Lexing China tells you what you need to know about Ernie Bot, China tech giant Baidu’s ChatGPT-like Chatbot.
1. Background
On March 16, 2023, Baidu, the Chinese AI giant introduced Ernie Bot which is defined as “new generation large language model and generative AI product” and it was fully open to general public on August 31, 2023. After the announcement of Ernie Bot, over 75,000 corporate consumers requested API access on the first day. By November 10, 2023, Ernie Bot has registered over 70M users and its ecosystem involves more than 8 million developers and serves around 220,000 corporate consumers.
Ernie Bot’s functions are focused on Chinese language comprehension, multi-model generation (text, image, audios and videos), literary creation (conversational Q&A, summary and analysis), business copywriting, mathematic reasoning.
Ernie Bot incorporates Baidu’s ERNIE and PLATO models such as supervised fine-tuning, reinforcement learning and prompt learning. The training data of ERNIE Bot includes trillions of web pages, billions of search and image data, billions of speech calls and a knowledge graph of 550 billion of facts.
2. GAI Measures
Ernie Bot along with other Chinese chat bots developed by Chinese AI companies has prompted the Chinese regulator to enhance regulation on generative AI service and in this context a specific regulatory text on Generative AI service (“Interim Measures for the Administration of Generative AI Services” (“GAI Measures”)) [1] was jointly released by 7 Chinese ministerial departments on July 13, 2023.
According to GAI Measures, both the generative AI service provider and the user shall observe applicable law and regulations and respect social moral and ethics. The service provider shall take appropriate measures in case where it becomes aware of any illegal content or illegal activities committed by the users when using the services. In particular, the service provider shall take prompt measures to stop generating, transmitting illegal contents and eliminate the same and report the incident to the competent authorities. In case where the users are found engaging in illegal activities by using generative AI, the service supplier concerned, once alerted, shall take measures such as issuing warning (to the user concerned), restricting functions, suspending or terminating the service and keeping relevant records and reporting the same to the competent authorities [2].
The GAI Measures address the discriminatory practice with respect to algorithms design, training data, supply of service. The filing of algorithms and security assessment are conditionally required under applicable regulatory texts in order to prevent and combat the discriminatory practices feared by the regulator.
The GAI Measures also provide a 5-fold criterion with respect to training data including in particular lawful provenance of data and measures susceptible to enhance the veracity, accuracy, objectivity and diversity of training data.
This regulatory text is also reflective of the regulator’s consistent position on protection of minors which is translated in article 10 requiring the service provider to explicitly make its target users group, purpose of its services known to the public and take effective measures to prevent minors from overreliance upon or being addictive to the services based on generative AI.
3. Our reading of the “user’s instructions”
Obviously, the official launch of Ernie Bot on August 31, 2023 right after the release of GAI Measures was not coincidental. The official launch shall have been preceded by intensive efforts to have the algorithms duly filed and comply with a regulatory body composed of numerous national laws, regulatory texts (including GAI Measures), national standards.
The “user’s instructions” of Ernie Bot appears to largely echo the GAI measures. According to the said instructions, Baidu shall be entitled to monitor by manual or technical means the acts and information involved in the use of the service by the user, including but not limited the monitoring the inputs and outputs of contents, setting up risk filtering mechanism etc. The same instructions provide more than 30 prohibited/restricted practices for illustrative purpose. Baidu also requests its users to observe “applicable law, regulation, public order, social ethics, socialist value, national interests, others’ legitimate interests, moral value and information veracity when inputting, outputting, editing, commenting, uploading, publishing and diffusing information”. It is interesting to observe that Baidu explicitly requests “information veracity” from the user whilst obviously the regulator does not request the same from the generative AI service provider.
Article 5.3 of the said “Instructions” requests the user of services to ensure that his/her input contents not infringe/compromise others’ IPRs, dignity, name, privacy and the user shall be otherwise held liable for all direct or consequential damages arising from such infringement.
As per the same “user’s instructions”, the user is granted the right to use the generated contents for non-business purpose on a non-transferable, non-exclusive basis and such generated contents, if disclosed or made known to the public, shall be imperatively accompanied by a notice “generated by Ernie Bot app” which appear intended to comply with the transparency requirement posed by the GAI Measures.
Jun Yang
Jade & Fountain PRC Lawyers
Lexing China
[1] Interim Measures for the Administration of Generative Artificial Intelligence Services (生成式人工智能服务管理暂行办法) 13 July 2023
[2] Lexing International Newsletter “Lexing Insights” No. 38 “Artificial Intelligence”– November 2023
