GenAI is an umbrella term used to describe a particular type of artificial intelligence that generates output in response to input. However, it is not that different from "predictive AI" that has been embedded in popular products for a while. Netflicks and Amazon use predictive AI to tell you what shows you might like based on what you've watched or things you might want to buy based on what you've purchased previously.
GenAI tools - like Open AI's ChatGPT 4o and DAllE-3, Microsoft's Bing Chat, Anthropic's Claude, Google's Gemini, Perplexity, MidJourney, and Pi- use the patterns found in the trillions of bytes of data on which they've been trained to generate output (text, visuals, music) by predicting the next best word in the sentence, pixel in the image, or note in the music. The output from these tools is "original" (in the sense that it doesn't directly plagiarize from one source) and is dependent on the prompt it's given.
The user gives it a prompt and the tool responds. If I said to you - "It's sunny today, the sky is _____", how would you finish that sentence? You'd likely say "blue", "clear", "bright", or maybe even "beautiful", right? GenAI text generators work the same way. They take their data, look for patterns in the data, and then choose the next best word to use. Yes, it seems more impressive than that. But that's it. Want to see it in action?
The output generated by these tools is only as good as the input it receives from a human user - input also known as "the prompt". The human prompts the tool to give it an output. The more guidance and parameters provided in the prompt, the better the output.
A good prompt will help the AI focus or hone in on the relevant data in its "memory". Good prompting includes the following:
a Role that you want the machine to adopt
For example, "you are an expert in crafting learning outcomes for undergraduate biology courses."
a Context that you want the machine to stay within
"The course is an introductory course at a research university that enrolls about 400 students, 2/3 of which are freshmen biology majors and 1/3 of which are freshman or sophomores from other science majors."
a Task that you want the machine to complete
"You will craft 5 learning outcomes for this introductory course on human genetics."
a Chain of Thought for the machine to follow
"First, you will review the course syllabus that I provide you. You will not respond other than to tell me when you're done reviewing it. Second, you will identify and classify the learning outcomes found in the syllabus based on Bloom's Taxonomy of Learning. You will explain your reasoning. Third, you will rewrite those learning outcomes if you think they could be articulated more clearly. Fourth, etc...."
an Invitation to Ask Questions. This is one of the most helpful prompting tricks - inviting the tool to ask you for information that will help it complete the task.
"Before you begin, do you have any questions to help you complete this task?"
GenAI has many limitations that are relevant to education:
It only predicts: it doesn't understand, reason, think or comprehend. It appears to do those things, but certainly not as humans do.
It confabulates: this means that it makes stuff up because it can't verify that the pattern-matching prediction it made is right, and it doesn't have a semantic understanding that humans do so it just doesn't know that what it's saying isn't true. For example, have you ever read a student paper where they talk about an article, book, or author that doesn't exist or at least has never written about that topic? That's GenAI.
It uses a LOT of energy: it's predicted that the energy needed for one ChatGPT query is equivalent to powering a lightbulb for 20 minutes, and generating an image takes as much energy as fully charging your smartphone
It reinforces and amplifies human biases: these machines were trained on content created by humans, so naturally there are biases within the data. Also, they were trained largely on western content, so there's a western-bias bent. And, the history of humanity is full of biases, so the patterns within the data reflect that. For example, since the majority of doctors in the western world over time have been white men, the pattern-matching prediction machine, will see "white man" as the most likely match to the term "doctor."
Garbage in, Garbage out: one of the sources of training data for Open AI's ChatGPT was Reddit! Of course, the general internet was also a data source. So, it's no wonder why these machines say some crazy things. And now, these machines are being fed back into with their own output, so the quality of the output is likely to get worse (not better) over time.
It threatens the value of intellectual property: did you know that your students are uploading your intellectual property (eg., syllabus, lecture notes, lecture transcript) to ChatGPT? They're also uploading the assigned readings. Users can trick DALLE-3 to generate images that it doesn't own the copyright to, even though the developers tried to put guardrails around that. And, the data on which these tools were trained? Not bought and paid for. (Although now academic publishers are selling their - your - intellectual property to OpenAI).
So, before you use it, or before you ask or require your students to use it, you might want to consider these limitations and, at the very least, have a conversation with your students about them.