is the generative AI "rules" Concerns about these concerns are like a shadow, forcing relevant countries and regions to speed up the legislative process and build a "guardrail" for the healthy development of generative AI.
In June this year, the European Parliament passed the draft "Artificial Intelligence Act" with a high vote, which means that the European Parliament, EU member states and the European Commission will start "tripartite negotiations" to determine the final terms of the bill. The United States is also actively discussing issues related to AI regulation.
returned to China. In April this year, the Cyberspace Administration of China drafted the "Administrative Measures for Generative Artificial Intelligence Services (Draft for Comment)". In July this year, the revised "Interim Measures for the Administration of Generative Artificial Intelligence Services" (the "Measures") was officially promulgated and will come into force on August 15, 2023. This is considered the world's first generative AI legislation. The "Measures" make specific provisions on generative artificial intelligence in terms of technology development and governance, service standards, supervision and inspection, and legal responsibilities, setting the tone for the future development of generative AI in my country.
Source: China Netcom
Industry insiders believe that after the release of the "Measures", compliance has become a must for large-scale model manufacturers. What changes will the "Measures" bring to large model manufacturers? The "IT Times" reporter asked nearly ten large model manufacturers such as Tencent, Alibaba Cloud, and Baidu, but none of them gave a positive response.
Generative AI has laws to follow
On July 13, the revised "Interim Measures for the Management of Generative Artificial Intelligence Services" was officially released and will come into effect on August 15, 2023.
Compared with the "Generative Artificial Intelligence Service Management Measures (Draft for Comment)" released 3 months ago, it can be found that the "Measures" have made major changes.
For example, the "Measures" newly proposed the principle of "adhering to equal emphasis on development and security, and combining innovation promotion with law-based governance", advocating inclusiveness and prudence, classified and hierarchical supervision, and expanding the encouragement clause from one to three in the draft for comments.
Another example, the "Measures" no longer emphasizes that the content generated by generative artificial intelligence should be true and accurate, but only points out that the accuracy and reliability of the generated content should be improved, respecting the new characteristics of large-scale model technology.
At the same time, the "Measures" changed the requirement that users must provide real identity information to "generative artificial intelligence service providers should sign service agreements with service users who register their services, clarifying the rights and obligations of both parties", expanding the scope of services, etc. .
In the eyes of many people in the industry, for the Internet industry, the "Measures" have made great progress compared with the draft for comments released in April. "This is not common in the formulation of general laws and regulations." Shi Yuhang, a lawyer of Huiye Law Firm, believes that the changes to the "Measures" have absorbed the opinions of relevant practitioners, indicating that the country is promoting the development of generative AI. positive attitude.
However, the "Measures" did not fully explain the concerns of the industry. Shi Yuhang pointed out that the "Measures" did not give detailed explanations on how to determine the legality of data sources and how to understand the intellectual property rights enjoyed by others in accordance with the law. They can only be explored in subsequent judicial and law enforcement practices. Jurisprudence to clarify the direction. However, in Shi Yuhang's view, the introduction of the "Measures" at least gives domestic generative AI manufacturers "a law to follow."
compliance has become "required"
"Measures", what practical impact will it bring to large model manufacturers?
According to Shi Yuhang, the introduction of the "Measures" means that technology will no longer be the only factor that determines the development of large-scale model manufacturers, and compliance will also affect the development speed of enterprises. and compliance capabilities will directly determine how far generative AI products and services can go.
In the latest "Measures", Article 17 clearly states that to provide generative artificial intelligence services with public opinion attributes or social mobilization capabilities, security assessments shall be carried out in accordance with relevant national regulations, and shall be managed in accordance with the "Internet Information Service Algorithm Recommendation Management" "Regulations" to perform the procedures of algorithm filing, modification, and cancellation filing. In other words, if generative AI for the public wants to be put into use, in most cases, it needs to go through the filing and safety assessment.
Source: Dongfang IC
"This requires companies to consider compliance issues in advance." Shi Yuhang explained that because the compliance chain of generative AI is very long, the project establishment, research and development, testing, release, update, and delisting of the project every Each link involves a large number of compliance control points and multiple filing and application checkpoints, involving laws and regulations in different fields. In the filing stage, companies need to submit compliance certificates for each process involved in generative AI. A large-scale model start-up team told reporters that they are currently studying the "method" and integrating it into the design of their own products.
Specific to the enterprise, the changes brought about by compliance will not only stay at the theoretical level, but will involve all aspects of the enterprise. In Shi Yuhang's view, the deployment of compliance will inevitably touch changes in the underlying mechanism and content of the enterprise. From the internal system reform of the enterprise, the setting of positions to the establishment of the organizational structure, from product design, testing to R&D layout, all links need to be adjusted according to compliance requirements. "The company's process structure must be sound, at least not just considering the investment in technology research and development." Shi Yuhang said.
But for some businesses, compliance issues may present new problems. Because compliance itself requires cost investment, for some companies with insufficient funds, it is necessary to make a trade-off between research and development and compliance. Shi Yuhang said: "This may drag down the progress of enterprise research and development."
implementation faces challenges Implementation of the regulations is still difficult.
Lin Hui (pseudonym) has recently received orders for compliance solutions from several companies.
"We were also forced to do it by several customers." After the release of the new regulations, at the request of customers, Lin Hui and his team began to dismantle the new regulations and create customized compliance solutions based on the customer's products.
"Guiding opinions do not represent standard answers." Lin Hui told reporters that the biggest difficulty that manufacturers currently encounter in compliance issues is the lack of specific application scenario rules. Although the "Measures" has set up a framework, it is still a guiding opinion and cannot cover all application scenarios. The specific boundaries and scales still need to be explored by enterprises according to their own application scenarios. "This is a delicate matter. For example, if you formulate 10 compliance requirements according to the "Measures", if there is a problem with the excess, is it considered a violation?" Lin Hui said.
Generative AI compliance not only involves laws and regulations, but also involves technical operations, which also brings difficulties to the implementation of compliance solutions. In Lin Hui's view, in order to deduce a suitable plan based on the existing laws and regulations, planners need to understand the law, security, and technical architecture, and at the same time, be responsible for all regulatory issues such as public security, network security, network information, and finance. The regulations are very clear. Only in this way can we formulate plans based on the regulatory habits of relevant agencies. "Even so, there is no guarantee that the compliance plan formulated is correct." Lin Hui said.
In specific operations, a compliance plan for often includes product compliance, business rectification, and privacy statement descriptions. There is also an ambiguity in the filing process of
. Shi Yuhang believes that since the algorithm itself is difficult to describe, during the filing process, enterprises need to explore the boundary of describing the algorithm and its security. In terms of compliance review, the company often adopts the self-certification method, and the company also needs to prepare materials according to its actual situation. He said: "There is no unified standard yet, and companies still need to explore it themselves."
In June this year, the National Internet Information Office issued the "Announcement on the Release of In-depth Synthesis Service Algorithm Filing Information". In , a total of 41 generative AIs have completed the filing, including DingTalk's intelligent customer service algorithm and Dharma Academy's open-domain natural dialogue synthesis algorithm. , Baidu Wenshengtu content generation algorithm, Tencent meeting virtual background algorithm, Xunfei speech recognition algorithm, etc.
However, most large-scale model manufacturers are still secretive about the content of filing and compliance. "IT Times" reporters tried to interview Tencent, Ali Nearly ten large model manufacturers such as Cloud and Baidu did not give a positive response. Even though has passed the relevant filing, a large factory responded that the filing process and other aspects are still in an indescribable stage and cannot make a positive response.
Typesetting / Ji Jiaying
picture / China Netcom East IC
source / "IT Times" public account vittimes
E N D
