39:15 - 00:22:43:05
I think I need to start checking on how my teenager's doing her homework.
00:22:44:22 - 00:23:14:12
An interesting use case is from a group at Stanford University who developed a generative A.I. tool called BioMedLM which uses the data from PubMed and PMC trained to interpret biomedical language. So, to test their system they used complex questions from the biomedical board exam with highly accurate results. The world around us is changing rapidly, and we're changing how we work to meet these new needs.
00:23:14:14 - 00:23:47:23
And I hope that what you have heard today is that NLM is embracing this new world in our research and our information products! We are evolving to keep up with these changes, especially in information management practices. I encourage our listeners to consider their information management strategies, especially as they relate to data quality and making data accessible and interoperable so we can harness the full potential of AI and the future scientific landscape.
00:23:48:01 - 00:24:05:14
So, Dianne, before you go, I mentioned this at the beginning of our chat. Every episode of NNLM Discovery has original artwork. You will be the first person to give us input on what our artwork should be for this episode. What prompts should we use when we generate this episode's art?
00:24:05:15 - 00:24:21:05
Well, I'm excited to help with this artwork. So one of my favorite artists is Georgia O'Keeffe. So how about in the style of Georgia O'Keeffe paint me a landscape that shows a future with library users embracing AI technology?
00:24:22:06 - 00:24:31:12
I can't wait to see what the AI pumps out. Remember to look back at all of our AI artwork this season. Dianne, thanks for joining us today.
00:24:31:14 - 00:24:36:05
You are so welcome. I've really enjoyed the podcast series so far.
00:24:36:17 - 00:24:47:22
Thank you for joining us for this special episode of NNLM Discovery. Subscribe to our podcast and follow us online and remember to rate, share, and comment on our show wherever you get your podcasts.