ChatGPT: What Does It Think of Itself?
ChatGPT is an Artificial Intelligence (AI) powered chatbot launched by the company OpenAI. Although the program was only publicly released in late November 2022, it has already amassed 100 million active users last month (according to a UBS study). The capabilities of ChatGPT have been discussed widely in major publications and social media, causing controversy regarding AI’s impact on the near future. But is the tool as robust as we think?
In a previous post, I demonstrated a small portion of ChatGPT’s ability. At first, I was amazed by this tool. Not only could it provide an answer or solution to anything I asked, but it was also done so in an incredibly efficient manner.
However, after spending ample time using the tool, I was quick to notice some drawbacks. Sometimes when I asked for a piece of code, it would not run, and sometimes answers were completely wrong.
Unfortunately, I cannot summarize the cleverness as well as drawbacks of the tool in a single article effectively. However, I do want to talk about one flaw in particular that I have not seen many people talk about.
The one thing that stood out to me was how ChatGPT would deliver incorrect information whilst still sounding plausible. Let’s dive into an example.
Whilst playing around with ChatGPT, I asked the chatbot:
“Tell me some strengths of ChatGPT, and support facts with evidence.”
There was one thing I really wanted to establish with this answer: how will the chatbot support facts although it does not have access to the web?
So what did I find from the chatbot’s reply? I was provided with answers, as well as facts supported by, surprisingly, statistics. I have provided two screenshots of instances when statistics were used.
The first two lines of each screenshot strike me as very confident. Claims such as “highly coherent”, and “highly adaptable” reminded me of thinking of action words for my resume.
This reply also intrigued me since I knew that ChatGPT did not have access to the web and these claims were supported by statistics. So naturally, my next question was:
From the screenshot above, it is evident that ChatGPT simply made up these statistics to support its claims. Although I did not come across other instances of this behavior, I find it very interesting that the model is very confident in its claims that are supported by fake evidence.
Now, I could do some research and comment on how ChatGPTs model is built, but that is not the purpose of this article. I wanted to write about the front-end operation and how the program presents outputs to the user. One question that I still ponder is “where did this confidence come from?”
Straight away I thought about the Dunning–Kruger effect. This is quite ironic since ChatGPT at its core is a neural network (which tries to simulate the brain).
So does this mean that ChatGPT answers can’t be trusted? Well not exactly. My intention in this article is not to dive into the backend of how ChatGPT works. However, I did find in this article, that making up unexisting or wrong facts is a common manifestation of flaws in large language models. It is simply the state of the technology as of now. Some other flaws are:
- Not explicitly following the user’s instructions
- Generating biased or toxic output
My takeaway from using ChatGPT is that yes, it can provide solutions to problems. But by manipulating the way you ask for something, you will get a better or worse answer. That being said, it may be necessary to cross-check or validate the responses you receive.
Moving forward, I believe that ChatGPT is a great automation tool than a chat tool (maybe something to talk about in the future). It will be very interesting to see how this tool improves, and what other chatbots will become available to the public.
Thanks for reading!
If you want unlimited access to articles on Medium, click here to become a member. If reading the occasional article is good for now, make sure to subscribe here to receive my articles in your inbox!