Taylor Sorensen, Pushkar Mishra, Roma Patel, Michael Henry Tessler, Michiel A. Bakker, Georgina Evans, Iason Gabriel, Noah Goodman, Verena Rieser. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

How do multi-modal generative models describe images of recent historical events and figures, whose legacies may be nuanced, multifaceted, or contested? This task necessitates not only accurate visual recognition, but also socio-cultural knowledge and cross-modal reasoning. To address this evaluation challenge, we introduce Century -- a novel dataset of sensitive historical images. This dataset consists of 1,500 images from recent history, created through an automated method combining knowledge graphs and language models with quality and diversity criteria created from the practices of museums and digital archives. We demonstrate through automated and human evaluation that this method produces a set of images that depict events and figures that are diverse across topics and represents all regions of the world. We additionally propose an evaluation framework for evaluating the historical contextualisation capabilities along dimensions of accuracy, thoroughness, and objectivity. We demonstrate this approach by using Century to evaluate four foundation models, scoring performance using both automated and human evaluation. We find that historical contextualisation of sensitive images poses a significant challenge for modern multi-modal foundation models, and offer practical recommendations for how developers can use Century to evaluate improvements to models and applications.

The development of highly-capable conversational agents, underwritten by large language models, has the potential to shape user interaction with this technology in profound ways, particularly when the technology is anthropomorphic, or appears human-like. Although the effects of anthropomorphic AI are often benign, anthropomorphic design features also create new kinds of risk. For example, users may form emotional connections to human-like AI, creating the risk of infringing on user privacy and autonomy through over-reliance. To better understand the possible pitfalls of anthropomorphic AI systems, we make two contributions: first, we explore anthropomorphic features that have been embedded in interactive systems in the past, and leverage this precedent to highlight the current implications of anthropomorphic design. Second, we propose research directions for informing the ethical design of anthropomorphic AI. In advancing the responsible development of AI, we promote approaches to the ethical foresight, evaluation, and mitigation of harms arising from user interactions with anthropomorphic AI.

The development of increasingly agentic and human-like AI assistants, capable of performing a wide range of tasks on user's behalf over time, has sparked heightened interest in the nature and bounds of human interactions with AI. Such systems may indeed ground a transition from task-oriented interactions with AI, at discrete time intervals, to ongoing relationships -- where users develop a deeper sense of connection with and attachment to the technology. This paper investigates what it means for relationships between users and advanced AI assistants to be appropriate and proposes a new framework to evaluate both users' relationships with AI and developers' design choices. We first provide an account of advanced AI assistants, motivating the question of appropriate relationships by exploring several distinctive features of this technology. These include anthropomorphic cues and the longevity of interactions with users, increased AI agency, generality and context ambiguity, and the forms and depth of dependence the relationship could engender. Drawing upon various ethical traditions, we then consider a series of values, including benefit, flourishing, autonomy and care, that characterise appropriate human interpersonal relationships. These values guide our analysis of how the distinctive features of AI assistants may give rise to inappropriate relationships with users. Specifically, we discuss a set of concrete risks arising from user--AI assistant relationships that: (1) cause direct emotional or physical harm to users, (2) limit opportunities for user personal development, (3) exploit user emotional dependence, and (4) generate material dependencies without adequate commitment to user needs. We conclude with a set of recommendations to address these risks.

Generative AI systems produce a range of ethical and social risks. Evaluation of these risks is a critical step on the path to ensuring the safety of these systems. However, evaluation requires the availability of validated and established measurement approaches and tools. In this paper, we provide an empirical review of the methods and tools that are available for evaluating known safety of generative AI systems to date. To this end, we review more than 200 safety-related evaluations that have been applied to generative AI systems. We categorise each evaluation along multiple axes to create a detailed snapshot of the safety evaluation landscape to date. We release this data for researchers and AI safety practitioners (https://bitly.ws/3hUzu). Analysing the current safety evaluation landscape reveals three systemic ”evaluation gaps”. First, a ”modality gap” emerges as few safety evaluations exist for non-text modalities. Second, a ”risk coverage gap” arises as evaluations for several ethical and social risks are simply lacking. Third, a ”context gap” arises as most safety evaluations are model-centric and fail to take into account the broader context in which AI systems operate. Devising next steps for safety practitioners based on these findings, we present tactical ”low-hanging fruit” steps towards closing the identified evaluation gaps and their limitations. We close by discussing the role and limitations of safety evaluation to ensure the safety of generative AI systems.

Large language models produce human-like text that drive a growing number of applications. However, recent literature and, increasingly, real world observations, have demonstrated that these models can generate language that is toxic, biased, untruthful or otherwise harmful. Though work to evaluate language model harms is under way, translating foresight about which harms may arise into rigorous benchmarks is not straightforward. To facilitate this translation, we outline six ways of characterizing harmful text which merit explicit consideration when designing new benchmarks. We then use these characteristics as a lens to identify trends and gaps in existing benchmarks. Finally, we apply them in a case study of the Perspective API, a toxicity classifier that is widely used in harm benchmarks. Our characteristics provide one piece of the bridge that translates between foresight and effective evaluation.