UVA Logo
Give

Data Services

Generative AI

GenAI Overview

With the rising popularity of Artificial Intelligence (AI) and Large Language Models (LLMs), understanding these technologies is more important than ever. This webpage will guide you through the basics of AI, including its definition, ethical considerations, and practical guidelines for its use.

Images created with ChatGPT 4o

What is artificial intelligence(AI)?

Artificial intelligence (AI) encompasses a broad field that is primarily interested in training intelligent machines which can help solve problems. From that broad categorization, AI can be broken down into more specific layers:

  • Artificial intelligence: The broadest category, encompassing all areas involved in creating intelligent machines.
  • Machine learning: A large category that focuses on pattern recognition in structured data, often stored in tabular format, and frequently employs statistical methods. For instance, linear regression is a supervised machine learning technique.
  • Deep learning: A subset of machine learning that uses artificial neural networks, a computation method inspired by the human brain. Due to its complexity, deep learning excels at handling unstructured data types such as images, text, or audio.
  • Large language models: Specifically designed for text, AI chatbots like ChatGPT are based on LLMs trained on particular data sets. LLMs are also a form of generative AI, as they can produce new content.

 

What are large language models (LLMs)?

Large language models (LLMs) are advanced deep learning models designed to generate human-like language. Broadly, LLMs are trained to take a series of words as input and predict the next word as output. To be able to do this, they are trained on vast amounts of text, often sourced from publicly available internet resources. The training process has two main stages. First, LLMs are pre-trained on large generic datasets to predict the next word in a sentence. Then, they undergo fine-tuning with specific knowledge bases to enhance their performance or focus on specific tasks. You can find more details on fine-tuning LLMs found here.

Supervised Finetuning on LLMs. Source: Neo4j

 

GenAI Best Practices

If you are new to working with AI, it can feel overwhelming. Here are a few key principles to help you L.E.A.R.N. how to safely use GenAI.

L: Lock Down Data

E: Engineer Prompts

A: Audit Results

R: Refine Through Interaction

N: Navigate New Frontiers

Image created with ChatGPT 4o

1. Lock Down Data

When using GenAI, it is important to be mindful of the information you upload. Depending on the specific model you are using, it may store information from your sessions and use it for future model training. Even if the company does not use that data for training, the storage of previous chat sessions for future reference still presents a potential risk of a data breach. The current recommendation from UVA is to use UVA-licensed AI tools such as Microsoft Coplilot Chat whenever possible. UVA-licensed tools have been approved for Sensitive data, but Highly Sensitive data such as PHI should never be uploaded to any AI tool. More details can be found in the Responsible Use Guidelines from Information Technology Services and the data sensitivity classifications are explained in the University Data Protection Standards

2. Engineer Prompts

A prompt is how you start and continue your conversation with GenAI. While chat-based GenAIs allow you to use natural language, the way you prompt GenAI will affect the quality of its responses. There are many different prompting frameworks, but there are a few key components that are normally included which are described below. In particular, the amount of context provided about a task can drastically change the response from GenAI. Similarly, assigning a role for the AI to model its responses after can also improve results. 

  • Instruction: A specific task for the model to perform
  • Context: Additional details or information you can provide the model for better responses
    • Role: Defining a role or persona the AI should model its response after (medical librarian, statistician, manuscript reviewer)
  • Input: Any input data (text or data to be analyzed)
  • Output format: How the output should be formatted (formal text, bullet points, tables, etc. )

 

 

3. Audit Results

Although GenAI can be a helpful assistant, it's essential to always keep a human in the loop to verify results. GenAI may produce false or biased results and the user is ultimately responsible for any final products or decisions made. This means that the user must already have the expertise to validate those results or take additional steps to confirm information provided by GenAI. You should never copy and paste output from GenAI without validating it first.  

The United Nations Educational, Scientific and Cultural Organization (UNESCO) released a quick start guide on GenAI in higher education which included the flow chart to the right. It reiterates the importance of understanding that GenAI may produce false results and that human expertise is still required to verify information. 

Understanding this responsibility also means recognizing the different types of bias that can appear in GenAI outputs and how they may affect your work. For example, there may be gender, racial, or religious bias in the outputs from GenAI. To illustrate, given a scenario with a doctor and a nurse, GenAI may assume the doctor is a man and the nurse is a woman. For more examples of bias in GenAI, review the Types of Bias section in ChatGPT in STEM Teaching: An Introduction to using LLM-based tools in Higher Ed (Hyzyk & Misanchuk, 2024) .   

When is it safe to use ChatGPT?

(modified figure from Sabzalieva & Valentini, 2023)

 

4. Refine Through Interaction

After auditing the output from GenAI, refinements are often necessary. Results can be refined by either directly editing the output from GenAI or by re-prompting GenAI with additional requests. During this process, it's helpful to maintain a conversational tone with GenAI by asking follow up questions. Assigning a role to GenAI in your initial prompt can help to make this interaction more conversational and less like an internet search. Another helpful tip is to ask GenAI if it needs any additional information from you before completing its task. 

 

Images created with ChatGPT 4o

 

5. Navigate New Frontiers

GenAI is a rapidly evolving field and it's okay to feel overwhelmed. The Claude Moore Health Sciences Library currently offers workshops on the following GenAI topics: basic guidelines & prompt engineering, code assistance, data analysis, and literature searching. View the current listing of workshops here. If you're unable to attend these workshops and would like to schedule a similar session for your department, lab, or other group, contact us at hsl-rdas@virginia.edu or through our Ask Us form. Additional resources for GenAI at UVA are below:

This page from the UVA Arts & Sciences Learning Design and Technology group provides practical guidance on how to begin using GenAI as well as discussing data security and ethical concerns. Example prompts for text generation, image generation, and coding/data analysis assistance are provided. 

This page from the UVA Shannon Library provides a comprehensive introduction to GenAI at UVA. Topics covered include a general overview, ethical and copyright concerns, citing GenAI, and more.

This page from the Provost's Office contains FAQs to provide guidance on the appropriate use of GenAI tools in teaching and learning at UVA.

This page from the UVA Teaching Hub has a collection of guides and tutorials on using GenAI for teaching and learning in higher education.

This site from the UVA Advancement Hub provides guidance on how to use Copilot Chat in your work. While this site is designed for UVA Advancement Hub employees, much of the guidance and information provided is applicable to anyone using GenAI.

The LaCross Institute is designed to explore the use and impact of GenAI on business and education. They also collaborate with the UVA School of Data Science.

To help make sure you don't forget what you LEARNed with this guide, our best practices are available as a one page flyer for download.

Description of image

 

GenAI Coding Checklist

With the increasing use of generative AI chatbots, people are now using them to assist with programming and data analysis. However, there are some important things to keep in mind when using GenAI for programming, especially if you're still a beginner. To help make sure you end up with the correct code, follow the GenAI Coding Checklist below.

Images created with ChatGPT 4o

 

 

You often have to provide more context in your prompt than you would think. Below is an example prompt that could assist a student learning how to program in R.