Rochester Institute of Technology professor responds to new ChatGPT artificial intelligence
ROCHESTER, N.Y. (WROC) — ChatGPT is an artificial intelligence (AI) recently launched by OpenAI that can produce high quality responses to a variety of questions on topics such as sports, literature, poems, cover letters, and more.
Rochester Institute of Technology Software Engineering Department Assistant Professor Dr. Ashique KhudaBukhsh says that ChatGPT is able to produce these responses because it is built on top of a language model. More specifically, GPT 3.5, which Dr. KhudaBukhsh says is a very high-performance language model.
“Essentially, a language model is trained on a vast amount of data,” Dr. KhudaBukhsh said. “The chargeability is trained using reinforcement learning. So, the idea is, suppose it’s requested to produce five or 10 responses, and then the responses are ranked. It gets rewards if it produces a response that has a better score, and it gets a penalty if it produces a response that has a poor score. It’s the typical way like when we train animals, or even kids with rewards when they behave well.”
Dr. KhudaBukhsh also says the sources that ChatGPT gets its information from come from all over.
“It uses a language model that has been trained on a vast amount of text coming from lots of different sources. It could be Wikipedia, it could be lots of books. It has a lot of knowledge about a lot of different areas,” Dr. KhudaBukhsh said.
Dr. KhudaBukhsh says that ChatGPT is not a one and done generator. If the user wants to make edits to what ChatGPT generated, they can tell the AI to change it, and KhudaBukhsh has tested this out. He asked the AI to produce a movie synopsis on the 2001 Oscar nominated film Lagaan and took the process even further by asking for modifications after the first response.
“It produced one. Then I said, This is too long, can you make it like a three-sentence synopsis? It shot in it quite well, and retained the essence of the five or seven sentences that it wrote before? Then I said, Okay, I want a single sentence description. It also did that quite well. Then finally, I asked that, can you give a one-word description of the movie? and then it says “inspirational,” Dr. KhudaBukhsh. “I find that the responses are quite high quality. And for some certain kinds of tasks, it might be very difficult to figure it out that if it’s created by chargeability, or if it’s generated by an actual human. So, there is an inherent risk there.”
While this technology is new and exciting to try, Dr. KhudaBukhsh does recognize some of the downfalls that can come from using a system like this in places like the classroom.
“One problem could be that students start using it for writing essays,” Dr. KhudaBukhsh said. “For the homework assignments, if they’re supposed to write an essay, do research for that, and spend hours polishing the text, and instead of doing that use ChatGPT to produce one, it becomes challenging.”
The challenge that teachers and schools will now have to face is creating a more robust system to hold students accountable for turning in original work.
“In high school settings, the in-class exams will become even more important because there, the student will have to write an essay or an answer inside the classroom. So, if there is a huge gap between their homework assignments and exams, we will kind of know that maybe some wrong practices are being adopted,” Dr. KhudaBukhsh said. “I think it makes high school and even college level teaching challenging. But with every new technology, we always will have certain changes to the society, and eventually I think we will find an equilibrium.”
Outside of the classroom, Dr. KhudaBukhsh also warns of issues that ChatGPT and AI’s similar to it can pose.
“We actually had a recent work where we showed that automated captioning systems, such as closed captioning systems that just can convert audio into text. These kinds of systems, also known as ASR systems, are often trained on conversations between adults. So, their vocabulary is very different from conversation between kids,” Dr. KhudaBukhsk said. “We showed that when you have videos on YouTube Kids, and you use any automated speech recognition system on them, sometimes they produce very unsafe errors.” For example, if a kid is saying that I want to go to market and I want to buy some corn, and the ASR system will say I want to go to the market and I want to buy some [redacted]. So, these kinds of unsafe errors happen.”
Dr. KhudaBukhsh says despite some blind spots throughout the process, systems like ChatGPT and other artificial intelligence are extremely useful.
“It is extremely useful,” Dr. KhudaBukhsh said. “Technology will always have its blind spots. So, we have to be very vigilant about those kinds of errors and the new kinds of risks that it poses but, in the end, this is a big scientific progress. And as an NLP researcher and AI researcher, I’m very excited about it.”
Leave a Reply