Navigating Generative AI in Higher Education

Generative AI is a type of machine learning, called "deep learning", that is trained on massive datasets to create new content based on a user's prompt or instructions. Deep learning models are able to recognize patterns and relationships in datasets, which mimic human behaviors. This allows them to create new content similar to what one might expect a human to produce. But it is important to remember that these models are not as sophisticated as a human brain. While they use neural networks to create new content based on past relationships in data, they cannot create new content that is not already included somewhere in its training data. For example, ChatGPT was trained on datasets of content available before April 2023. If you ask ChatGPT to write a poem about the gymnastics winners in the 2024 Summer Olympics, it will not be able to perform the task because it does not know who won in gymnastics at the 2024 Summer Olympics. 

The most common type of deep learning models are "large language models" which are used for text-based generative AI tools, such as ChatGPT. But there are many different types of deep learning models, including those trained to create images, music, and video. 

There are many Generative AI tools available, including those that create images, audio, help organize files, summarize information, and more. Below are some of the more popular text generation tools.

It is important to be aware of the potential for hallucinations and social bias in all content provided by these tools. Hallucinations occur when a generative AI tool provides an answer that is false or made-up. Users should always double-check that the information generated by AI can be verified through other reliable sources. 

ChatGPT

ChatGPT is one of the most talked about generative AI tools available. The current model, ChatGPT 4o, was trained using an enormous dataset made up of text available on the internet. While it is unclear what exactly is in the dataset, OpenAI, the developer of ChatGPT, has stated that the dataset only consists of information through September 2021. 

Microsoft CoPilot

Microsoft Copilot is another OpenAI generated tool (Microsoft is an investor of OpenAI) and uses the same GPT-4 model as ChatGPT. Because it also contains an algorithmic search engine, the information acquired through Copilot is likely to be more accurate than ChatGPT. It also provides sources for some searches, which allows users to see where some of the data came from. As with ChatGPT, the actual dataset used to train this tool is unavailable to the public. You may have heard that sometimes Copilot can generate "weird" or unsettling content. Safeguards have been put in place to limit this type of content generation. 

Perplexity

Perplexity is an AI search engine that also utilizes GPT-4 and is connected to the internet. Users can utilize the "pro" feature to see more detail on how information is gathered and, occasionally, the tool will ask clarifying questions to ensure the generated content is relevant. Perplexity also includes sources and uses a more advanced algorithm for searches to try to provide more "trustworthy" results. 

Google Gemini

Google Gemini is the only tool listed here that requires an account log-in. This tool uses a different type of training model, called the Language Model for Dialogue Applications (LaMDA), to generate responses to user prompts. According to Google, "[t]he language model learns by “reading” trillions of words that help it pick up on patterns that make up human language so it’s good at predicting what might be reasonable responses." The difference in this model from GPT-4 is that it is meant to be more "conversational." Though also tied to the traditional Google search engine, the results tend to be less relevant and reliable than other tools.