ChatGPT: Opening Your "I"(ntelligence) to What's True and Truly Artificial
This opinion piece was written by Professor James Geller - Chair of the Department of Data Science
The ChatGPT program has been a topic of discussion in every single professional meeting that I have recently attended, and in most cases without being on the agenda. The professional mailing lists of HS teachers and Medical Informatics research leaders that I am on are abuzz with messages about it. Reports about catastrophic failures and amazing successes are balanced by blank fear. Will ChatGPT write the essays and solve the homework problems of our students? How close is ChatGPT to passing the medical licensing exam? Will ChatGPT put all of us out of our jobs? Google is so worried that ChatGPT will replace its search engine that they are bringing the founders Page & Brin back. I started in Artificial Intelligence research in 1983 and finished a dissertation on Knowledge Representation in 1988, and I dare say that ChatGPT is the biggest breakthrough that I have seen since I graduated. I even venture to go on record that it introduces a technology that will be as influential as the first Web Browser. Yet it has dangerous flaws.
Probably the best way to characterize ChatGPT is as a person with a phenomenal memory who has read and memorized and even organized thousands of books, yet … does not really understand anything s/he has read. Whatever topic comes up, s/he can find relevant phrases and sentences from memory and form them into impressively beautiful prose, or computer programs, or poems, etc. that look and sound convincing, are often correct, and address the questions or requests received. Yet many times the results of ChatGPT can be spectacularly wrong without giving the slightest indication that would raise the user’s suspicions. I have asked it to solve an introductory programming homework question and it gave me a correct answer, a slightly wrong answer, and a completely wrong answer for the same problem. In some cases when it does not know an answer it seems to make up an impressive sounding, yet completely untrue, response.
Google does not need to worry (yet), ChatGPT does not do any search and acknowledges that all its knowledge is from 2021 and earlier. Any uses of the word “search” in the context of ChatGPT are misleading. However,…
Alan Turing, the father of modern computing (which he used for cracking German encrypted messages during the Second World War), and the father of the idea that an Artificial Intelligence is possible, formulated a test that would determine whether a “hidden conversation partner” is intelligent or not, which is aptly called the Turing Test. In simple words, if the conversation partner can trick a human into believing that s/he is also a human, then s/he is intelligent. (To learn about Alan Turing in an entertaining way, I recommend the movie “The Imitation Game.”)
For any superficial observer who does not check the correctness of all responses, ChatGPT gets closer to passing the Turing Test than any other program in history. It smoothly avoids questions about personal preferences and does admit that it is “only a language model” when it gets a difficult question. It has a limited context model when the user refers back to previous questions, yet it compares to the annoying chatbot on your banking home page as an Abrams Tank would compare to a .45 caliber handgun.
Microsoft will invest millions more into OpenAI, the maker of ChatGPT. They will be able to add true search functionality and to improve the context model, making it even better at pretending to be intelligent. Yet, it is not intelligent in a human sense, and it is an open question whether any of its successors will ever reach this goal, a question that has been hotly debated by AI researchers and philosophers for decades. In my view, it is a fantastic new tool, just as the first Web Browser was a fantastic new tool, but if you ask ChatGPT an important question, make sure to check the answer for correctness before acting on it, or believing it. Now go play with ChatGPT.
-- James Geller, Professor