Immutable Explainability: Towards Verifiable and Auditable Affective AI

Abstract: Affective artificial intelligence has made substantial advances in recent years; yet two critical issues persist, particularly in sensitive applications. First, these systems frequently operate as 'black boxes', leaving their decision-making processes opaque. Second, audit logs often lack reliability, as the entity operating the system may alter them. In this work, we introduce the concept of Immutable Explainability, an architecture designed to address both challenges simultaneously. Our approach combines an interpretable inference engine - implemented through fuzzy logic to produce a transparent trace of each decision - with a cryptographic anchoring mechanism that records this trace on a blockchain, ensuring that it is tamper-evident and independently verifiable. To validate the approach, we implemented a heuristic pipeline integrating lexical and prosodic analysis within an explicit Mamdani-type multimodal fusion engine. Each inference generates an auditable record that is subsequently anchored on a public blockchain (Sepolia Testnet). We evaluated the system using the Spanish MEACorpus 2023, employing both the original corpus transcriptions and those generated by Whisper. The results show that our fuzzy-fusion approach outperforms baseline methods (linear and unimodal fusion). Beyond these quantitative outcomes, our primary objective is to establish a foundation for affective AI systems that offer transparent explanations, trustworthy audit trails, and greater user control over personal data.