Lei Li, Yue Zhu
The expanding deployment of artificial intelligence in criminal sentencing has intensified concerns over algorithmic opacity and the erosion of procedural fairness. Existing governance approaches tend to address transparency and accountability in isolation, often privileging a single stakeholder perspective while neglecting the structurally divergent needs of judges, defendants, algorithm developers, and the public. This paper proposes an integrated governance framework comprising three components: a multi-stakeholder demand analysis grounded in stakeholder salience theory that formalizes differentiated transparency requirements through quantitative need mapping and conflict intensity metrics; a multi-dimensional transparency evaluation model spanning technical, process, outcome, and interaction dimensions, operationalized via AHP-weighted fuzzy comprehensive assessment; and a dynamic accountability mechanism structured across ex ante, in-process, and ex post phases with blockchain-anchored traceability and closed-loop feedback. Empirical validation using the COMPAS recidivism dataset and a structured survey of 300 respondents across four stakeholder groups confirms the evaluation model’s discriminative robustness, with system rankings stable in 97.3% of sensitivity trials, and demonstrates a 70.5% improvement in accountability response latency over static audit baselines. Structural equation modeling reveals that accountability completeness is more strongly associated with stakeholder satisfaction than transparency is, with much of the transparency association operating through accountability as a mediating pathway. These findings suggest that disclosure mandates yield diminished legitimacy returns absent enforceable institutional consequence structures, offering actionable guidance for jurisdictions developing governance frameworks for sentencing AI.