Overview
This guide compiles available USC resources on ChatGPT and other generative AI tools.
About ChatGPT
What is generative artificial intelligence (AI)?
If you've ever used predictive text or autocomplete on your phone, to draft an email, or type a query into Google search, you've used a form of generative AI! But what is generative AI exactly?
In simple terms, generative AI is a subfield of artificial intelligence in which computer algorithms are used to generate outputs that resemble human-created content, be it text, images, graphics, music, computer code or otherwise.
In generative AI, algorithms are designed to learn from training data that includes examples of the desired output. By analyzing the patterns and structures within the training data, generative AI models can produce new content that shares characteristics with the original input data. In doing so, generative AI has the capacity to generate content that appears authentic and human-like.
From Generative AI Defined, TechRepublic
For more detailed information about AI and generative AI, refer to the following resources:
What is ChatGPT?
ChatGPT is a chatbot created by OpenAI and is one of the more well known generative AI tools. Creating an account with OpenAI is required, but once you're logged in, using ChatGPT is fairly straightforward: you type text into its conversational interface and it responds to you with instantly-generated text output. You may have to tweak your input—also known as your "prompt"—a few times over the course of the conversation to get the results you're aiming for.
It's important to keep in mind that ChatGPT doesn't actually understand your input or its own output. Simply, ChatGPT has been trained on a tremendous amount of online human-generated text from a specific time range. Powered by that training, it looks at what you've typed and guesses what text is most likely to come next. It's handy for brainstorming, drafting, editing, and troubleshooting, but you probably don't want to use what ChatGPT generates without first reviewing it carefully. We'll also discuss other considerations in later sections of this guide.
As USC faculty member Jonathan May puts it, ChatGPT:
takes in the current conversation, forms a probability for all of the words in its vocabulary given that conversation, and then chooses one of them as the likely next word. Then it does that again, and again, and again, until it stops.
So it doesn’t have facts, per se. It just knows what word should come next. Put another way, ChatGPT doesn’t try to write sentences that are true. But it does try to write sentences that are plausible.
From ChatGPT is great — you're just using it wrong, Jonathan May
Resources for instructors
The following resources are tailored toward instructors and provide suggestions for how to incorporate generative AI in classroom assignments:
Brightspace's Turnitin no longer provides the AI detection tool which was previously and only temporarily available in Blackboard. For more information about Turnitin:
The following resources come from a USC Libraries workshop. Note that the information is specific to the health sciences:
- Fakin' the Citation: prevent and detect ChatGPT-generated citations in writing assignments, USC Libraries workshop by Amy Chatfield
These guides can be provided to students using generative AI for assignments or research:
Resources for University employees
The following recorded webinars were hosted by USC OCISO and present useful information to keep in mind when using ChatGPT or other generative AI. To view the recordings, you must be logged into your USC Zoom account:
ChatGPT: Privacy, Intellectual Property, and You, presented by Lee Kim
- FERPA and ChatGPT in education
- ChatGPT & Intellectual Property – Changing Landscape
- AI/LLM & Governance
- AI & Workforce Development
- AI & Board of trustees
- Academic freedom – creativity, analysis, and ChatGPT
VulnerabilityGPT: Cybersecurity in the Age of AI, presented by Matthew Tossain
- Risks posed by LLM, specifically focusing on the role of AI in cyber deception operations
- How threat actors are leveraging generative AI to create more convincing malicious campaigns
- Latest trends in LLM-enabled cybercrime and the strategies required to mitigate these risks
- How AI technologies can enhance the efficiency of cybersecurity operations through innovative applications of AI in threat intelligence, vulnerability management, and incident response
Resources for learning to use ChatGPT
OpenAI offers its own guide for obtaining better results from ChatGPT:
This cheatsheet from The Neuron provides succinct and easy-to-apply guidance:
The following LinkedIn Learning video courses may be relevant:
Handling sensitive data
Be cautious about inputting information into ChatGPT, especially if you may be working with confidential data. As a general rule of thumb, don't type anything into ChatGPT that you don't intend to be publicly available:
Information submitted to these tools should be treated as being posted publicly online. Be cautious not to submit anything confidential as the information could be sold, obtained by unwanted persons, and/or used against you or USC in a malicious way. Some common examples of inadvertent confidential information disclosure might include a student submitting personal financial aid information to request assistance with financial aid application documentation, a researcher submitting a confidential data set for assistance with data analysis, or an administrator submitting confidential Personally Identifiable Information for assistance with data reporting.
From Precautions for Use of Generative AI, USC OCISO
Example of FERPA compliance
In the example below, Lee Kim, an attorney and IANS faculty member, compares a prompt that would be considered FERPA compliant (left) to one that would not (right):
From ChatGPT: Privacy, Intellectual Property, and You, an IANS webinar hosted by USC OCISO
Turning off chat history and/or model training in ChatGPT
Information in the prompts you submit to ChatGPT (or any other generative algorithm) has the potential to get incorporated into its training model, which then informs what ChatGPT outputs to others who use it in the future.
To an extent, ChatGPT permits you to opt out of this. Just be aware that:
- Settings changes apply to new conversations with ChatGPT, not existing or previous conversations
- Regardless of your settings, OpenAI can review your conversations
- This information is current as of January 2024 but could change in the future
Turn off your chat history and model training
Browser instructions:
- In the left sidebar, click your account at the bottom.
- Click Settings.
- Click Data controls.
- Click the Chat history & training toggle to turn it off.
- Repeat these steps if you use ChatGPT on multiple browsers (Safari, Firefox, Chrome, etc).
iOS instructions
Android instructions
Turn off model training but retain your chat history
Per OpenAI's Data Controls FAQ, ChatGPT Team and ChatGPT Enterprise content is not used for model training.
If you are using ChatGPT Plus or ChatGPT Free, however, you will need to go to https://privacy.openai.com/policies?modal=take-control. Select Do not train on my content and proceed through the on-screen prompts.
Using ChatGPT-generated output
ChatGPT can be helpful for generating a range of output. For instance, you can use it to:
- Brainstorm essay prompts or ideas
- Draft emails, announcements, or general copy
- Provide proofreading, alternative phrasing, and other editing suggestions
- Write code to automate repetitive tasks, or troubleshoot existing code or scripts
Keep in mind that ChatGPT-generated content may not be accurate, so you should review what it produces with a critical eye:
Information received from generative AI tools can be factually incorrect, copyrighted, or confidential to another party. Therefore, you should be cautious when using this information for academic, research or administrative purposes.
From Precautions for Use of Generative AI, USC OCISO
As Princeton faculty member Arvind Narayanan states, ChatGPT:
is trained to produce plausible text. It is very good at being persuasive, but it’s not trained to produce true statements.
From Decoding the Hype About AI, The Markup
Citing ChatGPT
The USC Libraries provides the following information on citing ChatGPT according to MLA, APA, and Chicago Style guidelines:
Sources
In addition to USC resources, the following articles were consulted for this guide:
-
Generative AI Defined: How It Works, Benefits and Dangers, Owen Hughes
-
Decoding the Hype About AI, Julia Angwin with Arvind Narayanan
-
What is ChatGPT and what can it be used for?, S. Shyam Sundar and Shomir Wilson
-
ChatGPT is great — you're just using it wrong, Jonathan May
-
Introduction to Large Language Models, Armand Ruiz
-
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Emily M. Bender, Timnit Gehbru, Angelina McMillan-Major, Margaret Mitchell
-
Language Models are Few-Shot Learners, Tom B. Brown et al (OpenAI)