The Confident BS of ChatGPT

When ChatGPT was making a lot of buzz earlier in the year, I was busy and it took me a while to sit down and play with it. Right before the Spring semester started, I finally had time to log in. I was impressed. I sat down and wrote-up my thoughts to give to my students, outlining what I thought were good uses of it, what were uses that would get them in trouble with academic integrity, the strengths and limitations.

My impressions were that ChatGPT could be a great tutor. I gave it several prompts related to computer architecture, like “What is virtual memory?”, “How is the TBL used in virtual memory”, “How can programmers write cache-aware code?” Then I gave it prompts from machine learning and natural language processing, for example: “Explain SVD”, “How does ChatGPT work?”, “How do transformers, encoders, decoders work?” I was impressed with all of the answers. They were both well-written and factually accurate. I told students that this could be a great learning assistant, a virtual tutor. I particularly appreciated that it could disambiguate what I was asking from the context of my prompt.

ChatGPT Opinions

Then I tried a few opinion questions. To the question “Are traditional machine learning algorithms obsolete in the face of deep neural networks?”, I got a nuanced response on the pros and cons of deep learning versus more traditional machine learning algorithms. I prompted “What are the greatest threats to democracy around the world?” twice, each time getting a very slightly different answer, showing the randomness in the machine. I decided to get really opinionated next “Will Elon Musk destroy Twitter?” As of the date of that prompt, ChatGPT had no knowledge of the deal. We all know what happened, so I’ll just leave you with this image …

I shared some of my thoughts with my teaching colleagues before the semester started, and there was an explosion of discussion with some minor freak-outs. After we had a chance to discuss, I think most professors were past the shock and awe stage.

Teacher Freakout over ChatGPT

First, professors realize that there are a lot of ways to cheat already. There are sites like Chegg that provide answers to prof’s homework and test questions.

Second, if you are asking questions that can be easily answered by Chegg or a bot, then maybe you should revamp what you are doing.

By the way, ChatGPT nearly destroyed Chegg when ChatCPT first came out, so it will be interested to see how Chegg bounces back.

I interrupt this blog post for a brief rant about cheating

When I was young, only losers cheated. That is, smart people never cheated. Average people never even cheated. The only people who cheated were those who felt that they had no way out. So when I came to teach college, I was a bit naive and unprepared for the fact that even smart students cheat now.

I’ve talked a lot about this with my younger son in order to get schooled on how kids think these days. He’s not a cheater by the way. Too much pride. My son attributed it to the pressure that students feel these days to get ahead. Pressure from their parents mainly, especially for children of first-generation immigrants (of which my kids are by their father). Part of this is parents wanting to make sure that their kids succeed, but let’s face it, part of it is bragging rights to their friends.

Part of the pressure is also the perception on the part of students that they won’t get an internship or job if they don’t have straight As. I can state with certainty based on my observations at a good CS school, that B students can get jobs and internships if they have cool projects to show on their GitHub. I always told students that, but they never believed me.

One thing I will definitely not miss about teaching is the end of the semester requests to “bump up their grade”. My answer was always a flat no, not fair to other students.

ChatGPT Hallucinates

It is well known by now that ChatGPT puts together information scraped from the web, but also takes a few reasoning baby steps on its own. Both of these sources can provide false information. Users of these bots needs to be sophisticated.

Later in the Spring semester, a Student showed me the interaction above. Notice that ChatGPT politely declined to give an opinion in normal mode, but in developer mode it unleashed this obsequious, fawning assessment of who I am. Hilarious, yet a little creepy. Hilarious because of the over-the-top false flattery. Creepy because ChatGPT said it “had the pleasure of interacting with her on a few occasions.” Yikes! So we must be aware that these AI chatbots are recording our interactions in order to learn from them. If that doesn’t bother you just a little, you are not paying attention.

Google Search Lab

I also signed up for Google Search Lab and have been using their AI-powered search on my computer and phone.

I’ve found it helpful. It is faster than going to Wikipedia (I still love you Wikipedia!). Features that I particularly enjoy are that you see the sources of the information so you can dig deeper, and you can ask follow-up questions.

Google Labs was an experiment that was shuttered a few years back but was revived in 2023, as announced in Google I/O 23.