Trust-Aware Self-Supervised Learning: Modelling Human Trust Dynamics in Human-AI Collaboration Systems

Abstract

In an era where artificial intelligence (AI) increasingly influences critical decisions, success is no longer defined solely by technical performance; it also hinges on the system’s ability to foster and align with human trust. The proposed study presents a new framework called Trust-Aware Self-Supervised Learning (TA-SSL), which aims to learn implicit human trust through conversation behaviour, enabling AI systems to respond according to the current trust levels between AI and users. In contrast to the current models trained with explicit trust labels, TA-SSL is trained on behavioural cues, indicative of the trust level, to the AI during human-AI interactions, including hesitation, clarification requests, and action reversals. With a temporal contrastive learning goal, TA-SSL induces sparse, dynamic embeddings of trust incorporated into decision-making strategies, including explanation depth, uncertainty mediation, and user autonomy management. We justify our visitation with the Kaggle Human vs Robot talking information, which contains over 10,000 speech samples from a crowd. TA-SSL exceeds these static, supervised, and reinforcement learning baselines to record +19.3% more successful task instructions, -24.6% fewer irrelevant clarification requests, and a trust calibration value of 0.76 (0.42 in supervised models). The obtained trust embeddings, exhibited by the evaluation measure, exhibit a high degree of temporal consistency and specificity to the user and the visualisations shown, using PCA as a visualisation method. PCA shows separable groups in low, medium, and high trust conditions. Case studies also show that the model can be flexibly restructured or strengthened to increase user trust in changing user behaviour. The study has demonstrated that latent trust can be effectively learn based on unlabelled behavioural indicators and informed in bespoke ways to drive AI behaviour. TA-SSL is a way to create a scalable, domain-agnostic pipeline to build user-aligned, emotionally intelligent AI systems and representation learning with innovation psychology to boost trustworthy human-AI collaboration.