Cognitive, industrial & philosophical depths of Artificial Intelligence-Generated Content (AIGC): perspectives from China

Leading minds in academia and industry share insights on AIGC's wide-ranging applications and current challenges, as well as the trustworthiness of technology

Against the explosive rise of artificial intelligence-generated content (AIGC) in 2022, six experts from China engaged in a discourse at Tencent Research Institute in January, 2023 over challenges and opportunities presented by AIGC. Tencent, the umbrella company and Chinese tech giant, has just released its own large language AI model “Hunyuan混元” last Thursday, marking its entry into the rapidly evolving AI race. The experts were:

• Yao Xin, Chair Professor and Head of the Department of Computer Science and Engineering at Southern University of Science and Technology

• Duan Weiwen, Director of the Research Office of the Philosophy of Technology, Institute of Philosophy, Chinese Academy of Social Sciences

• Wang Yuntao, Deputy Chief Engineer at Cloud Computing & Big Data Research Institute, China Academy of Information and Communications Technology

• Wu Baoyuan, Associate Professor at the School of Data Science, The Chinese University of Hong Kong, Shenzhen

• Yin Jun, Director of the Digital Content Technology Center, R&D Efficiency and Capability Department of Tencent Games CROS

• Shi Shuming, Director of the Natural Language Processing Center at Tencent AI Lab; Producer of Effidit

Notably, these experts' perspectives encompassed an amazing array of realms — cognition, psychology, education, art, economy, manufacturing, and internet governance. Although they agreed that Artificial General Intelligence (AGI) is still a distant goal, the vast potentials of AIGC applications were uniformly recognized. Beyond its well-discussed impacts on media, e-commerce, film, and entertainment, AIGC has potential to spark creativity in content creation, redefine cognitive perception, revolutionize learning and research, and boot work productivity. It can even address issues like autism and social anxiety.

What makes their insights even more interesting is the diverse tapestry of philosophies and disciplines they drew upon, ranging from computer science to geometry, from cybernetics to Saul Kripke's possible worlds theory and Marxism. From the lens of the possible worlds theory, for example, AIGC is capable of creating the metaverse—a space where countless "possible worlds" coexist and interact, offering limitless opportunities for human creativity and exploration. When placed within the framework of Marxist theory, AIGC emerges as a revolutionary power that transforms forces of production and productivity, potentially reshaping the relations of production and impacting human society.

This is an extract of their discussions. For the full-text transcript, please check out the WeChat blog of Tencent Research Institute.

---

Cumulative efforts behind the explosion of AIGC

Yao Xin: The explosion of AIGC is a cumulative success. While AI has a long history of content generation, limitations in data, computing power, and algorithm technology have held it back from generating massive interest and surpassing people's expectations as it has done recently. Nonetheless, it's essential to take a closer inspection on what precise technological breakthroughs have taken place.

First, it’s relatively easy to exceed expectations in the technological realm. For instance, AIGC can emulate a certain artist and create convincing images, dialogues, and even music in that artist’s style. However, I am doubtful of the common belief that AIGC may be an effective pathway to artificial general intelligence (AGI). This is because AIGC generates content by utilizing vast amounts of data and immense computing power, rather than engaging in abstract thinking that is often seen in human creativity. For example, when I observe numerous images, my brain goes through an abstract process before returning to the actual space to create a new image. I’m not sure whether AIGC currently has such abstract functions; I think a machine is unable to generate concepts such as four legs, fur, etc. even if it has seen a million photos of cats and dogs.

Second, additional challenges will arise as AIGC expands deeper into science and technology or fields related to people’s welfare. It will become increasingly difficult to ensure that AIGC-generated content always adheres to predefined constraints. Let me use an analogy: we can now provide many molecules to an AI system, and it may generate new molecules in response. While developing drugs from these molecules is conceivable, there may be a gap between the theoretical concept and practical application. There’s a long way ahead to achieving the so-called AGI.

Yao Xin, Chair Professor and Head of the Department of Computer Science and Engineering at Southern University of Science and Technology

Yin Jun: I couldn’t agree more with Professor Yao. There has been a substantial amount of research on both the broader concept of AIGC or the specific focus on Large Language Models (LLMs). The current explosion of AIGC is the result of a multitude of factors: accumulation of extensive high-quality open datasets, theoretical breakthroughs such as the Diffusion Model, and the availability of new computing hardware that supports larger and more efficient models such as GPT-3. Even today, the cost associated with training such models remains notably high. The combination of these factors has made the generation of images or text much more attainable than before. This, I think, explains the explosive rise of AIGC this year.

It's crucial to recognize that in comparison to earlier deep neural networks (DNNs), AIGC has not yet undergone a fundamental paradigm shift. It will be unwise to jump to the conclusion that ChatGPT has attained AGI based on its proficient performance in generating conversations and its appearance of intelligence. ChatGPT doesn't possess genuine understanding of the content it generates; it merely gives the illusion of comprehension. This is still a considerable distance from human intelligence.

I think AI needs to focus on several essential areas in the next phase of development. First, it should achieve autonomous learning. Like humans, some of its reasoning processes should be made explainable or understandable, and it should demonstrate the capability to perform tasks across diverse domains. Second, on the industrial front, AI should be able to assist game content generation. The current capabilities of AIGC models still fall significantly short of professional standards in the gaming industry, whether in generating images, text, 3D models, character animations, or using ChatGPT to create game scripts or non-player character (NPC) dialogues.

Application practices of AIGC

Wang Yuntao: The current practice of AIGC can be categorized into four main processing modes: text, audio, image and video, and virtual space.

1. Text: This mode primarily involves generating or editing text content, including tasks such as article generation, text style conversion, and question-and-answer dialogues. Common applications include writing robots, chatbots, etc.

2. Audio: AIGC can performs tasks like text-to-speech (TTS), speech-to-text, and speech editing. It also extends to non-speech content such as music generation and sound scene editing. Typical applications include intelligent dubbing, virtual performances, automatic music composition, and generating songs.

3. Image and Video: This mode includes facial generation, facial replacement, character attribute editing, face manipulation, and posture manipulation. Technologies such as image generation, enhancement, and restoration are also relevant. Representative applications involve beauty filters, face swapping, image replays, style modification, and AI painting.

4. Virtual Space: This mainly involves 3D reconstruction and digital simulation, and generating/editing digital characters or virtual scenes. Prominent applications are metaverse, digital twins, game engines, 3D modeling, and virtual reality (VR).

AIGC applications offer significant advantages in providing more diverse, dynamic, and interactive content experience. It has already led to substantial improvements in industries with a high-level digital culture and diverse content demands:

1. AIGC + Media: AIGC has played an important role in media convergence through human-machine collaboration. The time for a robot to generate a comprehensive report has been dramatically reduced from 30 seconds to under two seconds.

2. AIGC + E-commerce: The core application in this realm is the generation of 3D models, which enhances online shopping experience through virtual product displays and trials. Additionally, AIGC is leveraged to create virtual anchors, bolstering live-streaming sales by providing engaging and informative content.

3. AIGC + Film & Television: AIGC has expanded creative possibilities in the film and television industry, elevating the quality of productions. It has already been used for generating script ideas for several theatre hits in China, with services like intelligent conversion from fiction to script. Furthermore, AIGC has aided image processing and post-production in a number of TV and film productions in China.

4. AIGC + Entertainment: This refers to creating entertaining images and audiovisual content. It's also driving consumer-focused applications to explore the metaverse, opening up new dimensions of entertainment experiences.

AIGC has also shown promise in medical and industrial sectors. However, this is still in the early stages of industrial integration and business logic implementation.

Wang Yuntao, Deputy Chief Engineer at Cloud Computing & Big Data Research Institute, China Academy of Information and Communications Technology

Shi Shuming: AIGC has witnessed remarkable progress in overall technological advancement. I recall five years ago, only TTS generation was deemed practical in the AIGC domain; a mere three years ago, the idea of AI producing high-quality and contextually relevant images based on text inputs was unimaginable.

Text generation relied heavily on specific models with limited applicability, but with the advent of LLMs and continuous refinement of language technology, AIGC has become increasingly impressive. Both Stable Diffusion and ChatGPT have astounded people with their robust text comprehension and content generation capabilities.

Of course, China needs more efforts in advancing AIGC. A predominant share of AIGC research and development was conducted by a few institutions in the United States. They have been the leaders of AI technology, and China must also endeavor to contribute more to the further development of AI.

As for commercial applications, AI-enhanced human capabilities such as AI-assisted creation stand out. While a large volume of images autonomously generated by AIGC are indeed meaningless, it can prove useful if given prompts, or a combination of hints, in generating high-quality contents. Through iterative testing and interaction, AIGC can assist most people with limited artistic abilities in creating visual content. The same principle applies to text generation. AIGC's assistance can significantly improve efficiency in text generating and rewriting, while also sparking creativity.

Objectively speaking, AIGC undeniably enhances productivity and work efficiency, the most direct application being AI-assisted creation, but there are still uncharted territories to explore. Of course, some are wondering whether ChatGPT can replace search engines; I believe it is unlikely at the current stage. ChatGPT may fulfill part of the functions of search engines but cannot entirely replace them.

In a word, AIGC technology is advancing rapidly and exceeding expectations; second, there is ample room for innovation in the realm of commercialization, but the most crucial aspects are yet to be explored.

Shi Shuming, Director of the Natural Language Processing Center at Tencent AI Lab; Producer of Effidit

Value of AIGC

Duan Weiwen: Overall, AIGC will inspire innovations in the following four aspects:

1. AIGC innovates not only content creation but also cognitive perception. With AIGC, individual “ideas” can take on greater significance in content creation.

2. AIGC is a new tool for learning and research, empowering individuals with more advanced creative abilities. This is evident in the recent controversy that many undergraduates are using AIGC to finish theses, which some consider as plagiarism. But the truth is, most undergraduate theses are copy and paste; some students do it better than others. What AIGC does is streamline literature search and processing, thereby enhancing learning efficiency. AIGC may become a routine research tool in the collaborative process of human-machine cognition.

3. AIGC can play a role in the development of the metaverse, that is, the creation of a virtual world. AI art brings together myriad possibilities akin to Saul Kripke's possible worlds theory. [According to American philosopher Saul Kripke, modal facts are construed as facts about possible worlds, where the actual world is just one in which as many aspects of the world as possible as similar to ours.] In the past, individual brains had limited access to these potential data resources, but AIGC can seamlessly integrate these resources based on one's imagination, effectively becoming a tool for crafting possible worlds. This significantly expands the metaverse's potential, allowing for the amalgamation of humanity's collective creative ideas, cultural heritage, and spiritual wealth.

4. Addressing autism is also an interesting application. Nowadays, many suffer from social anxiety, and AIGC can provide a platform for interaction with one’s digital alter ego. For example, an artist trained AI with her childhood diary and engaged in conversation with her younger self, eventually gaining insight into her adolescent concerns and achieving a therapeutic effect. Therefore, AIGC contributes to spiritual self-awareness, self-healing, and may have the potential to become a valuable companion, empowering individuals to embark on a journey toward self-renewal and greater spiritual strength.

Duan Weiwen, Director of the Research Office of the Philosophy of Technology, Institute of Philosophy, Chinese Academy of Social Sciences

Wang Yuntao: I am more often involved with industries, so the first association that comes to mind when I think of AIGC is "supercomputing." AIGC may introduce new challenges to the current computing architecture, including the computing system. As we strive to effectively implement heterogeneous AI systems, we are bound to uncover various limitations in computing devices, data storage, and hardware-software coordination, all of which will necessitate innovative solutions.

As for applications, the most prominent challenge posed by AIGC to traditional industries is for content technology. The landscape of content creation has shifted from a centralized platform model to a more user-generated, decentralized one. In this transformation, AI technology is playing an increasingly revolutionary role, influencing aspects like content generation, distribution, and moderation.

Trustworthiness of AIGC

Wu Baoyuan: ChatGPT has gained significant popularity, but it also has problems like the potential to generate erroneous information or content that appears correct but fundamentally false. However, it's important not to overly fixate on these challenges, as doing so can unduly hinder technological progress. For instance, in Deepfake research, academics must declare ethical implications and potential technical risks when studying generative adversarial attacks but not when studying defense and detection mechanisms, leading to a disproportionate focus on the latter. This may limit the exploration of innovative approaches, even though attacks can also inspire advances in defense.

Also, in the context of the digital economy, AIGC has the potential to serve as a valuable tool for generating data, safeguarding privacy, substantially reducing the cost of data collection, and even creating entirely new data. In conclusion, AIGC has its social issues and innovative challenges, but it could benefit from the development of more diverse application scenarios to bring out its positive side

Wu Baoyuan, Associate Professor at the School of Data Science, The Chinese University of Hong Kong, Shenzhen

Duan Weiwen: The age of search engines and the platform economy saw the transformation of the world into data, a process known as the "datafication of the world"; corresponding concerns in privacy, ethics, and law were addressed with new approaches. Now, as the world transitions into the AIGC era, content is being generated on the basis of "datafication", marking the advent of the second-order datafication age. This process is akin to how Euclidean geometry was developed based on measurement of the world, which eventually grew into a novel form of cognition and knowledge production.

New legal and ethical challenges in the AIGC age necessitates the establishment of new social contracts and consensus regarding acceptable and unacceptable behaviors. Regulators, industry leaders, legal and ethical scholars, and the broader community must collaborate in devising a predictable governance model that safeguards AIGC innovation — not just safeguarding economic interests but also ensuring that society can embrace technology with confidence, which is only possible if ethical and legal issues have been taken into consideration from the very beginning.

For example, AIGC has the potential to contribute to the identification and purification of data toxicity, a concept frequently discussed today and one that mirrors real-life issues like biases and discrimination. However, it's crucial to understand that there is no universal standard to definitively define what is pure or contaminated. Consequently, this purification process cannot be absolute, as striving for absolute purity would contradict some of the core principles in modern society. Progress can only be made through the recognition of complexity of these issues and ongoing exploration to determine what is acceptable and what is still evolving.

Yao Xin: The development of safe, trustworthy, and responsible AIGC has lagged behind in certain aspects. There are several key issues at play:

1. Data Quality Concerns: A significant portion of the data used for training AI models is sourced from the internet, and much of it may be incorrect or inaccurate. These data imperfections get propagated through the training process, leading to potential errors that accumulate in large AI models. These errors can become entrenched and challenging to rectify.

2. Proactive Safety and Trustworthiness: To effectively apply AIGC in industry or applications closely related to people, safety and trustworthiness should be integral considerations throughout the model development and training process. It should be a continuous concern from the outset.

3. Ethical Education and Knowledge Diversity: The utilization of ChatGPT in student theses raises questions about what educational institutions should teach students and how they should teach them. Over-reliance on a single AI model for knowledge generation could lead to a loss of diversity in knowledge production, which may have profound implications for our society. These issues should be taken into account right from the inception of AIGC. Failure to do so could result in a future where we find ourselves constrained by the limitations and biases of large AI models, similar to how recommendation algorithms have narrowed our worldview.

Future of AIGC

Yin Jun: The advent of AIGC and the broader landscape of AI are bound to bring about a profound transformation in the current forces of production and productivity, which will reshape the relations of production [quoting Marx’s theory of production, which sees the production of material life as pivotal in shaping human history], potentially yielding significant implications for the future of humanity and society.

Yin Jun, Director of the Digital Content Technology Center, R&D Efficiency and Capability Department of Tencent Games CROS

Wang Yuntao: AIGC will likely be a significant opportunity for the emerging digital-native world in the future. Just like the digitalization of the physical world, AIGC promises to play a pivotal role in ushering in a metaverse where individuals can conjure up new applications, formats, and business models seemingly out of thin air, with AIGC playing a pivotal role in this creative process. Nevertheless, it also presents formidable challenges, including implications for traditional economic theory. AIGC has the potential to alter the cost structure of production by significantly driving down the marginal cost of intelligence and elevating its marginal revenue. The future will become ever more intricate and varied.

Internet-based Philanthropy in China