Research Papers

Modern software relies heavily on data and machine learning, and affects decisions that shape our world. Unfortunately, recent studies have shown that because of biases in data, software systems frequently inject bias into their decisions, from producing better closed caption transcriptions of men's voices than of women's voices to overcharging people of color for financial loans. To address bias in machine learning, data scientists need tools that help them understand the trade-offs between model quality and fairness in their specific data domains. Toward that end, we present fairkit-learn, a toolkit for helping data scientists reason about and understand fairness. Fairkit-learn works with state-of-the-art machine learning tools and uses the same interfaces to ease adoption. It can evaluate thousands of models produced by multiple machine learning algorithms, hyperparameters, and data permutations, and compute and visualize a small Pareto-optimal set of models that describe the optimal trade-offs between fairness and quality. We evaluate fairkit-learn via a user study with 54 students, showing that students using fairkit-learn produce models that provide a better balance between fairness and quality than students using scikit-learn and IBM AI Fairness 360 toolkits. With fairkit-learn, users can select models that are up to 67% more fair and 10% more accurate than the models they are likely to train with scikit-learn.

Open Questions in Creating Safe Open-ended AI: Tensions Between Control and Creativity

Adrien Ecoffet, Jeff Clune, Joel Lehman

Artificial life originated and has long studied the topic of open-ended evolution, which seeks the principles underlying artificial systems that innovate continually, inspired by biological evolution. Recently, interest has grown within the broader field of AI in a generalization of open-ended evolution, here called open-ended search, wherein such questions of open-endedness are explored for advancing AI, whatever the nature of the underlying search algorithm (e.g. evolutionary or gradient-based). For example, open-ended search might design new architectures for neural networks, new reinforcement learning algorithms, or most ambitiously, aim at designing artificial general intelligence. This paper proposes that open-ended evolution and artificial life have much to contribute towards the understanding of open-ended AI, focusing here in particular on the safety of open-ended search. The idea is that AI systems are increasingly applied in the real world, often producing unintended harms in the process, which motivates the growing field of AI safety. This paper argues that open-ended AI has its own safety challenges, in particular, whether the creativity of open-ended systems can be productively and predictably controlled. This paper explains how unique safety problems manifest in open-ended search, and suggests concrete contributions and research questions to explore them. The hope is to inspire progress towards creative, useful, and safe open-ended search algorithms.

Recommendation systems are an integral part of Artificial Intelligence (AI) and have become increasingly important in the growing age of commercialization in AI. Deep learning (DL) techniques for recommendation systems (RS) provide powerful latent-feature models for effective recommendation but suffer from the major drawback of being non-interpretable. In this paper we describe a framework for explainable temporal recommendations in a DL model. We consider an LSTM based Recurrent Neural Network (RNN) architecture for recommendation and a neighbourhood-based scheme for generating explanations in the model. We demonstrate the effectiveness of our approach through experiments on the Netflix dataset by jointly optimizing for both prediction accuracy and explainability.

On the Origin of Samples: Attribution of Output to a Particular Algorithm

Roman V. Yampolskiy

With unprecedented advances in genetic engineering we are starting to see progressively more original examples of synthetic life. As such organisms become more common it is desirable to be able to distinguish between natural and artificial life forms. In this paper, we present this challenge as a generalized version of Darwin's original problem, which he so brilliantly addressed in On the Origin of Species. After formalizing the problem of determining origin of samples we demonstrate that the problem is in fact unsolvable, in the general case, if computational resources of considered originator algorithms have not been limited and priors for such algorithms are known to be equal. Our results should be of interest to astrobiologists and scientists interested in producing a more complete theory of life, as well as to AI-Safety researchers.

The rapid advancement of Generative Artificial Intelligence (GenAI) has introduced new opportunities for transforming higher education, particularly in fields that require analytical reasoning and regulatory compliance, such as cybersecurity management. This study presents a structured framework for integrating GenAI tools into cybersecurity education, demonstrating their role in fostering critical thinking, real-world problem-solving, and regulatory awareness. The implementation strategy followed a two-stage approach, embedding GenAI within tutorial exercises and assessment tasks. Tutorials enabled students to generate, critique, and refine AI-assisted cybersecurity policies, while assessments required them to apply AI-generated outputs to real-world scenarios, ensuring alignment with industry standards and regulatory requirements. Findings indicate that AI-assisted learning significantly enhanced students' ability to evaluate security policies, refine risk assessments, and bridge theoretical knowledge with practical application. Student reflections and instructor observations revealed improvements in analytical engagement, yet challenges emerged regarding AI over-reliance, variability in AI literacy, and the contextual limitations of AI-generated content. Through structured intervention and research-driven refinement, students were able to recognize AI strengths as a generative tool while acknowledging its need for human oversight. This study further highlights the broader implications of AI adoption in cybersecurity education, emphasizing the necessity of balancing automation with expert judgment to cultivate industry-ready professionals. Future research should explore the long-term impact of AI-driven learning on cybersecurity competency, as well as the potential for adaptive AI-assisted assessments to further personalize and enhance educational outcomes.

Generative AI is rapidly moving from research to deployment, elevating the need for responsible development, evaluation, and governance. We conduct a PRISMA guided review of 232 studies (November 2022 - December 2025), spanning large language models, vision language models, diffusion models, and agentic pipelines. We make four contributions: (1) the first survey bridging governance principles, technical evaluation, and domain deployment across all four system types; (2) a ten-criterion rubric (C1-C10) scoring major AI safety benchmarks on risk-surface coverage, paired with a policy crosswalk mapping benchmarks to regulatory requirements; (3) twelve lifecycle KPIs, explainability guidance for foundation models, and a testbed catalogue; and (4) domain-specific analysis across healthcare, finance, education, arts, agriculture, and defense. Three findings emerge: benchmark coverage is dense for bias and toxicity but sparse for privacy, provenance, deepfakes, and system-level failures in agentic settings; evaluations remain largely static and task local, limiting audit portability; and inconsistent documentation complicates cross-release comparison. We outline a research agenda prioritizing adaptive multimodal evaluation, privacy and provenance testing, deepfake risk assessment, calibration reporting, versioned artifacts, and continuous monitoring. This survey offers a structured path to align generative AI evaluation with governance needs for safe and accountable deployment.

The many initiatives on trustworthy AI result in a confusing and multipolar landscape that organizations operating within the fluid and complex international value chains must navigate in pursuing trustworthy AI. The EU's AI Act will now shift the focus of such organizations toward conformance with the technical requirements for regulatory compliance, for which the Act relies on Harmonized Standards. Though a high-level mapping to the Act's requirements will be part of such harmonization, determining the degree to which standards conformity delivers regulatory compliance with the AI Act remains a complex challenge. Variance and gaps in the definitions of concepts and how they are used in requirements between the Act and harmonized standards may impact the consistency of compliance claims across organizations, sectors, and applications. This may present regulatory uncertainty, especially for SMEs and public sector bodies relying on standards conformance rather than proprietary equivalents for developing and deploying compliant high-risk AI systems. To address this challenge, this paper offers a simple and repeatable mechanism for mapping the terms and requirements relevant to normative statements in regulations and standards, e.g., AI Act and ISO management system standards, texts into open knowledge graphs. This representation is used to assess the adequacy of standards conformance to regulatory compliance and thereby provide a basis for identifying areas where further technical consensus development in trustworthy AI value chains is required to achieve regulatory compliance.

Hide and Seek: Fingerprinting Large Language Models with Evolutionary Learning

Dmitri Iourovitski, Sanat Sharma, Rakshak Talwar

As content generated by Large Language Model (LLM) has grown exponentially, the ability to accurately identify and fingerprint such text has become increasingly crucial. In this work, we introduce a novel black-box approach for fingerprinting LLMs, achieving an impressive 72% accuracy in identifying the correct family of models (Such as Llama, Mistral, Gemma, etc) among a lineup of LLMs. We present an evolutionary strategy that leverages the capabilities of one LLM to discover the most salient features for identifying other LLMs. Our method employs a unique "Hide and Seek" algorithm, where an Auditor LLM generates discriminative prompts, and a Detective LLM analyzes the responses to fingerprint the target models. This approach not only demonstrates the feasibility of LLM-driven model identification but also reveals insights into the semantic manifolds of different LLM families. By iteratively refining prompts through in-context learning, our system uncovers subtle distinctions between model outputs, providing a powerful tool for LLM analysis and verification. This research opens new avenues for understanding LLM behavior and has significant implications for model attribution, security, and the broader field of AI transparency.

Recent research in explainability has given rise to numerous post-hoc attribution methods aimed at enhancing our comprehension of the outputs of black-box machine learning models. However, evaluating the quality of explanations lacks a cohesive approach and a consensus on the methodology for deriving quantitative metrics that gauge the efficacy of explainability post-hoc attribution methods. Furthermore, with the development of increasingly complex deep learning models for diverse data applications, the need for a reliable way of measuring the quality and correctness of explanations is becoming critical. We address this by proposing BEExAI, a benchmark tool that allows large-scale comparison of different post-hoc XAI methods, employing a set of selected evaluation metrics.

Artificial intelligence (AI) is currently based largely on black-box machine learning models which lack interpretability. The field of eXplainable AI (XAI) strives to address this major concern, being critical in high-stakes areas such as the finance, legal and health sectors. We present an approach to defining AI models and their interpretability based on category theory. For this we employ the notion of a compositional model, which sees a model in terms of formal string diagrams which capture its abstract structure together with its concrete implementation. This comprehensive view incorporates deterministic, probabilistic and quantum models. We compare a wide range of AI models as compositional models, including linear and rule-based models, (recurrent) neural networks, transformers, VAEs, and causal and DisCoCirc models. Next we give a definition of interpretation of a model in terms of its compositional structure, demonstrating how to analyse the interpretability of a model, and using this to clarify common themes in XAI. We find that what makes the standard 'intrinsically interpretable' models so transparent is brought out most clearly diagrammatically. This leads us to the more general notion of compositionally-interpretable (CI) models, which additionally include, for instance, causal, conceptual space, and DisCoCirc models. We next demonstrate the explainability benefits of CI models. Firstly, their compositional structure may allow the computation of other quantities of interest, and may facilitate inference from the model to the modelled phenomenon by matching its structure. Secondly, they allow for diagrammatic explanations for their behaviour, based on influence constraints, diagram surgery and rewrite explanations. Finally, we discuss many future directions for the approach, raising the question of how to learn such meaningfully structured models in practice.

Why do explanations fail? A typology and discussion on failures in XAI

Clara Bove, Thibault Laugel, Marie-Jeanne Lesot +2

As Machine Learning models achieve unprecedented levels of performance, the XAI domain aims at making these models understandable by presenting end-users with intelligible explanations. Yet, some existing XAI approaches fail to meet expectations: several issues have been reported in the literature, generally pointing out either technical limitations or misinterpretations by users. In this paper, we argue that the resulting harms arise from a complex overlap of multiple failures in XAI, which existing ad-hoc studies fail to capture. This work therefore advocates for a holistic perspective, presenting a systematic investigation of limitations of current XAI methods and their impact on the interpretation of explanations. % By distinguishing between system-specific and user-specific failures, we propose a typological framework that helps revealing the nuanced complexities of explanation failures. Leveraging this typology, we discuss some research directions to help practitioners better understand the limitations of XAI systems and enhance the quality of ML explanations.

Deepfakes represent one of the toughest challenges in the world of Cybersecurity and Digital Forensics, especially considering the high-quality results obtained with recent generative AI-based solutions. Almost all generative models leave unique traces in synthetic data that, if analyzed and identified in detail, can be exploited to improve the generalization limitations of existing deepfake detectors. In this paper we analyzed deepfake images in the frequency domain generated by both GAN and Diffusion Model engines, examining in detail the underlying statistical distribution of Discrete Cosine Transform (DCT) coefficients. Recognizing that not all coefficients contribute equally to image detection, we hypothesize the existence of a unique ``discriminative fingerprint", embedded in specific combinations of coefficients. To identify them, Machine Learning classifiers were trained on various combinations of coefficients. In addition, the Explainable AI (XAI) LIME algorithm was used to search for intrinsic discriminative combinations of coefficients. Finally, we performed a robustness test to analyze the persistence of traces by applying JPEG compression. The experimental results reveal the existence of traces left by the generative models that are more discriminative and persistent at JPEG attacks. Code and dataset are available at https://github.com/opontorno/dcts_analysis_deepfakes.