Research Papers

Evaluating the Critical Risks of Amazon's Nova Premier under the Frontier Model Safety Framework

Satyapriya Krishna, Ninareh Mehrabi, Abhinav Mohanty +4

Nova Premier is Amazon's most capable multimodal foundation model and teacher for model distillation. It processes text, images, and video with a one-million-token context window, enabling analysis of large codebases, 400-page documents, and 90-minute videos in a single prompt. We present the first comprehensive evaluation of Nova Premier's critical risk profile under the Frontier Model Safety Framework. Evaluations target three high-risk domains -- Chemical, Biological, Radiological & Nuclear (CBRN), Offensive Cyber Operations, and Automated AI R&D -- and combine automated benchmarks, expert red-teaming, and uplift studies to determine whether the model exceeds release thresholds. We summarize our methodology and report core findings. Based on this evaluation, we find that Nova Premier is safe for public release as per our commitments made at the 2025 Paris AI Safety Summit. We will continue to enhance our safety evaluation and mitigation pipelines as new risks and capabilities associated with frontier models are identified.

Creating systems that are aligned with our goals is seen as a leading approach to create safe and beneficial AI in both leading AI companies and the academic field of AI safety. We defend the view that misaligned AGI - future, generally intelligent (robotic) AI agents - poses catastrophic risks. At the same time, we support the view that aligned AGI creates a substantial risk of catastrophic misuse by humans. While both risks are severe and stand in tension with one another, we show that - in principle - there is room for alignment approaches which do not increase misuse risk. We then investigate how the tradeoff between misalignment and misuse looks empirically for different technical approaches to AI alignment. Here, we argue that many current alignment techniques and foreseeable improvements thereof plausibly increase risks of catastrophic misuse. Since the impacts of AI depend on the social context, we close by discussing important social factors and suggest that to reduce the risk of a misuse catastrophe due to aligned AGI, techniques such as robustness, AI control methods and especially good governance seem essential.

Multi-criteria Rank-based Aggregation for Explainable AI

Sujoy Chatterjee, Everton Romanzini Colombo, Marcos Medeiros Raimundo

Explainability is crucial for improving the transparency of black-box machine learning models. With the advancement of explanation methods such as LIME and SHAP, various XAI performance metrics have been developed to evaluate the quality of explanations. However, different explainers can provide contrasting explanations for the same prediction, introducing trade-offs across conflicting quality metrics. Although available aggregation approaches improve robustness, reducing explanations' variability, very limited research employed a multi-criteria decision-making approach. To address this gap, this paper introduces a multi-criteria rank-based weighted aggregation method that balances multiple quality metrics simultaneously to produce an ensemble of explanation models. Furthermore, we propose rank-based versions of existing XAI metrics (complexity, faithfulness and stability) to better evaluate ranked feature importance explanations. Extensive experiments on publicly available datasets demonstrate the robustness of the proposed model across these metrics. Comparative analyses of various multi-criteria decision-making and rank aggregation algorithms showed that TOPSIS and WSUM are the best candidates for this use case.

The introduction of the AI Act in the European Union presents the AI research and practice community with a set of new challenges related to compliance. While it is certain that AI practitioners will require additional guidance and tools to meet these requirements, previous research on toolkits that aim to translate the theory of AI ethics into development and deployment practice suggests that such resources suffer from multiple limitations. These limitations stem, in part, from the fact that the toolkits are either produced by industry-based teams or by academics whose work tends to be abstract and divorced from the realities of industry. In this paper, we discuss the challenge of developing an AI ethics toolkit for practitioners that helps them comply with new AI-focused regulation, but that also moves beyond mere compliance to consider broader socio-ethical questions throughout development and deployment. The toolkit was created through a cross-sectoral collaboration between an academic team based in the UK and an industry team in Italy. We outline the background and rationale for creating a pro-justice AI Act compliance toolkit, detail the process undertaken to develop it, and describe the collaboration and negotiation efforts that shaped its creation. We aim for the described process to serve as a blueprint for other teams navigating the challenges of academia-industry partnerships and aspiring to produce usable and meaningful AI ethics resources.

A Participatory Strategy for AI Ethics in Education and Rehabilitation grounded in the Capability Approach

Valeria Cesaroni, Eleonora Pasqua, Piercosma Bisconti +1

AI-based technologies have significant potential to enhance inclusive education and clinical-rehabilitative contexts for children with Special Educational Needs and Disabilities. AI can enhance learning experiences, empower students, and support both teachers and rehabilitators. However, their usage presents challenges that require a systemic-ecological vision, ethical considerations, and participatory research. Therefore, research and technological development must be rooted in a strong ethical-theoretical framework. The Capability Approach - a theoretical model of disability, human vulnerability, and inclusion - offers a more relevant perspective on functionality, effectiveness, and technological adequacy in inclusive learning environments. In this paper, we propose a participatory research strategy with different stakeholders through a case study on the ARTIS Project, which develops an AI-enriched interface to support children with text comprehension difficulties. Our research strategy integrates ethical, educational, clinical, and technological expertise in designing and implementing AI-based technologies for children's learning environments through focus groups and collaborative design sessions. We believe that this holistic approach to AI adoption in education can help bridge the gap between technological innovation and ethical responsibility.

The integration of AI into daily life has generated considerable attention and excitement, while also raising concerns about automating algorithmic harms and re-entrenching existing social inequities. While the responsible deployment of trustworthy AI systems is a worthy goal, there are many possible ways to realize it, from policy and regulation to improved algorithm design and evaluation. In fact, since AI trains on social data, there is even a possibility for everyday users, citizens, or workers to directly steer the AI system's behavior through Algorithmic Collective Action, by deliberately modifying the data they share with a platform to drive its learning process in their favor. This paper considers how these grassroots efforts to influence AI interact with methods used by AI firms and governments to improve model trustworthiness. In particular, we focus on the setting where the AI firm deploys a differentially private model, motivated by the growing regulatory focus on privacy and data protection. We investigate how the use of Differentially Private Stochastic Gradient Descent (DP-SGD) affects the collective's ability to influence the learning process. Our findings show that while differential privacy protects individual data, it introduces challenges for effective algorithmic collective action. We establish this trade-off formally by characterizing lower bounds on the success of algorithmic collective action under differential privacy as a function of the collective's size and the firm's privacy parameters. We then verify these trends experimentally by simulating collective action during the training of deep neural network classifiers across several datasets. Finally, we perform a stylized economic analysis of privacy costs to integrate additional incentives, analyzing how utility and participation costs influence the formation of collectives under private training regimes.

As Natural Language Processing (NLP) models continue to evolve and become integral to high-stakes applications, ensuring their interpretability remains a critical challenge. Given the growing variety of explainability methods and diverse stakeholder requirements, frameworks that help stakeholders select appropriate explanations tailored to their specific use cases are increasingly important. To address this need, we introduce EvalxNLP, a Python framework for benchmarking state-of-the-art feature attribution methods for transformer-based NLP models. EvalxNLP integrates eight widely recognized explainability techniques from the Explainable AI (XAI) literature, enabling users to generate and evaluate explanations based on key properties such as faithfulness, plausibility, and complexity. Our framework also provides interactive, LLM-based textual explanations, facilitating user understanding of the generated explanations and evaluation outcomes. Human evaluation results indicate high user satisfaction with EvalxNLP, suggesting it is a promising framework for benchmarking explanation methods across diverse user groups. By offering a user-friendly and extensible platform, EvalxNLP aims at democratizing explainability tools and supporting the systematic comparison and advancement of XAI techniques in NLP.

Deep Learning (DL) models processing images to recognize the health state of large infrastructure components can exhibit biases and rely on non-causal shortcuts. eXplainable Artificial Intelligence (XAI) can address these issues but manually analyzing explanations generated by XAI techniques is time-consuming and prone to errors. This work proposes a novel framework that combines post-hoc explanations with semi-supervised learning to automatically identify anomalous explanations that deviate from those of correctly classified images and may therefore indicate model abnormal behaviors. This significantly reduces the workload for maintenance decision-makers, who only need to manually reclassify images flagged as having anomalous explanations. The proposed framework is applied to drone-collected images of insulator shells for power grid infrastructure monitoring, considering two different Convolutional Neural Networks (CNNs), GradCAM explanations and Deep Semi-Supervised Anomaly Detection. The average classification accuracy on two faulty classes is improved by 8% and maintenance operators are required to manually reclassify only 15% of the images. We compare the proposed framework with a state-of-the-art approach based on the faithfulness metric: the experimental results obtained demonstrate that the proposed framework consistently achieves F_1 scores larger than those of the faithfulness-based approach. Additionally, the proposed framework successfully identifies correct classifications that result from non-causal shortcuts, such as the presence of ID tags printed on insulator shells.

Over the past fifty years, numerous software defect prediction (SDP) approaches have been proposed. However, the ability to explain why predictors make certain predictions remains limited. Explainable SDP has emerged as a promising solution by using explainable artificial intelligence (XAI) methods to clarify the decision-making processes of predictors. Despite this progress, there is still significant potential to enhance the reliability of existing approaches. To address this limitation, we treat defect prediction and the corresponding interpretation as two distinct but closely related tasks and propose a joint learning framework that allows for the simultaneous training of the predictor and its interpreter. The novelty of our approach lies in two main aspects: 1. We design feedback loops that convey the decision-making logic from the predictor to the interpreter. This ensures a high level of conciseness in decision logic and feature engineering for both the predictor and the interpreter, enabling the interpreter to achieve reliable local and global interpretability. 2. We incorporate the interpretation results as a penalty term in the loss function of the joint-learning framework. This not only improves the accuracy of the predictor but also imposes a stronger constraint on the reliability of the interpreter. We validated our proposed method against several existing explainable SDPs across multiple datasets. The results demonstrate its effectiveness in both interpretation and defect prediction. The source code for the proposed method is available at: https://github.com/BugPredictor/software-defect-prediction.git

Explainable AI is increasingly employing argumentation methods to facilitate interactive explanations between AI agents and human users. While existing approaches typically rely on predetermined human user models, there remains a critical gap in dynamically learning and updating these models during interactions. In this paper, we present a framework that enables AI agents to adapt their understanding of human users through argumentation-based dialogues. Our approach, called Persona, draws on prospect theory and integrates a probability weighting function with a Bayesian belief update mechanism that refines a probability distribution over possible human models based on exchanged arguments. Through empirical evaluations with human users in an applied argumentation setting, we demonstrate that Persona effectively captures evolving human beliefs, facilitates personalized interactions, and outperforms state-of-the-art methods.

Towards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability

Shichang Zhang, Tessa Han, Usha Bhalla +1

The increasing complexity of AI systems has made understanding their behavior critical. Numerous interpretability methods have been developed to attribute model behavior to three key aspects: input features, training data, and internal model components, which emerged from explainable AI, data-centric AI, and mechanistic interpretability, respectively. However, these attribution methods are studied and applied rather independently, resulting in a fragmented landscape of methods and terminology. This position paper argues that feature, data, and component attribution methods share fundamental similarities, and a unified view of them benefits both interpretability and broader AI research. To this end, we first analyze popular methods for these three types of attributions and present a unified view demonstrating that these seemingly distinct methods employ similar techniques (such as perturbations, gradients, and linear approximations) over different aspects and thus differ primarily in their perspectives rather than techniques. Then, we demonstrate how this unified view enhances understanding of existing attribution methods, highlights shared concepts and evaluation criteria among these methods, and leads to new research directions both in interpretability research, by addressing common challenges and facilitating cross-attribution innovation, and in AI more broadly, with applications in model editing, steering, and regulation.

As artificial intelligence (AI) becomes integral to economy and society, communication gaps between developers, users, and stakeholders hinder trust and informed decision-making. High-level AI labels, inspired by frameworks like EU energy labels, have been proposed to make the properties of AI models more transparent. Without requiring deep technical expertise, they can inform on the trade-off between predictive performance and resource efficiency. However, the practical benefits and limitations of AI labeling remain underexplored. This study evaluates AI labeling through qualitative interviews along four key research questions. Based on thematic analysis and inductive coding, we found a broad range of practitioners to be interested in AI labeling (RQ1). They see benefits for alleviating communication gaps and aiding non-expert decision-makers, however limitations, misunderstandings, and suggestions for improvement were also discussed (RQ2). Compared to other reporting formats, interviewees positively evaluated the reduced complexity of labels, increasing overall comprehensibility (RQ3). Trust was influenced most by usability and the credibility of the responsible labeling authority, with mixed preferences for self-certification versus third-party certification (RQ4). Our Insights highlight that AI labels pose a trade-off between simplicity and complexity, which could be resolved by developing customizable and interactive labeling frameworks to address diverse user needs. Transparent labeling of resource efficiency also nudged interviewee priorities towards paying more attention to sustainability aspects during AI development. This study validates AI labels as a valuable tool for enhancing trust and communication in AI, offering actionable guidelines for their refinement and standardization.