In this post, I will address the pressing challenges posed by ChatGPT to our current post-secondary education systems in universities and colleges. Additionally, I will explore the far-reaching impacts that ChatGPT implies for the future of online learning and how we will engage in it.

I firmly believe that ChatGPT, as a groundbreaking IT product, will eventually become a daily necessity for billions of users and transform various aspects of our lives, including how we live, work, learn, and teach. In this follow-up blog post (after my last post - What is ChatGPT), I will first discuss the immediate impacts of ChatGPT on current teaching methods in universities and colleges. Following that, I will explore other effects that ChatGPT will soon have on how we learn in our daily lives.

As a professor with over 20 years of experience working in a university, I would like to use my personal experience as an early user of ChatGPT to explain the imminent and real threats that we as teachers in universities and colleges are facing. Shortly after ChatGPT’s launch in December 2022, I put it to the test by submitting numerous questions that I had used in my computer science courses, including programming, operating systems, machine learning, and advanced linear algebra. I specifically selected challenging questions that I had personally created from scratch and had never released anywhere on the internet, making it unlikely that ChatGPT had encountered them during its training stage. To my surprise, ChatGPT was able to answer all of these questions, sometimes with ‘B/B+’ level responses, but often with ‘A+’ level responses within just three attempts. What truly astounded me was that ChatGPT not only produced correct answers but also provided clear and logical explanations. In many cases, the answers generated by ChatGPT were of higher quality than those submitted by my own students throughout my 20-year teaching career. Based on my experiences, I am confident that ChatGPT can accurately answer the majority, if not all, of the exam and assignment questions that are typically given to students in most university and college courses in mathematics, natural science, computer science, and engineering.

The capabilities of ChatGPT extend beyond answering questions in technical courses. It can also generate reports in business courses and essays in liberal arts or social science courses. The text documents that ChatGPT produces are fluent and logical and often indistinguishable from those written by humans, even professionals. While the raw reports and essays generated by ChatGPT without editing by humans may not be outstanding, they should be sufficient to receive marks of at least a ‘B/B+’ from most instructors. However, due to the probabilistic nature of ChatGPT’s text generation mechanism, it can produce vastly different answers in different trials for the same query. Sometimes, these answers generated under the same prompt may differ semantically in meaning, while other times they are just different wordings to rephrase the same meaning. All of these answers are equally fluent and acceptable to a human reader. Additionally, all queries can be slightly modified to force ChatGPT to produce even more diverse or specific outputs. All these present a significant challenge for detecting academic plagiarism in open-book assignments and projects in schools. This issue is widespread in almost all university or college courses and affects almost all academic programs, including mathematics, science, engineering, social science, liberal arts, business, laws, and medicine.

There are ongoing debates on social media about how to address the possibility of academic plagiarism caused by ChatGPT in universities and colleges. Some individuals suggest revising all assignments and projects to make them substantially more challenging than what ChatGPT is capable of handling. However, I am skeptical about the feasibility of this approach. We should not underestimate the level of intelligence and versatility demonstrated by ChatGPT at present. It is highly unlikely that anyone can come up with questions that are both reasonable for students and difficult for ChatGPT to answer. Even if such questions were discovered, they would eventually become part of ChatGPT’s training data once students submit them to query the system. As a result, it is highly likely that ChatGPT would be able to correctly answer such questions in future versions. I have already observed that ChatGPT has been constantly evolving and making steady progress in the past two months. In December 2022, I asked ChatGPT to prove the statement that “every even natural number greater than 2 is the sum of two prime numbers.” It generated a long paragraph filled with mathematical notations and equations, giving the appearance of a mathematical proof. I did not bother to read it closely as I knew it could not be a legitimate proof. However, when I asked ChatGPT the same question about one or two months later, it did not attempt to create a fake proof. Instead, it informed me that the statement is known as Goldbach’s conjecture, which is one of the oldest and most well-known unsolved problems in number theory.

There is another idea being floated in response to the issue of potential academic plagiarism caused by ChatGPT, which suggests that we stop counting any open-book works by students and instead rely solely on closed exams for course evaluation. However, I believe that this approach is not the best solution for universities, given the advanced subjects and topics being taught. The range and complexity of questions that can be asked of students in an exam lasting only a few hours are limited. This type of evaluation could lead to a focus on rote memorization rather than meaningful learning. Therefore, we need to find better ways to address the issue of plagiarism while still ensuring that students are challenged and engaged in their learning.

There is a proposal to welcome the use of ChatGPT instead of restricting it in universities. The idea is to teach students how to use ChatGPT to their advantage in completing assignments and projects. However, the skills required to compose good questions for ChatGPT to answer are different from the skills required to understand and write an answer independently. It is comparable to the difference between using a calculator and being able to perform mental arithmetic. While teaching how to use ChatGPT could be a useful short course, it does not replace the value of a four-year higher education. Furthermore, there is a risk in assuming that ChatGPT can replace the need for learning and knowledge acquisition. While it is true that ChatGPT possesses a vast array of knowledge that far surpasses that of most individuals, and can provide better answers to many questions than the majority of people, this does not mean that future generations do not need to know anything as long as they know how to ask ChatGPT for answers. This perspective is concerning, and it is important to recognize the limitations of relying solely on technology for education.

Regardless of our stance on the issue, it is undeniable that the proliferation of technologies such as ChatGPT in recent years poses significant threats to our existing education systems. As educators, it is imperative that we acknowledge these challenges and collaborate to transform our traditional approaches to education, in order to keep pace with the ever-evolving landscape of technology.

The internet has become a ubiquitous learning tool, not just within schools. Undoubtedly, Google search plays an indispensable role in online learning by helping us quickly locate relevant web pages with a simple query consisting of keywords or phrases. However, ChatGPT offers a better user experience in several ways. Firstly, we can ask ChatGPT using natural language, making it appear much smarter than Google and providing a more natural way for people to communicate with computers, especially for those who are not tech-savvy. Secondly, ChatGPT can provide a coherent summary written in natural language for any query, saving us the time and effort of reading through multiple long web pages. Impressively, ChatGPT can compose fluent and coherent text in natural language that is indistinguishable from human writing. Unlike Google, which simply excerpts a paragraph from one relevant page, ChatGPT can use its own words to address our questions directly. For complex questions, ChatGPT provides much better experiences than Google’s excerpts. While ChatGPT may occasionally provide incorrect answers, it can already answer most factual questions from any discipline with accuracy not inferior to that of any human teacher. Of course, there is still significant room for improvement before ChatGPT can answer all questions perfectly on par with dictionaries and textbooks. I am confident that these issues will be gradually resolved in future generations of ChatGPT. At present, ChatGPT poses a real threat to Google’s search engine and business model, and I am curious to see how this battle will unfold in the coming years. Regardless, it will significantly change the way we search for information and learn new knowledge in the future.

Over the course of the last few decades, the internet has become home to an extensive and diverse range of high-quality content that is available for free to anyone. Most of this content takes the form of text, which provides the fuel for ChatGPT and similar programs. These text documents have largely been written by humans and have gradually accumulated over time. It takes days or even weeks for a journalist to write an article, and several minutes for a user to write a product review. However, ChatGPT can produce an article of the same length in under a second, and generate thousands or millions of reviews in mere seconds. To most readers, text generated by ChatGPT is indistinguishable from content created by humans. As ChatGPT continues to improve, it may become increasingly difficult to discern which text is human-generated and which is machine-generated. It is reasonable to expect that more computer-generated text, as well as computer programs/codes, will be posted to the internet in the future. With the ease and speed of text generation by ChatGPT, it is likely that this portion of text content will rapidly expand, quickly overtaking human-written content to become the dominant force in online content. Consequently, we will inevitably begin to read more machine-generated content when we surf the internet. It is difficult to predict how this will affect our learning in the long term. Additionally, we can envision future chatbots learning from the vast amounts of “fake” text, largely produced by ChatGPT, which may lead to a self-perpetuating cycle of learning. It remains to be seen whether this self-sustaining feedback loop will result in a continuous enhancement of chatbots’ intelligence level. In any case, we should prepare ourselves to read more text generated by computers like ChatGPT that is customized for individual readers based on their preferences, backgrounds, and biases since text generation by computers is a fast and inexpensive process.