News on March 18, with the rapid breaking of ChatGPT, AIGC (AI Generated Content) is regarded as a new production method that uses AI technology to automatically produce content after PGC and UGC. The rapid iterative evolution of AIGC has made the application of large models in a new field, and domestic enterprises have followed up the research and application of large models. While the big model brings transformative experience, it also brings many hidden worries such as content security, privacy protection, infringement, and legitimacy of results.
Big breakthroughs bring new problems
AIGC pushes artificial intelligence from perceiving and understanding the world to producing and creating the world. The content creation model has developed from early professional production (PGC) and user production (UGC) to AI-assisted user production (AIUGC), and then to the current AIGC (artificial intelligence to create content). The recently released GPT-4 by OpenAI, which has attracted much attention, is a multimodal large model (accepting image and text input and generating text).
In China, large Internet companies have long entered the field of large-model research, and with the popularity of AIGC, the layout of domestic enterprises in large-model has also surfaced. The mainland super-large model not only does not lag behind similar foreign products, but also surpasses it in some fields.
Experts predict that by 2030, synthetic data will replace real data as the primary source of data used by AI models. "The general multimodal large model with unified architecture, unified modality and unified task has gradually become a major trend in artificial intelligence research." He Xiaodong, vice president of Jingdong Group and IEEE Fellow, said in an interview with the "Communications Industry News" all-media reporter that multimodal large models are an important way for artificial intelligence technology to move from weak artificial intelligence in limited fields to general artificial intelligence. General-purpose multimodal large models will gradually play the role of basic models in the field of artificial intelligence.
The three elements of current artificial intelligence technology include data, algorithms, and computing power, but data is often static and algorithms are single-task-driven. He Xiaodong believes that in the future, it will be upgraded to the new three elements of scene, system and computing power, including complex interactive intelligent scenarios and the dynamic data they generate, multi-task collaboration and multi-algorithm fusion systems, and new computing power that can support complex scenarios and systems, so that computers can deeply understand the real world and solve major real problems, and improve the versatility, adaptability and task completion rate of intelligent systems. The new generation of fusion intelligence needs to be iterative and developed based on "live" interaction scenarios, including the interaction between humans and agents, and the interaction between multiple agents, which will become a new development trend of artificial intelligence.
Although there are still some technical challenges and imperfections in the big model itself, its power has been seen and has even adversely affected existing governance.
Privacy, bias, infringement...
Lian Shiguo, a top 2% of the world's top scientists and chief AI scientist of Unicom Digital, told the "Communications Industry News" all-media reporter that on the one hand, it does not learn in the way humans create content, does not learn the essence of creation, seems to have the style imitation and element reorganization of existing content, and the content it produces may have copyright infringement risks. For example, painting in the style of Picasso, writing a novel in the style of a writer, synthesizing the voice of a celebrity character, replacing the face of a character in a picture or video, generating a high-imitation digital avatar of someone, etc.
On the other hand, it is mainly built data-driven, and there are still problems such as the unexplainability and unfairness of deep learning, and the content it produces may be untrue and non-compliant. For example, ChatGPT will give different answers to the same question and different methods, drawing AI will produce different paintings, and even generate unfactual, discriminatory text/picture content.
Compared with PGC and UGC, AIGC improves content productivity, but when AIGC content is uncontrollable, the corresponding application risk will increase. For example, AIGC technology makes content creation easier, which also reduces the cost of infringement, counterfeiting, etc., and increases the cost of supervision; AIGC technology makes it possible to generate content quickly in batches, which can pose the potential risk of deliberately controlling the direction of public opinion.
He Xiaodong said that with the deepening of the application of artificial intelligence, its own technical shortcomings and the problems brought by privacy ethics, decision-making bias, and use security have caused a crisis of trust. Technically, the danger of weak and vulnerable algorithms; The black-box model makes the algorithm opaque, making it impossible for people to intuitively understand the reasons behind the decision. In application, the bias in the training data leads to the lack of fairness; The frequent use of biometric information represented by face recognition technology increases the possibility of privacy leakage. Ethically, the decision-making of AI systems is complex, and it is difficult to define the responsible subject, which brings ethical safety issues.
Govern AI with AI
Wu Xia, a legal observer in the communications industry, said that although the regulatory level has closely tracked new technologies and new formats, and issued relevant management regulations for Internet information service algorithm recommendation and deep synthesis, with the acceleration of practical innovation, legislation and law enforcement need to strengthen anticipation in combination with specific scenarios.
"Experts in the field of artificial intelligence are already thinking about technical means to identify AI-generated content." Lian Shiguo said that the legitimacy of content is judged through manual rules, and authentication is carried out through statistical analysis of text content and adding text watermarks. We believe that in addition to technical means, relevant digital governance systems will also be initiated.
"We need to build a safe and trustworthy AI governance system," He said.
The so-called "the road is one foot high, the magic is one foot high", and the technological evolution needs to be followed up by corresponding governance supporting technologies. Lian Shiguo believes that the healthy development of AIGC depends on improving controllability by improving technological maturity, upgrading regulatory means, and improving policies and regulations, including AIGC explainability technology innovation, AIGC content compliance detection, AIGC content piracy monitoring, AIGC content traceability, AIGC applicable laws and regulations, AIGC user risk awareness training, etc.
He Xiaodong believes that it is necessary to strengthen technical research on robustness, explainability, privacy protection and other aspects, solve the problems encountered in the current application of artificial intelligence, enhance the public's trust in artificial intelligence, ensure that applications and services minimize bias, and promote the healthy and high-quality development of artificial intelligence.
In addition, Wang Lei, an expert of the State Intellectual Property Office, said that the healthy and orderly development of large models requires work from two perspectives in terms of patents. The first is to use patent documents to sort out AIGC's technological development context, technical hardcore, and future development trend of technology. It is of great value to sort out patent documents, which can help the Chinese industry to look at the entire technical system from the Big Picture and find a favorable entry point and germination point. Second, a patent is an exclusive protection right, and its exploitation must obtain the permission of the patentee. The development of technology requires early attention to the distribution of patent rights, avoiding minefields and forbidden areas, or purchasing patent licenses under favorable conditions to avoid patent risks for future products and services.
In the application and development of emerging technologies such as AI, it is particularly important to coordinate development and security. Wu Xia believes that in terms of cybersecurity, when humans cannot easily distinguish whether there is a machine or a real person behind the conversation, the difficulty and pressure of cracking down on illegal crimes such as telecom network fraud will also increase sharply. Therefore, at a deeper level, in the rapid development of AIGC, we should also pay attention to cultural security.
"There is reason to believe that the rapid development of AIGC will eventually promote the progress of mankind," Lian Shiguo said.
Science and technology for good is the basic ethics of technological development, we have reason to believe that AI big models can "kind" understanding people, we should also embrace big models, improve in development, and standardize in progress. There is no problem with technology, the problem will only appear in the people who use it.
Written by: Cui Liangliang
Editing, proofreading: bright
Guide: New article