>>our publications

Abstract: Text-to-image (TTI) diffusion models have achieved remarkable visual quality, yet they have been repeatedly shown to exhibit social biases across sensitive attributes such as gender, race and age. To mitigate these biases, existing approaches frequently depend on curated prompt datasets - either manually constructed or generated with large language models (LLMs) - as part of their training and/or evaluation procedures. Beside the curation cost, this also risks overlooking unanticipated, less obvious prompts that trigger biased generation, even in models that have undergone debiasing. In this work, we introduce Bias-Guided Prompt Search (BGPS), a framework that automatically generates prompts that aim to maximize the presence of biases in the resulting images. BGPS comprises two components: (1) an LLM instructed to produce attribute-neutral prompts and (2) attribute classifiers acting on the TTI’s internal representations that steer the decoding process of the LLM toward regions of the prompt space that amplify the image attributes of interest. We conduct extensive experiments on Stable Diffusion 1.5 and a state-of-the-art debiased model and discover an array of subtle and previously undocumented biases that severely deteriorate fairness metrics. Crucially, the discovered prompts are interpretable, i.e they may be entered by a typical user, quantitatively improving the perplexity metric compared to a prominent hard prompt optimization counterpart. Our findings uncover TTI vulnerabilities, while BGPS expands the bias search space and can act as a new evaluation tool for bias mitigation.

Abstract: As machine learning models are increasingly deployed in high-stakes settings, e.g. as decision support systems in various societal sectors or in critical infrastructure, designers and auditors are facing the need to ensure that models satisfy a wider variety of requirements (e.g. compliance with regulations, fairness, computational constraints) beyond performance. Although most of them are the subject of ongoing studies, typical approaches face critical challenges: post-processing methods tend to compromise performance, which is often counteracted by fine-tuning or, worse, training from scratch, an often time-consuming or even unavailable strategy. This raises the following question: "Can we efficiently edit models to satisfy requirements, without sacrificing their utility?" In this work, we approach this with a unifying framework, in a data-driven manner, i.e. we learn to edit neural networks (NNs), where the editor is an NN itself - a graph metanetwork - and editing amounts to a single inference step. In particular, the metanetwork is trained on NN populations to minimise an objective consisting of two terms: the requirement to be enforced and the preservation of the NN's utility. We experiment with diverse tasks (the data minimisation principle, bias mitigation and weight pruning) improving the trade-offs between performance, requirement satisfaction and time efficiency compared to popular post-processing or re-training alternatives.

Abstract: The use of Generative AI (GenAI) in developing large brainwave foundation models for Brain-Computer Interfaces (BCIs) offers enormous potential but also comes with several key safety and ethical concerns. This work identifies these challenges and highlights cases of potential misuse of GenAI in BCIs, including synthetic neural activity, behaviour profiling, privacy and equality risks. Finally, it emphasizes the importance of essential safeguarding techniques to mitigate these risks, such as innovative technological solutions and proper regulatory and ethical frameworks.

Abstract: Over the past five years, innovative AI systems, including applications based on Generative AI and Large Language Models (LLMs), have emerged. Text, image, video, and symbol manipulation are now within the capabilities of AI. Generative AI is already present in contemporary newsrooms where the deployment of AI is relevant, as LLMs excel at text manipulation. Unlike the traditional top-down approach to introducing new technologies, this deployment involves a complex, multifaceted, and under-researched procedure. This paper draws from labour process theory, the social construction of technology (SCOT) paradigm, and literature on the quantitative shift in journalism and the introduction of algorithms into newswork, while situating the examination of Generative AI within the Greek media system. This research investigates the integration of Generative AI in newsrooms through semi-structured interviews with journalists. The goal is to capture insights into their experiences, challenges, and perceptions regarding the introduction of Generative AI in newswork. Findings indicate that Greek journalists, drawing from their experiences with the existing labour process and the economic dynamics of newsrooms, have significant concerns regarding Generative AI while simultaneously being sceptical of news organisations’ capabilities to integrate new technologies in a manner that respects journalists’ status.

Abstract: This paper explores the interplay between AI metrics and policymaking by examining the conceptual and methodological frameworks of global AI metrics and their alignment with National Artificial Intelligence Strategies (NAIS). Through topic modeling and qualitative content analysis, key thematic areas in NAIS are identified. The findings suggest a misalignment between the technical and economic focus of global AI metrics and the broader societal and ethical priorities emphasized in NAIS. This highlights the need to recalibrate AI evaluation frameworks to include ethical and other social considerations, aligning AI advancements with the United Nations Sustainable Development Goals (SDGs) for an inclusive, ethical, and sustainable future.

Abstract: In the face of what is called “Existential risks/threats” or “polycrisis” – where digitalization, biomedicalization, and environmental degradation intertwine as existential threats – this article argues for a critical philosophy of technology as a necessary framework for ethical engagement. To address this polycrisis, we elaborate on an approach that integrates ethics and technology through serious gaming. Specifically, we introduce Tethics, a board game designed to engage players in ethical reflection on technological governance. By embedding critical philosophical insights into interactive gameplay, Tethics fosters a dynamic, experiential understanding of ethical dilemmas. Drawing on Andrew Feenberg’s critical theory of technology, we outline how game design can serve as a method for both refining philosophical inquiry and fostering public engagement. We argue that serious games may offer a unique means of interrogating and reconfiguring our relationship with technology, or can successfully act as educational tools for the ethics of technology today. Finally, we argue that Tethics performs a philosophically therapeutic function, debunking the metaphysical underpinnings of an instrumentalist and determinist viewpoint that dominates the current understanding of technology.

Abstract: We recommend enabling inclusive urban transport planning through civic artificial intelligence. To achieve this policy recommendation, we propose the following: (1) Encourage and provide resources for experimentation with new technologies that enable local community participation in urban transport planning; (2) Recognize the potential of Artificial Intelligence (AI) to assist in complex urban transport planning decisions; (3) Acknowledge that AI is embedded in society, instead of treating it as a neutral technology; and (4) Foster community engagement in transport planning and evaluation, via a Civic AI framework that directly integrates preferences and feedback into planning.

Abstract: The prospect of the use of Large Language Models, like ChatGPT, in work environments raises important questions regarding both the potential for a dramatic change in the quality of jobs and the risk of unemployment. The answers to these questions, but, also, the posing of questions to be answered, may involve the use of ChatGPT. This, in turn, may give rise to a series of ethical considerations. The article seeks to identify such considerations by presenting a research on a questionnaire that was developed by means of ChatGPT before it was answered, first, by a group of humans (H) and, then, through the use of a machine (M), ChatGPT. The language model was actually used to respond to the questionnaire twice. First, based on its data (M1), and, second, based on it being asked to imitate a human (M2). Based on the significant differences between the H and M answers, and, further, on the noticeable differences occurring within the M answers (the differences between the M1 and M2 answers), the article concludes by registering a cluster of three ethical considerations.

Abstract. This article provides an intellectual history of artif icial intelligence in the
electronic era of computing, that is, from the postwar decades to the present. We
argue for the existence of two periods; a f irst period, def ined by the discourse of a
post-industrial society and an information age, and a second one, characterised
by the discourse of a fourth industrial revolution. Discourses of a post-industrial
society and a fourth industrial revolution are constitutively related to discourses
of computer automation, which, in turn, are def ined by artif icial intelligence.
This paper provides a canvas of an intellectual history of artif icial intelligence
in the electronic era through the examination of discourses of this period on
computer automation.

Abstract: The development and rapid integration of digital technologies into everyday activities is often accompanied by narratives about the impact of algorithms on social reality. In recent years, concerns have grown regarding the future of algorithmic governance, as the expansion of automated decision-making systems in education, work, and institutional processes has exposed the racial, gender, and class biases embedded in the design and data of these algorithms. This article maps the related public discourse and examines the algorithmic imaginaries that characterize different fields of public debate and academic research, with the aim of reconstructing the critical discourse surrounding algorithms. In an effort to move beyond narratives of technophobia and technophilia, the article places scientific communication at the center of the discussion and highlights the role it can play in fostering dialogue about the digital transformation of society.

Abstract: Deep learning has catalysed progress in tasks such as face recognition and analysis, leading to a quick integration of technological solutions in multiple layers of our society. While such systems have proven to be accurate by standard evaluation metrics and benchmarks, a surge of work has recently exposed the demographic bias that such algorithms exhibit–highlighting that accuracy does not entail fairness. Clearly, deploying biased systems under real-world settings can have grave consequences for affected populations. Indeed, learning methods are prone to inheriting, or even amplifying the bias present in a training set, manifested by uneven representation across demographic groups. In facial datasets, this particularly relates to attributes such as skin tone, gender, and age. In this work, we address the problem of mitigating bias in facial datasets by data augmentation. We propose a multi-attribute framework that can successfully transfer complex, multi-scale facial patterns even if these belong to underrepresented groups in the training set. This is achieved by relaxing the rigid dependence on a single attribute label, and further introducing a tensor-based mixing structure that captures multiplicative interactions between attributes in a multilinear fashion. We evaluate our method with an extensive set of qualitative and quantitative experiments on several datasets, with rigorous comparisons to state-of-the-art methods. We find that the proposed framework can successfully mitigate dataset bias, as evinced by extensive evaluations on established diversity metrics, while significantly improving fairness metrics such as equality of opportunity.

Abstract: Deep learning-based methods have pushed the limits of the state-of-the-art in face analysis. However, despite their success, these models have raised concerns regarding their bias towards certain demographics. This bias is inflicted both by limited diversity across demographics in the training set, as well as the design of the algorithms. In this work, we investigate the demographic bias of deep learning models in face recognition, age estimation, gender recognition and kinship verification. To this end, we introduce the most comprehensive, large-scale dataset of facial images and videos to date. It consists of 40K still images and 44K sequences (14.5M video frames in total) captured in unconstrained, real-world conditions from 1,045 subjects. The data are manually annotated in terms of identity, exact age, gender and kinship. The performance of state-of-the-art models is scrutinized and demographic bias is exposed by conducting a series of experiments. Lastly, a method to debias network embeddings is introduced and tested on the proposed benchmarks.

Abstract: The article introduces the essentialist 1940s demarcation between a digital-superior and an analog-inferior computer as a key moment in severing the computing machine from the human worker labouring with it, and, accordingly, to keeping the history of computing technology and the history of computing labour apart. It starts with a section that further argues that the introduction of this key demarcation is strongly linked to the transition from the prewar use of the concept “computer”, which referred to a human worker, to its postwar use, which refers to a computing machine. The argument comes full circle by connecting the concealed analog to hidden female computing labour, a connection suggested by a revisiting of the paradigmatic display of the ENIAC as a digital machine. There follow two sections, one on the history of the female labour concealed by presenting the digital computer as superior and the other on the history of the male labour neglected by ignoring the analog computer as inferior.

Abstract: This paper is an attempt at a critical introduction to the pursuit of constructing machines capable of artificial intelligence, thinking machines. This has defined modernity, historical capitalism, through an endless introduction and use of computing tools and mechanisms (merchant capitalism) and computing machines (industrial capitalism), mechanical, electrical, and, since the 1940s, also, electronic. It is popularly assumed that presenting machines as capable of artificial intelligence is a rather recent phenomenon. The examples offered here suggest that computing artifacts were ideologized as intelligent from the beginning of capitalism. Attention is invited to both offline and online computers, going back to mechanical calculators and steam engine governors respectively. The case of the steam engine governor, an online analog of the steam engine that was connected to the circuit of the steam engine in a ‘negative feedback’ manner, is used to introduce to the limits of artificial intelligence, as these are linked to the unprecedented environmental crisis that all creatures of nature are now living with.