Author Introduction:
Yu Mingfeng, male, associate professor in the Department of Philosophy at Tongji University, PhD.
Research Projects:
Shanghai Philosophy and Social Science Planning General Project “Research on Nietzsche and Modern German Aesthetic Education Thought” (2020BZX008); National Social Science Foundation Major Project “Research on Ethical Risk Prevention of Artificial Intelligence” (20&ZD041).
Abstract
The breakthrough of ChatGPT in natural language processing has opened the door to the generalization of artificial intelligence. The current breakthroughs are not based on the rule-based early artificial intelligence but on a new type of artificial intelligence based on machine learning. Machine learning is the aggregation of computing power, data, and algorithms, and thus the civilizational interface of the machine learning era has transformed from text recognizable by humans to data recognizable by machines. Currently, discussions about ChatGPT and educational artificial intelligence often focus on the short-term effects of machine learning. It is necessary for us to understand the challenges posed by ChatGPT from the perspective of the machine learning era, using this as an opportunity to examine the restructuring of human learning. While machine learning brings various challenges, it may also liberate educational institutions from entanglement with survival institutions, returning to the civilized concept of “learning to become an adult”.
The so-called philosophy of education has the primary task of reflecting on what education and learning are. This reflection is not achieved overnight, as the specific meanings of education and learning have their own temporality and are bound to change over time. When Zhu Xi annotated the “Analects”, the real meaning of “learning and practicing it at the right time” was already vastly different from Confucius’s time, because Zhu Xi’s era had already seen the advent of printing technology and the imperial examination system. At that time, the main focus of learning was no longer the practice of rites but the memorization and annotation of classics. (Although what Zhu Xi emphasized was the awakening of the mind, his “Commentary on the Four Books” still became the standard textbook for hundreds of years. Regarding “learning and practicing it at the right time”, Zhu Xi’s explanation is less about tracing the original meaning of Confucius’s time and more about confirming the principles of the study of Li: “Human nature is inherently good, but the awakening occurs in stages; those who awaken later must emulate what the earlier awakened have done to clarify goodness and return to their original state.” [See Zhu Xi: “Commentary on the Four Books”, Zhonghua Book Company, 2011 edition, page 49.]) Entering the 21st century, its meaning is even more different. The main task of learning is no longer to delve into the Four Books and Five Classics but to integrate into the modern knowledge system and become citizens of modern states. Therefore, the reflection of educational philosophy must inevitably include a historical philosophical dimension, redefining education and learning in the context of dramatic changes in the times, with the primary task being a redefinition of the times themselves. (However, does education and learning not have an essence that transcends the times? Shouldn’t philosophy focus on the essence in this sense? Such a dimension of essence transcending change does indeed exist in our cognitive structure. Yet even if we assume it exists in reality, its specific meaning must be understood in a historical context. In this regard, starting from historical philosophy remains a practical path. From the reflection of historical philosophy, we may approach this essence more effectively than through systematic approaches. Nietzsche adopted a more radical historical approach, which directly negated the dimension of essence. Whether we agree with Nietzsche’s approach or not, and whether his approach inherently contains paradoxes, his critique of the inherent non-historical tendency of philosophy is thought-provoking: “All the statements of philosophers about man are merely assessments of man in a very limited period. A lack of historical awareness is the hereditary defect of all philosophers.” [See Nietzsche: “Human, All Too Human”, translated by Wei Yuqing, East China Normal University Press, 2008 edition, page 19.])
“Changing with the times” in this sense does not mean going with the flow, but rather seeking a clear position amidst the accelerated flow and restless currents of the era. The perspective and sense of direction provided by historical philosophy can make us sensitive to new phenomena like AlphaGo and ChatGPT on one hand while preventing us from falling into the panic created by media exaggeration on the other. Thinkers in this sense actively intervene while remaining calm, focusing on understanding the deep, quiet currents beneath the surface of the tumultuous river.
ChatGPT is such a current in 2023. The breakthroughs in artificial intelligence in natural language processing are significant. In discussions about educational issues, there are two points worth emphasizing.
First, natural language processing (NLP) is still a specialized field of artificial intelligence, yet natural language is the basic medium of human civilization. It not only carries human knowledge and thoughts but is also the primary channel for interaction between humans and the world, and between people. The breakthrough in natural language processing means that the door to the “generalization” of artificial intelligence has been opened. This generalization refers, on one hand, to the fact that GPT-4 has exhibited certain characteristics of general artificial intelligence. Microsoft Research published a lengthy paper titled “Sparks of Artificial General Intelligence: Early Experiments with GPT-4”, concluding that “Given the breadth and depth of GPT-4’s capabilities, we believe it should be reasonably regarded as an early (but still incomplete) version of a general artificial intelligence system.” General artificial intelligence (AGI) is not limited to specialized areas like chess, driving, or painting, but is “capable of performing various tasks like humans” and has cross-domain problem-solving abilities.[1] Microsoft is a major funding source for OpenAI, so Microsoft Research is not a neutral third party. The astonishing conclusion of this lengthy paper will certainly face skepticism and challenges from peers; however, based purely on performance, GPT-4, which excels in chatting, painting, programming, and other fields, can indeed be considered “the spark of general artificial intelligence”. Regarding this, distinguishing between the concepts of general artificial intelligence and strong artificial intelligence may resolve some debates. Strong artificial intelligence possesses “real mental capabilities”, while general artificial intelligence may not; it merely gives the impression of possessing “real mental capabilities” based on its performance. (For a discussion on this distinction, see Xu Yingjin: “Fifteen Lectures on the Philosophy of Artificial Intelligence”, Peking University Press, 2021 edition, pages 92-93. Of course, we can further question: if there is no real mental capability behind it, to what extent can general artificial intelligence be generalized? Will it remain only in a spark state?) In this regard, we must be cautious not to overestimate large language models like ChatGPT. It is not truly “chatting” because there is no “real mental capability” behind it; it does not understand the sentences it outputs but is merely performing impressively accurate frequency statistics of words. Therefore, aside from the debates about general artificial intelligence, this so-called “generalization” also primarily refers to the fact that breakthroughs in natural language processing have further removed barriers to human-machine communication, which means that artificial intelligence has fully infiltrated the world of life. The era of artificial intelligence has truly arrived. In this regard, ChatGPT can indeed be regarded as a milestone in the history of artificial intelligence development, and its significance cannot be underestimated.
Secondly, the impending era of artificial intelligence is, more accurately, an era of machine learning. The current breakthroughs in artificial intelligence are not based on rule-based early artificial intelligence but on a new type of artificial intelligence based on machine learning. The so-called early artificial intelligence based on rules is the “traditional way of making computers perform a task”, that is, “writing down algorithms” or “a series of instructions sent to the computer” to directly stipulate how the computer should perform tasks.[2] In contrast, machine learning “involves a large amount of data input to predict entirely new results, rather than direct commands yielding direct outputs”.[3] In other words, “Every algorithm has inputs and outputs: data is input into the computer, the algorithm processes it, and then outputs the result. Machine learning reverses this situation: it inputs data and expected results, and outputs the algorithm that transforms the former into the latter”.[4] The algorithms used in machine learning are “algorithms that generate other algorithms”; it is the automation of automation itself.
There are three types of methods in machine learning: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning, according to the definition by Stuart J. Russell and Peter Norvig, is to obtain a function mapping from input to output by “observing some examples of input and output”.[5] Therefore, supervised learning deals with labeled data, while unsupervised learning identifies patterns from vast inputs without labeled data or explicit feedback. Reinforcement learning allows machines to “learn from a series of reinforcements, i.e., rewards and punishments”.
The rapid development of machine learning in the last decade is especially due to “deep learning” based on “artificial neural networks”. The so-called “depth” refers to the hidden intermediate layers set between the input layer and output layer. As some argue: “The hidden layers are key to the capabilities of artificial neural networks, but they also pose a problem. It is difficult to understand how artificial neural networks find solutions.”[6] This is also the “emergence” and “explainability” issues that have attracted significant attention and controversy in discussions about ChatGPT and GPT-4. Regardless of whether it is called “emergence”, machines are indeed learning. Strictly speaking, machine learning is different from human learning; as long as strong artificial intelligence has not yet emerged, we can only discuss “machine learning” in a metaphorical sense. However, after layers of training, machines have indeed generated a fairly reliable algorithm from vast data; this process can be referred to as “learning”, although its learning path still exhibits a black box nature.
We do not need to delve into the relevant technical details here, but rather be content with outlining the technical profile and understanding several basic characteristics of the current era from this outline, explaining why we advocate naming this era as the “era of machine learning”.
First, the widespread application of machine learning can be seen as a significant transformation in the technological era since the mid-19th century. (For a broad and narrow definition of the technological era, see Yu Mingfeng: “Reduction and Infinity”, Shanghai Sanlian Bookstore, 2022 edition, pages 60 and 92-93.) Currently, terms like “AI era”, “intelligent era”, “data era”, and “algorithm era” are popular, but if we want to more accurately identify this development, we should call it the “era of machine learning”. This is because machine learning is the current form of artificial intelligence and the key to why data has become so important and algorithms can penetrate the world of life. Computing power, data, and algorithms are all indispensable; emphasizing only one of the three cannot capture the essence of the success of machine learning: “The reason AI has been reborn and has grown exponentially in the past decade is due to significant advancements in the field of machine learning (as we have noted, this is based on faster computer processors, vast available data, and new computing methods).” (See Wayne Holmes, Maya Biliak, Charles Fadel: “Artificial Intelligence in Education”, translated by Feng Jianchao et al., East China Normal University Press, 2021 edition, page 94. For discussion on the importance of data, see Susskind’s remarks: “Data is crucial for machine learning; if there is too little data, the development of machine learning algorithms will be hindered. If there is enough data, ‘a learning program with only a few hundred lines can easily produce millions of lines of programs and can be reused across different problems’. This is why data is called ‘the new coal mine’, and those who collect data are called ‘data miners’.” [Jamie Susskind: “The Power of Algorithms”, Beijing Daily Press, 2022 edition, page 10.] However, data is important because it has corresponding computing power and algorithms; conversely, the opposite is also true. Therefore, emphasizing only one of the three is one-sided.) Machine learning can be seen as the aggregation of computing power, data, and algorithms. As historians of technology have said, to judge the direction of technological development, one should not look at one-sided breakthroughs but at “the gathering momentum of a wave”.[7] Machine learning is precisely the gathering momentum of current information technology.
Secondly, from the perspective of machine learning, we find that the forms of knowledge and the interfaces of information exchange are transforming from text recognizable by humans to data recognizable by machines. On one hand, the past knowledge base of humanity has been digitized, and on the other hand, the sources of data have expanded from human thoughts to everything that the internet can record and sensors can monitor, including seemingly meaningless human behaviors, bodily organ pulsations, and animal behaviors, etc. Smartphones, smart driving, and various sensors implanted in humans and animals are collecting such data every day. Machines thus bypass conscious subjects, summarizing patterns and predicting behaviors through data. This means that the interface of current civilization has significantly shifted to data that machines can recognize. It is estimated that around the year 2000, only one-fourth of information existed in data form; by 2013, this proportion had exceeded 98%. (See Kenneth Cukier & Viktor Mayer-Schönberger [2013]. The Rise of Big Data. Foreign Affairs, May/June 2013 https://www.foreignaffairs.com/articles/2013-04-03/rise-big-data. Susskind quotes this statement in his book, but the Chinese version mistakenly translates the “one-fourth” of 2000 as “one-third”. Regarding this process, Susskind provides substantial explanations: “Four factors contributed to this process. First, an increasing number of social activities are conducted through digital systems or platforms, thus more data can be collected. Second, over the past fifty years, the cost of storing data has halved every two years, while its density has increased by 50 million times. Third, the explosive growth of computing power has enabled us to process the stored content. Fourth, the digital information can be replicated almost at no marginal cost, allowing it to be reproduced millions of times at a relatively low cost. These factors together explain why such a massive data explosion occurred during the transition from print-based information systems to digital systems.” See “The Power of Algorithms”, page 34.) It is conceivable that in the past decade, data has nearly covered all information. This transformation can be regarded as the most important information revolution since the invention of printing, marking a shift in the interface of civilization. Today, it is still difficult to make a sufficiently thorough assessment of this, as we are still in the early stages of this transformation; however, from the perspective of machine learning, we can at least open a more appropriate view to understand this transformation.
Thirdly, naming this era as “machine learning” highlights the unique knowledge conditions and subjectivity situations of this era, which is especially worth paying special attention to when discussing educational issues. Throughout the long history of civilization, humans have been the only subjects of knowledge production. However, in the era of machine learning, not only has the primary form of knowledge or information exchange shifted from text to data, but the sources of data have expanded to various sensors, and the processing of data and generation of knowledge will also bypass the human brain, occurring within the black box of machine learning. Do machines not thus become cognitive subjects beyond humans? With the arrival of the machine learning era, humans are externalizing their subjectivity at an accelerating pace that cannot be underestimated. If we continue to adhere to traditional subject theory models, we will be unable to grasp the real situation of humanity in the current world. “This world is no longer human-centered.”[8] Humans are no longer the only cognitive subjects or action subjects. Although artificial intelligence does not yet possess self-awareness and is unlikely to “emerge” self-awareness through advancements in computing power and algorithms (as this would mean an ontological leap), it must be stated that artificial intelligence has already acquired a quasi-subjectivity status.
Moreover, machine learning has indeed diminished the subjectivity of individual humans: “When we infiltrate, wear, or carry sensors, or when sensors are implanted in our bodies, what changes occur in the concept of our bodily experience when we output our physiological and behavioral historical data to the sensors?”[9] If machine learning can bypass human self-awareness through data processing to derive more reliable knowledge about humans and make more reliable predictions, then are we not experiencing a transformation of ourselves into objects? Regardless, we urgently need an expanded theory of subjectivity today, incorporating the subjectivity forces beyond humanity into our examination. The concept of machine learning is an important component of such a theory of subjectivity.
The emphasis on the concept of the “machine learning era” is, on one hand, to more accurately define this era, especially its knowledge conditions and living situations. On the other hand, it is to propose the concept of human learning accordingly. Future humans will no longer be able to understand themselves without machines, and future human learning will also be impossible to understand without machine learning. This future has even already arrived.
ChatGPT first drew global attention due to the strong reactions in the educational field. In early 2023, reports indicated that a student at Northern Michigan University received the first prize for a paper written by ChatGPT, prompting the educational field to sound the first alarm of “the wolf is coming”. Why education first? Because the application of ChatGPT in various aspects of social life often requires further integration, but its application in education is the most direct. It can browse vast literature at lightning speed and write papers in a relatively organized form. Although various issues can still be found in practical use, from the reactions of North American universities and top journals, we can already sense the weight of this impact. However, most discussions on this topic merely focus on the short-term impacts brought by ChatGPT, especially issues of evaluation in higher education, intellectual property in academic production, and plagiarism. As some argue: “The most controversial issue regarding ChatGPT is not the quality of its responses but whether it will become a tool for those who need to write texts without putting in effort.”[10]
On the other hand, discussions more practically focus on “how to successfully integrate artificial intelligence into educational contexts to benefit teachers and students while promoting responsible and ethical use”.[11] Such discussions generally fall within the scope of educational artificial intelligence (AIEd). Similar discussions have a long history; the book “Artificial Intelligence in Education” even traces its history back to the 1920s with psychologist Sidney Pressey and behaviorism’s father B.F. Skinner (see Wayne Holmes, Maya Biliak, Charles Fadel: “Artificial Intelligence in Education”, translated by Feng Jianchao et al., East China Normal University Press, 2021 edition, pages 99-100. Of course, this can only be considered as the prehistory of educational artificial intelligence, as artificial intelligence did not yet exist at that time), while the main current application is the so-called Intelligent Tutoring Systems (ITS). It can be imagined that if ChatGPT is appropriately utilized, it would be the most effective intelligent tutoring system to date; further developments in machine learning will also lead to deeper integration of artificial intelligence into teaching.
All these discussions, like those on the other side, are necessary. However, related discussions often merely regard artificial intelligence as a usable tool, limiting the exploration to how to apply artificial intelligence in education and develop teaching aids. This only focuses on the short-term effects of machine learning and has not yet reflected on educational philosophy from the perspective of such a grand judgment as the “machine learning era”.
In such a time of great change, we must reflect on human learning in the machine learning era with an educational philosophical perspective, beyond short-term considerations. Reflections on this are large-scale thoughts centered on the future of humanity, possessing a nature of “pre-thinking” and “speculation”. Therefore, the author here merely satisfies by raising the following four points to spark discussion.
First, if the arrival of the machine learning era signifies the restructuring of human knowledge systems, labor divisions, and lifestyles, then the overall content and overall patterns of human learning will also face restructuring. However, human learning may face severe challenges in this restructuring, involving deep-seated issues such as the nature of knowledge and the purpose of education, which are especially worthy of in-depth educational philosophical reflection. In the face of the challenges posed by artificial intelligence, Joseph E. Aoun, former president of the American Council on Education (ACE), proposed the idea of “robot-proof education”.[12] This does not mean to eliminate the involvement of machines in education, but emphasizes that the goal of education should shift towards cultivating abilities that cannot be replaced by machines. For example, he advocates for cultivating students’ divergent thinking because in terms of convergent thinking, machines far surpass humans: “When we use convergent thinking, we weigh data and alternatives to achieve the best, clearest results. This cognitive activity is precisely the field where high-end computers and machines are becoming increasingly proficient. Divergent thinking, on the other hand, exists in the form of freely flowing thoughts through creative generation of multiple reflections.”[13] However, the problem lies in the fact that training convergent thinking and mastering related knowledge often serve as the premise for effective divergent thinking. Current education has various well-known drawbacks; for instance, basic education places too much emphasis on assessing the “degree of knowledge mastery”, while higher education suffers from excessive specialization and knowledge-centric issues. In this regard, the power of machine learning seems to have a liberating potential, liberating us from the burdens of mechanical memory and simple accumulation of knowledge, allowing educators to return to the essence of education. However, will this liberation remove the ladder, causing future generations to lose patience and possibilities for gradual ascent? When humans no longer need to remember phone numbers, we may also forget a series of phone numbers. Similarly, when humans can easily retrieve knowledge from machines, to what extent will we distance ourselves from classical texts and the sources of civilization? Regardless, machine learning is successfully digitizing and informing all knowledge, and will thought thus transform into information processing? In the era of machine learning, human learning should aim at cultivating the ability to think beyond data and information. Therefore, whether we call it “divergent thinking”, “creative thinking”, “critical thinking”, or “systematic thinking”[14], or “high-road transfer”[15], the restructuring of learning must be contextualized in the machine era with the goal of transcending machines in human life.
Secondly, in the era of machine learning, it is necessary to re-examine the divide between the humanities and sciences. On one hand, technological literacy and data literacy should become basic competencies for future humans, “because coding is the universal language of the digital world, everyone should be familiar with it”.[16] If one does not understand programming languages, it means one cannot enter the civilizational interface of the machine learning era and cannot form intuition and judgment about the operational mechanisms of the era, thus failing to become a qualified citizen of the future world. On the other hand, technical experts urgently need to break through their blind spots. As Susskind states: “Within technology companies, almost no engineers are tasked with thoroughly contemplating the systemic consequences of their work. Most of them only need to solve certain isolated technical problems to get by.”[17] If such technical experts are driving human development, does the so-called “progress” not represent a significant risk? Will astonishing technological breakthroughs like machine learning bring darkness in the name of light?
From another angle, software engineers can be seen as the infrastructure workers of the machine learning era. As machine learning becomes established, are software engineers not also striving to build a world that does not require them? “People often say that software is eating the world, so we need software developers. But once software has consumed the world, what we need becomes unclear. What happens when robots learn to code themselves?”[18] Technical experts also need to regain humanistic literacy and escape the status of being mere tools.
Thirdly, the restructuring of learning also implies a reconstruction of the educational system. The modern educational system was established based on the knowledge concepts of the Enlightenment and the organizational structure of industrial society; for instance, the division of majors is inseparable from the entire structure of industrial society. When technological advancements render these concepts and structures ineffective, the modern educational system established based on them is likely at a moment urgently needing reform. As Aoun states: “The American educational system was largely designed to meet the needs of the industrial economic system of the 19th and 20th centuries, often emphasizing the functions deemed most valuable in a world filled with factories, bureaucracies, and ledgers.”[19] Such a situation is evidently not limited to the American educational system but is universally reflected in educational systems worldwide. However, if the economic system established by the Second Industrial Revolution has gradually become a relic of history, then the educational system built on it has become a tree without roots. Is it not time for urgent reform?
Finally, in discussions concerning “artificial intelligence and education”, commentators often start from the employment crisis hidden within the artificial intelligence revolution, using “the future of work” as a focal point to contemplate and design “the future of education”. This is not without merit; in today’s context, it is indeed a responsible educational consideration. However, we must question whether this is precisely still bound by the educational and learning concepts of the industrial era? If machine learning contains certain liberating potential, could it possibly liberate us from modern work concepts? In fact, as early as the beginning of the Second Industrial Revolution, Nietzsche lamented the forgetting of education, because the true “educational institutions” were conflated with “survival institutions”; the former is “the delicate, precious fairy of heaven”, while the latter is “the practical maid”.[20] Can machine learning compel us to escape Nietzsche’s cursed “19th-century barbarism” and allow future humans to return to the civilized concept of “learning to become an adult”? This may be too optimistic a hope. However, let us invoke Nietzsche’s words to summarize this educational philosophical reflection, ensuring that this contemplation on machine learning does not become overly bleak: “Please note, my friends, do not confuse two types of matters. A person must learn a lot to survive and fight for survival, but everything he learns and does as an individual for this purpose has nothing to do with true education. True education only begins far above this predicament, away from the struggle for survival and the necessities of life.” (See Nietzsche: “On the Future of Our Educational Institutions”, Commercial Press, 2019 edition, page 74. The translation has been slightly modified from the German edition.)
References
[1] Xu Yingjin. Fifteen Lectures on the Philosophy of Artificial Intelligence [M]. Beijing: Peking University Press, 2021: 76, 93.
[2][17] Jamie Susskind. The Power of Algorithms [M]. Beijing: Beijing Daily Press, 2022: 8, vii.
[3][6][15] Wayne Holmes, Maya Biliak, Charles Fadel. Artificial Intelligence in Education [M]. Translated by Feng Jianchao et al. Shanghai: East China Normal University Press, 2021: 94, 233, 33.
[4] Domingos, P. (2015). The master algorithm: How the quest for the ultimate learning machine will remake our world. London: Allen Lane, 6.
[5] Russell, S. J., & Norvig, P. (2022). Artificial intelligence: A modern approach. Global Edition, Pearson, 671.
[7] Arnold Pacey, Bai Fulian. Technology in World Civilization [M]. Beijing: CITIC Press, 2023: 195, 366.
[8][9] Cosimo Accoto. The Age of Data [M]. Translated by He Daokuan. Beijing: Encyclopedia of China Publishing House, 2021: 87, 46.
[10] García-Peñalvo, Francisco José. (2023). The perception of Artificial Intelligence in educational contexts after the launch of ChatGPT: Disruption or Panic? Education in the Knowledge Society, 24, 3.
[11] Adiguzel, et al. (2023). Revolutionizing education with AI: Exploring the transformative potential of ChatGPT. Contemporary Educational Technology, 15(3), 1.
[12][13][14][16][18][19] Joseph E. Aoun. The Future of Education [M]. Translated by Li Haiyan, Wang Qinhui. Beijing: Machinery Industry Press, 2018: 62, 64, 79-84, 62, 52-53, 66.
[20] Nietzsche. On the Future of Our Educational Institutions [M]. Translated by Peng Zhengmei. Beijing: Commercial Press, 2019: 76.
(Editor: Song Pengyang)
Welcome to scan the QR code to follow us