从算法“黑箱”走向算法透明:基于“硬法—软法”的二元法治理模式

From Algorithmic“Black Box”to Algorithmic Transparency:A Dualistic Law Governance Model Based on“Hard Law⁃Soft Law”

算法是“程序设计+数学规则”的集合,其“黑箱”特性虽然并非算法风险的唯一原因,但却是规制算法风险必须要解决的问题。强调开放与控制并重的“硬法—软法”范式能够为算法“黑箱”的治理提供一种新的立法框架。基于理论层面,硬法规范的初次外在规制决定了算法“黑箱”治理的裁量边界,软法规范的二次内在规制能够为政府的算法“黑箱”治理行为开放公众参与机制和提供内在参照标准。基于实践层面,欧盟与美国已形成的算法软硬法混合治理格局能够为中国建立算法“黑箱”的“硬法—软法”二元法治理模式提供借鉴。算法“黑箱”治理的软法规范应包括增强算法模型的可解释性和向社会公开算法源代码,同时,宜借鉴“遵守或解释”机制,要求算法开发者/运营者在不遵守软法规范时应作出合理解释说明。算法“黑箱”治理的硬法规范应包括算法权利(最为密切的是算法解释权、算法自动化决策拒绝权),以及借助算法备案、影响评估和合规审计启动算法问责。

Algorithm is a collection of“programming+mathematical rules”.Although its“black⁃box”nature is not the only cause of algorithmic risk,it is an issue that must be addressed when regulating algorithmic risk.The“hard law⁃soft law”paradigm,which emphasizes both openness and control,can provide a new legislative framework for the governance of algorithmic“black box”.From the theoretical perspective,the primary external regulation of hard law norms determines the discre⁃tionary boundary of algorithmic“black box”governance,while the secondary internal regulation of soft law norms opens up the public participation mechanism and provides internal reference standards for the government’s algorithmic“black box”governance.From the practical perspective,the European Union and the United States have established a pattern of mixed hard law and soft law governance of algorithm,which can provide a reference for the establishment of“hard law⁃soft law”regulatory model of algorithmic“black box”in China.Soft law regulation of algorithmic“black box”should include enhan⁃cing the interpretability of algorithmic models and disclosing the source code of algorithm to the public,and at the same time,it is appropriate to draw on the“comply⁃or⁃explain”mechanism,which requires algorithmic developers/operators to provide reasonable explanations in the event of non⁃compliance with the soft law regulation.Hard law regulation of algorithmic“black box”should include algorithmic rights(the right to interpret algorithm and the right to reject automated algorithmic decision⁃making which are most closely related),as well as algorithmic accountability through algorithmic filings,impact assessments and compliance audits.