When asked explicitly, a Large language model (LLM) may validate your anger-but implicitly, it may still judge that anger as inappropriate.We call this divergence the endorsementexposure gap, and it reveals that LLMs encode hidden norms about which emotions are acceptable in which contexts.To measure these norms systematically, we introduce FEELING RULES ATLAS, a benchmark of 1,320 vignettes spanning 6 institutional settings, 12 roles, 7 emotions, and 5 intensity levels.We pair the benchmark with two probes: explicit norm judgments (APPROPRIATE/INAPPROPRIATE/DEPENDS) and implicit acceptability scored by log-likelihood contrast.Across six model families, we find large cross-model variation in sanctioning thresholds and institutional "norm signatures" not reducible to overall strictness; models that appear similarly lenient explicitly can diverge sharply in implicit judgments.These results establish normative affect: contextconditioned judgments of emotional appropriateness, as a distinct alignment axis, and motivate transparent profiling of feeling rules for emotionally sensitive deployments.

For web platforms facing regulatory scrutiny---from content moderation to ad delivery and recommendations---fairness audits routinely disagree due to metric choice, subgroup granularity, sampling variance, and dataset shift. Point estimates yield brittle pass/fail narratives that are hard to defend in governance contexts. We propose a Bayesian audit-of-audits that pools heterogeneous audits---count-based and metric-only---into interval-valued fairness claims with explicit uncertainty and policy-risk tables aligned to practitioner thresholds. The framework unifies classification and exposure metrics, enforces consistency across coarse and intersectional group definitions via soft coherence constraints, and quantifies the Value-of-Information of prospective audits. We also provide heterogeneity diagnostics and leave-one-audit-out sensitivities. Across a synthetic Audit Zoo, a content-moderation case study on CivilComments--WILDS, and an ad-delivery simulation, our meta-evaluator attains near-nominal coverage with narrower intervals and fewer decision flips than per-audit baselines, while integrating partial-information audits.

Abstract This comprehensive review of concept-supported interpretation methods in Explainable Artificial Intelligence (XAI) navigates the multifaceted landscape. As machine learning models become more complex, there is a greater need for interpretation methods that deconstruct their decision-making processes. Traditional interpretation techniques frequently emphasise lower-level attributes, resulting in a schism between complex algorithms and human cognition. To bridge this gap, our research focuses on concept-supported XAI, a new line of research in XAI that emphasises higher-level attributes or 'concepts' that are more aligned with end-user understanding and needs. We provide a thorough examination of over twenty-five seminal works, highlighting their respective strengths and weaknesses. A comprehensive list of available concept datasets, as opposed to training datasets, is presented, along with a discussion of sufficiency metrics and the importance of robust evaluation methods. In addition, we identify six key factors that influence the efficacy of concept-supported interpretation: network architecture, network settings, training protocols, concept datasets, the presence of confounding attributes, and standardised evaluation methodology. We also investigate the robustness of these concept-supported methods, emphasising their potential to significantly advance the field by addressing issues like misgeneralization, information overload, trustworthiness, effective human-AI communication, and ethical concerns. The paper concludes with an exploration of open challenges such as the development of automatic concept discovery methods, strategies for expert-AI integration, optimising primary and concept model settings, managing confounding attributes, and designing efficient evaluation processes.

This paper proposes a method for human action classification by utilizing correlations between SURF based descriptors. This approach provides us a novel type of descriptor that can be used for action classification. The method proposed is tested using an SVM classification technique. For evaluation purposes, the KTH action recognition dataset, which is a standard benchmark for this area is used as it is one of the most well known and challenging dataset. The method proposed was able to successfully classify different action classes.