By MARTY LEVINE
“Imagine a highly skilled creator who makes new things based on patterns it has already been exposed to,” said J.D. Wright. “It isn’t copying. But it isn’t thinking either. It predicts what should come next based on the patterns” — patterns it has discerned from its exposure to much of human creation.
That’s artificial intelligence, or AI, Wright said, as he began the latest addition to Faculty & Staff Development Program courses: Artificial Intelligence (AI) 101. “AI gives the illusion of thinking,” he said.
Wright, a consultant in the Center for Teaching and Learning, said the technology uses Large Language Learning — the process of recognizing patterns “in vast amounts of text data, enabling it to predict and arrange appropriate words and sentences in a human-like manner.
“It can’t really understand the words we use,” he added — or the words it uses — and yet it can produce basic business, research and other technical documents, as well as stories, jokes and songs, of a kind. It has evolved quickly since its inception in 2021 and will evolve even more quickly in the future. “You’re getting in on the ground floor,” Wright said.
He asked participants to head to openai.com, home of ChatGPT, perhaps the most well-known AI offering. He demonstrated how easy it is to enter a prompt or query, in ordinary, conversational language, asking ChatGPT to write (for instance) an announcement or social media post, and see it produce a result almost immediately — but not usually a satisfactory, finished product.
That forces you to add more prompts to refine the tone and audience of the piece, and requires you to read the results for things AI cannot see: errors and nonsense.
“It’s conversational; it’s very informal,” Wright said of AI. “It can understand.” And then he corrected himself: “It can process, is a better way of putting it. You are training it.”
This revising process is called “iterating” — creating new and more refined drafts.
As you revise your initial ChatGPT query — over hours or even months — the program remembers this conversation thread. You also can hold many conversations at once with it. Refining your initial prompt to get the best results “comes with practice,” Wright noted.
Class participants were amazed at the results of their single-sentence queries, and also somewhat skeptical of AI’s choices in spots.
One woman in the class had asked ChatGPT to create a paragraph to advertise a workshop, which resulted in an “excellent” product, she said: “It pretty much matches the flyer we did on our own.”
“With much less effort,” Wright added.
“It did very well with what little I gave it,” the woman said.
A second woman in the class asked ChatGPT to complete a similar task. “It is filling in information that I didn’t provide it,” she said, noting that it had added biographical material on the workshop presenter. “It’s giving all the right words and right things that I didn’t include. I’m just blown away with how it is able to regurgitate information it has apparently taken in from elsewhere.” Then she added, with hesitation in her voice: “None of it is untrue.”
“It’s very fun to use it, even though you know it is not thinking — there are some dangers to that,” Wright responded. To say “none of it is untrue” would suggest that the information ChatGPT chose to include doesn’t entirely match the human vision for this particular project.
The woman agreed: “I would definitely tweak mine,” she said of ChatGPT’s first try. “It helped immensely by cutting down a lot of thinking I’d have to do.” However, “it wasn’t my final version. … It gave me some words I wasn’t thinking,” but she felt they were just “fluff.”
“It’s really good at giving you a rough draft or a first draft that you can play around with,” Wright concluded.
A third participant, having asked ChatGPT for a humorous iteration of her memo, “Come learn about volunteering,” found the result to be “cute.”
“It’s witty. It’s clever. It’s not laugh-out-loud humor,” Wright allowed. But the tone was certainly “very inviting.”
Another participant was displeased when she asked for a visual rather than textual representation of her request: “Some of the information was incorrect,” she reported. “It doesn’t understand that the day and time are different things. Where and when isn’t exactly right.”
“Images are kind of touch and go” for AI in general, Wright said. The best image requests specify whether AI should use symbols or objects and how busy or simple the representation should be, or how colorful. Chat GPT “likes to go over the top,” he said. ChatGPT and other AI programs “don’t understand text as image,” sometimes substituting mere squiggles or misspelled verbiage for intelligible text. “But it is getting better all the time,” he said.
“With a well-designed, iterated prompt,” Wright said, AI can even summarize or critique texts, generate and correct computer code, create and play games or analyze data.
He has tried the latter task and received “charts, graphs, language and explanations” to use. “It was really stunning.” He has also used one AI service to create another, designing a chat bot to serve students in a particular class.
Wright was careful to end on a note of caution: Never give AI sensitive, confidential, private or proprietary information; beware of its well-known propensity to please by making up answers when the right one cannot be found (AI’s so-called “hallucinations”); and always be prepared to edit any AI product.
Think of it as a kind of co-intelligence or collaborator, he said. “Play around with this technology. It doesn’t come with a user manual. We’re still figuring this out. It should become more useful to you as you use it.”
Marty Levine is a staff writer for the University Times. Reach him at martyl@pitt.edu or 412-758-4859.
Have a story idea or news to share? Share it with the University Times.