面向产业链的治理:人工智能生成内容的技术机理与治理逻辑

Industry Chain-Oriented Governance:Technological Mechanisms and Governance Logic in the Management of Artificial Intelligence Generated Content

以ChatGPT为代表的大型生成式语言模型通过对海量自然语言文本数据的学习,运用统计学和概率分布预测下一个单词或句子,从而实现流畅连贯、语法正确、具备逻辑的文本生成。这一技术机理虽能提升模型性能,但却在内容公平性、真实性和可靠性方面带来治理挑战。概率生成文本的方式存在难以控制内容质量的问题。基于网络文本语料库的训练可能嵌入算法偏见,人类反馈强化学习可能加剧虚假信息的生成与传播风险,模型的涌现性还可能产生意料之外的治理风险。迈向“模型泛在”的人工智能2.0时代,生成式人工智能的治理应遵循敏捷治理与韧性治理并重、精准治理和参与式治理协同的原则。针对“模型即服务”的新型产业生态链,围绕信息生产变迁和产业链变迁的核心特性,构建基于应用程序接口模式的治理、基于插件模式的治理和基于模型垂直部署模式的治理架构。

Large-scale generative language models,exemplified by ChatGPT,predict the next word or sentence by learning from a vast amount of natural language text data,utilizing statistical methods and probability distributions.This mechanism produces text that is coherent,grammatically correct,and logical.However,it introduces governance challenges concerning content fairness,authenticity,and reliability.The probabilistic nature of text generation makes it challenging to control the quality of the generated content.Training based on network text corpora may introduce algorithmic bias,and human feedback reinforcement learning may exacerbate the risk of generating and disseminating false information.The emergent behavior of the model may lead to unforeseen governance risks.In the era of AI 2.0,the governance of generative artificial intelligence should adhere to the principles of agility and resilience,precision governance,and participatory governance.For the new industry ecology chain of“Model as a Service”,a governance framework should be established in line with the core characteristics of information production transformation and industry chain migration.This framework should incorporate governance based on Application Programming Interfaces(APIs),plugin-based governance,and governance of vertical deployment models.