Due to the recent rapid development of artificial intelligence (AI) and its expanding impact on the planet, green and sustainable AI research has increasingly gained attention. This systematic literature review searches main databases, including Scopus, Web of Science, and Google Scholar, using an organized methodological approach. Following a thorough screening process, 49 final studies published between 2016 and 2026 are selected from an initial identification of 325 original records. We identify and analyze four key categories of sustainable AI practices: (1) model-level algorithmic efficiency, (2) hardware- and system-level optimization, (3) lifecycle- and data-centric approaches, and (4) operational and policy-level sustainability. We also highlight and explain four dimensions at the intersection of AI and environmentally responsible behavior: AI for sustainable applications’ development in industries, ethical considerations and accountability in using AI, and opportunities enabled by generative AI. We then combine existing taxonomies, evaluation metrics, and challenges to identify areas for improvement and suggest future research directions. Based on our analysis, we emphasize the need for interdisciplinary cooperation to facilitate responsible AI innovation and match it with global sustainable development goals (SDGs). We also highlight the importance of developing adequate frameworks along with precisely defined and standardized metrics to assess the environmental impact of AI. This review aims to encourage more responsible and environmentally friendly AI practices by providing a structured framework for researchers, educators, and professionals engaged in sustainable AI.

BACKGROUND World Health Organization reports that chronic non-communicable diseases account for 74% of global deaths. Despite rapid advances in digital health technology, Artificial Intelligence tools for self-management remain deficient in two crucial elements: emotional connection with patients and trustworthiness. Concern around these two topics is of increasing interest and importance. With regards to trustworthiness of AI and emotional intelligence of AI however, the studies for these two concepts were developed completely separately and in an isolated manner and this in itself, is a considerable design gap. Within self-management for chronic conditions, it becomes necessary to build an approach to design with the integration of these two concerns. OBJECTIVE This systematic review aims to examine the extent to which the concepts of trustworthiness, emotional intelligence, situational awareness and personalization are integrated within artificial intelligence systems designed to facilitate self-management of chronic illness, and what the impact of integration is. METHODS From the beginning of February 2026, a thorough search was undertaken on 6 databases (PubMed, Scopus, IEEE Xplore, PsycINFO, Web of Science, and ACM Digital Library) along with connected papers, using Boolean strings linking together AI, chronic disease self-management, trust and emotional intelligence (tailored to each database's individual vernacular). Initial screening of identified articles was completed in two phases using the PRISMA 2020 criteria of pre-determined inclusion and exclusion criteria before critical appraisal using the MMAT v2018 and CASP tools. RESULTS After a systematic selection process of 1,486 studies, 45 studies were finally selected based on inclusion criteria. Four major theme areas emerged from the papers including: Technology in chronic illness self-management, Trust in human-AI interaction, Empathy and emotional intelligence in AI and The ethics, equity and ethical application of AI. The quality appraisal showed that 91% (41/45) of the selected studies were rated as high quality, with an average appraisal score of 93%. In addition, 73% (33/45) of the selected papers were published during 2024 and 2025 thus highlighting a high quality and contemporary compilation of literature on the subject. CONCLUSIONS Despite advances in AI for chronic disease management and in trust-empathy theory, these fields remain siloed. We identify 5 critical research gaps; the lack of a combined trust-empathy model, the under-specification of context awareness, the absence of equity consideration, the exclusion of overtrust consideration, and lack of long-term studies demonstrating effectiveness and safety of emotional AI systems. CLINICALTRIAL Not registered. The review protocol was developed before the search but was not prospectively registered in PROSPERO or an equivalent database. This limitation is discussed in Section 4.4.

This observational and exploratory study describes the agreement between peer assessments and professor assessments in two successive peer assessment workshops involving 96 second-year engineering students at the National School of Applied Sciences of Kenitra, part of Ibn Tofail University in Morocco. Equal-sized groups were formed based on student profiles using clustering and a balancing system. The first strategy aimed to maximise within-group profile similarity to form homogeneous groups by placing students from the same cluster in the same group, while the second aimed to maximise profile diversity to form heterogeneous groups by combining students from different clusters, subject to the available cluster sizes. For each workshop, we compared the average peer grade for each assignment with the professor’s grade using Pearson’s correlation, TOST equivalence tests, and Bland–Altman analysis. In the full samples, the first workshop showed a strong correlation (r = 0.90) and equivalence of means within the predefined ±4-point margin, whereas the second showed a weaker correlation (r = 0.43), a more negative mean bias, wider limits of agreement, and no equivalence of means within this margin. The ±4-point criterion relates to the mean difference and does not imply that acceptable agreement is established for every assignment. A sensitivity analysis restricted to assignments that received exactly four peer ratings supported the equivalence of means observed in both workshops, with mean differences of −0.2 and −0.4 points, respectively, whereas differences in correlation and dispersion persisted. The grouping strategy varied in conjunction with the task, the assessment grid, participation and the number of peer ratings, so its independent contribution cannot be determined. The inferential analyses did not account for dependence arising from shared assessors and groups, making the confidence intervals and conclusions regarding equivalence provisional. Controlled comparisons with simpler grouping methods are needed to determine whether profile-based grouping contributes to the quality of peer assessment.