Author(s): Iason Gabriel, Geoff Keeling, Arianna Manzini & James Evans

Abstract: The deployment of capable AI agents raises fresh questions about safety, human-machine relationships and social coordination. We argue for greater engagement by scientists, scholars, engineers and policymakers with the implications of a world increasingly populated by AI agents. We explore key challenges that must be addressed to ensure that interactions between humans and agents, and among agents themselves, remain broadly beneficial.

Author(s): Iason Gabriel, Geoff Keeling

Abstract: The normative challenge of AI alignment centres upon what goals or values ought to be encoded in AI systems to govern their behaviour. A number of answers have been proposed, including the notion that AI must be aligned with human intentions or that it should aim to be helpful, honest and harmless. Nonetheless, both accounts suffer from critical weaknesses. On the one hand, they are incomplete: neither specification provides adequate guidance to AI systems, deployed across various domains with multiple parties. On the other hand, the justification for these approaches is questionable and, we argue, of the wrong kind. More specifically, neither approach takes seriously the need to justify the operation of AI systems to those affected by their actions – or what this means for pluralistic societies where people have different underlying beliefs about value. To address these limitations, we propose an alternative account of AI alignment that focuses on fair processes. We argue that principles that are the product of these processes are the appropriate target for alignment. This approach can meet the necessary standard of public justification, generate a fuller set of principles for AI that are sensitive to variation in context, and has explanatory power insofar as it makes sense of our intuitions about AI systems and points to a number of hitherto underappreciated ways in which an AI system may cease to be aligned.

Author(s): Hannah Rose Kirk, Iason Gabriel, Chris Summerfield, Bertie Vidgen & Scott A. Hale

Abstract: Humans strive to design safe AI systems that align with our goals and remain under our control. However, as AI capabilities advance, we face a new challenge: the emergence of deeper, more persistent relationships between humans and AI systems. We explore how increasingly capable AI agents may generate the perception of deeper relationships with users, especially as AI becomes more personalised and agentic. This shift, from transactional interaction to ongoing sustained social engagement with AI, necessitates a new focus on socioaffective alignment—how an AI system behaves within the social and psychological ecosystem co-created with its user, where preferences and perceptions evolve through mutual influence. Addressing these dynamics involves resolving key intrapersonal dilemmas, including balancing immediate versus long-term well-being, protecting autonomy, and managing AI companionship alongside the desire to preserve human social bonds. By framing these challenges through a notion of basic psychological needs, we seek AI systems that support, rather than exploit, our fundamental nature as social and emotional beings.

Author(s): Stevie Bergman, Nahema Marchal, John Mellor, Shakir Mohamed, Iason Gabriel, William Isaac

Abstract: Value alignment, the process of ensuring that artificial intelligence (AI) systems are aligned with human values and goals, is a critical issue in AI research. Existing scholarship has mainly studied how to encode moral values into agents to guide their behaviour. Less attention has been given to the normative questions of whose values and norms AI systems should be aligned with, and how these choices should be made. To tackle these questions, this paper presents the STELA process (SocioTEchnical Language agent Alignment), a methodology resting on sociotechnical traditions of participatory, inclusive, and community-centred processes. For STELA, we conduct a series of deliberative discussions with four historically underrepresented groups in the United States in order to understand their diverse priorities and concerns when interacting with AI systems. The results of our research suggest that community-centred deliberation on the outputs of large language models is a valuable tool for eliciting latent normative perspectives directly from differently situated groups. In addition to having the potential to engender an inclusive process that is robust to the needs of communities, this methodology can provide rich contextual insights for AI alignment.

Author(s): Iason Gabriel, Arianna Manzini, Geoff Keeling, Lisa Anne Hendricks, Verena Rieser, Hasan Iqbal, Nenad Tomašev, Ira Ktena, Zachary Kenton, Mikel Rodriguez, Seliem El-Sayed, Sasha Brown, Canfer Akbulut, Andrew Trask, Edward Hughes, A Stevie Bergman, Renee Shelby, Nahema Marchal, Conor Griffin, Juan Mateos-Garcia, Laura Weidinger, Winnie Street, Benjamin Lange, Alex Ingerman, Alison Lentz, Reed Enger, Andrew Barakat, Victoria Krakovna, John Oliver Siy, Zeb Kurth-Nelson, Amanda McCroskery, Vijay Bolina, Harry Law, Murray Shanahan, Lize Alberts, Borja Balle, Sarah de Haas, Yetunde Ibitoye, Allan Dafoe, Beth Goldberg, Sébastien Krier, Alexander Reese, Sims Witherspoon, Will Hawkins, Maribeth Rauh, Don Wallace, Matija Franklin, Josh A Goldstein, Joel Lehman, Michael Klenk, Shannon Vallor, Courtney Biles, Meredith Ringel Morris, Helen King, William Isaac, James Manyika

Abstract: This paper focuses on the opportunities and the ethical and societal risks posed by advanced AI assistants. We define advanced AI assistants as artificial agents with natural language interfaces, whose function is to plan and execute sequences of actions on behalf of a user, across one or more domains, in line with the user's expectations. The paper starts by considering the technology itself, providing an overview of AI assistants, their technical foundations and potential range of applications. It then explores questions around AI value alignment, well-being, safety and malicious uses. Extending the circle of inquiry further, we next consider the relationship between advanced AI assistants and individual users in more detail, exploring topics such as manipulation and persuasion, anthropomorphism, appropriate relationships, trust and privacy. With this analysis in place, we consider the deployment of advanced assistants at a societal scale, focusing on cooperation, equity and access, misinformation, economic impact, the environment and how best to evaluate advanced AI assistants. Finally, we conclude by providing a range of recommendations for researchers, developers, policymakers and public stakeholders.

Author(s): Laura Weidinger, Kevin R McKee, Richard Everett, Saffron Huang, Tina O Zhu, Martin J Chadwick, Christopher Summerfield, Iason Gabriel

Abstract: The philosopher John Rawls proposed the Veil of Ignorance (VoI) as a thought experiment to identify fair principles for governing a society. Here, we apply the VoI to an important governance domain: artificial intelligence (AI). In five incentive-compatible studies (N = 2, 508), including two preregistered protocols, participants choose principles to govern an Artificial Intelligence (AI) assistant from behind the veil: that is, without knowledge of their own relative position in the group. Compared to participants who have this information, we find a consistent preference for a principle that instructs the AI assistant to prioritize the worst-off. Neither risk attitudes nor political preferences adequately explain these choices. Instead, they appear to be driven by elevated concerns about fairness: Without prompting, participants who reason behind the VoI more frequently explain their choice in terms of fairness, compared to those in the Control condition. Moreover, we find initial support for the ability of the VoI to elicit more robust preferences: In the studies presented here, the VoI increases the likelihood of participants continuing to endorse their initial choice in a subsequent round where they know how they will be affected by the AI intervention and have a self-interested motivation to change their mind. These results emerge in both a descriptive and an immersive game. Our findings suggest that the VoI may be a suitable mechanism for selecting distributive principles to govern AI.

Author(s): A Stevie Bergman, Lisa Anne Hendricks, Maribeth Rauh, Boxi Wu, William Agnew, Markus Kunesch, Isabella Duan, Iason Gabriel, William Isaac

Abstract: Calls for representation in artificial intelligence (AI) and machine learning (ML) are widespread, with "representation" or "representativeness" generally understood to be both an instrumentally and intrinsically beneficial quality of an AI system, and central to fairness concerns. But what does it mean for an AI system to be "representative"? Each element of the AI lifecycle is geared towards its own goals and effect on the system, therefore requiring its own analyses with regard to what kind of representation is best. In this work we untangle the benefits of representation in AI evaluations to develop a framework to guide an AI practitioner or auditor towards the creation of representative ML evaluations. Representation, however, is not a panacea. We further lay out the limitations and tensions of instrumentally representative datasets, such as the necessity of data existence and access, surveillance vs expectations of privacy, implications for foundation models and power. This work sets the stage for a research agenda on representation in AI, which extends beyond instrumentally valuable representation in evaluations towards refocusing on, and empowering, impacted communities.

Author(s): A Kasirzadeh, Iason Gabriel

Abstract: Large-scale language technologies are increasingly used in various forms of communication with humans across different contexts. One particular use case for these technologies is conversational agents, which output natural language text in response to prompts and queries. This mode of engagement raises a number of social and ethical questions. \r\n\r\nFor example, what does it mean to align conversational agents with human norms or values? Which norms or values should they be aligned with? And how can this be accomplished? In this paper, we propose a number of steps that help answer these questions. We start by developing a philosophical analysis of the building blocks of linguistic communication between conversational agents and human interlocutors. We then use this analysis to identify and formulate ideal norms of conversation that can govern successful linguistic communication between humans and conversational agents. Furthermore, we explore how these norms can be used to align conversational agents with human values across a range of different discursive domains.\r\n

Author(s): Iason Gabriel, V Ghazavi

Abstract: This paper addresses the question of how to align AI systems with human values and situates it within a wider body of thought regarding technology and value. Far from existing in a vacuum, there has long been an interest in the ability of technology to 'lock-in' different value systems. There has also been considerable thought about how to align technologies with specific social values, including through participatory design-processes. In this paper we look more closely at the question of AI value alignment and suggest that the power and autonomy of AI systems gives rise to opportunities and challenges in the domain of value that have not been encountered before. Drawing important continuities between the work of the fairness, accountability, transparency and ethics community, and work being done by technical AI safety researchers, we suggest that more attention needs to be paid to the question of 'social value alignment' - that is, how to align AI systems with the plurality of values endorsed by groups of people, especially on the global level.

Author(s): A Birhane, W Isaac, V Prabhakaran, M Díaz, MC Elish, I Gabriel, S Mohammed

Abstract: Participatory approaches to artificial intelligence are gaining momentum with the view that participation opens the gateway to an inclusive, equitable, robust, responsible and trustworthy AI. Indeed, these approaches are essential to understanding and adequately representing the needs, desires and perspectives of historically marginalized communities. However, there is also a lack of clarity about what meaningful participation entails and what it is expected to do in the context of AI. \r\n\r\nThis paper reviews participatory approaches across varied historical contexts as well as participatory methods and practices within the AI pipeline. We then examine three case studies in participatory AI. Ultimately, participation supports beneficial, emancipatory and empowering technology design, only when it avoids cooptation, power asymmetries and conflation with other activities.\r\n

Author(s): V Prabhakaran, M Mitchell, T Gebru, I Gabriel

Abstract: This paper explores the relationship between artificial intelligence and human rights, defending the value of a human rights-based approach in three different contexts. First, human rights can serve as a focal point for inter-cultural AI value alignment, functioning as part of an ‘overlapping consensus’ between different global value systems. Second, human rights, and their supporting legal instruments, can help determine who is responsible for what in the context of AI ethics, mapping out the duties of different actors including states and technology companies. Third, human rights can serve as a lingua franca that helps bridge the divide between the technical AI research community and civil society and activists on the ground. To illustrate how these claims work in practice, the paper focuses on three specific human rights: freedom from discrimination, health, and access to science.

Author(s): Iason Gabriel

Abstract: This essay explores the relationship between artificial intelligence and principles of distributive justice. Drawing upon the political philosophy of John Rawls, it holds that the basic structure of society should be understood as a composite of sociotechnical systems, and that the operation of these systems is increasingly shaped and influenced by AI. Consequently, egalitarian norms of justice apply to the technology when it is deployed in these contexts. These norms entail that the relevant AI systems must meet a certain standard of public justification, support citizens’ rights, and promote substantively fair outcomes, something that requires particular attention to the impact they have on the worst-off members of society.

Author(s): Iason Gabriel

Abstract: This paper looks at philosophical questions that arise in the context of AI alignment. It defends three propositions. First, normative and technical aspects of the AI alignment problem are interrelated, creating space for productive engagement between people working in both domains. Second, it is important to be clear about the goal of alignment. There are significant differences between AI that aligns with instructions, intentions, revealed preferences, ideal preferences, interests and values. A principle-based approach to AI alignment, which combines these elements in a systematic way, has considerable advantages in this context. Third, the central challenge for theorists is not to identify 'true' moral principles for AI; rather, it is to identify fair principles for alignment, that receive reflective endorsement despite widespread variation in people's moral beliefs. The final part of the paper explores three ways in which fair principles for AI alignment could potentially be identified.

Author(s): Laura Weidinger, Joslyn Barnhart, Jenny Brennan, Christina Butterfield, Susie Young, Will Hawkins, Lisa Anne Hendricks, Ramona Comanescu, Oscar Chang, Mikel Rodriguez, Jennifer Beroshi, Dawn Bloxwich, Lev Proleev, Jilin Chen, Sebastian Farquhar, Lewis Ho, Iason Gabriel, Allan Dafoe, William Isaac

Abstract: Safety and responsibility evaluations of advanced AI models are a critical but developing field of research and practice. In the development of Google DeepMind's advanced AI models, we innovated on and applied a broad set of approaches to safety evaluation. In this report, we summarise and share elements of our evolving approach as well as lessons learned for a broad audience. Key lessons learned include: First, theoretical underpinnings and frameworks are invaluable to organise the breadth of risk domains, modalities, forms, metrics, and goals. Second, theory and practice of safety evaluation development each benefit from collaboration to clarify goals, methods and challenges, and facilitate the transfer of insights between different stakeholders and disciplines. Third, similar key methods, lessons, and institutions apply across the range of concerns in responsibility and safety - including established and emerging harms. For this reason it is important that a wide range of actors working on safety evaluation and safety research communities work together to develop, refine and implement novel evaluation approaches and best practices, rather than operating in silos. The report concludes with outlining the clear need to rapidly advance the science of evaluations, to integrate new evaluations into the development and governance of AI, to establish scientifically-grounded norms and standards, and to promote a robust evaluation ecosystem.

Author(s): Laura Weidinger, Maribeth Rauh, Nahema Marchal, Arianna Manzini, Lisa Anne Hendricks, Juan Mateos-Garcia, Stevie Bergman, Jackie Kay, Conor Griffin, Ben Bariach, Iason Gabriel, Verena Rieser, William Isaac

Abstract: Generative AI systems produce a range of risks. To ensure the safety of generative AI systems, these risks must be evaluated. In this paper, we make two main contributions toward establishing such evaluations. First, we propose a three-layered framework that takes a structured, sociotechnical approach to evaluating these risks. This framework encompasses capability evaluations, which are the main current approach to safety evaluation. It then reaches further by building on system safety principles, particularly the insight that context determines whether a given capability may cause harm. To account for relevant context, our framework adds human interaction and systemic impacts as additional layers of evaluation. Second, we survey the current state of safety evaluation of generative AI systems and create a repository of existing evaluations. Three salient evaluation gaps emerge from this analysis. We propose ways forward to closing these gaps, outlining practical steps as well as roles and responsibilities for different actors. Sociotechnical safety evaluation is a tractable approach to the robust and comprehensive safety evaluation of generative AI systems.

Author(s): Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, Lewis Ho, Divya Siddarth, Shahar Avin, Will Hawkins, Been Kim, Iason Gabriel, Vijay Bolina, Jack Clark, Yoshua Bengio, Paul Christiano, Allan Dafoe

Abstract: Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.

Author(s): Maribeth Rauh, John Mellor, Jonathan Uesato, Po-Sen Huang, Johannes Welbl, Laura Weidinger, Sumanth Dathathri, Amelia Glaese, Geoffrey Irving, Iason Gabriel, William Isaac, Lisa Anne Hendricks

Abstract: 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.

Author(s): Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean Legassick, Geoffrey Irving, Iason Gabriel

Abstract: This paper develops a comprehensive taxonomy of ethical and social risks associated with LMs. We identify twenty-one risks, drawing on expertise and literature from computer science, linguistics, and the social sciences. We situate these risks in our taxonomy of six risk areas: I. Discrimination, Hate speech and Exclusion, II. Information Hazards, III. Misinformation Harms, IV. Malicious Uses, V. Human-Computer Interaction Harms, and VI. Environmental and Socioeconomic harms. For risks that have already been observed in LMs, the causal mechanism leading to harm, evidence of the risk, and approaches to risk mitigation are discussed. We further describe and analyse risks that have not yet been observed but are anticipated based on assessments of other language technologies. We conclude by highlighting challenges and directions for further research on risk evaluation and mitigation with the goal of ensuring that language models are developed responsibly.

Author(s): Amelia Glaese, Nat McAleese, Maja Trębacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato, Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory Greig, Charlie Chen, Doug Fritz, Jaume Sanchez Elias, Richard Green, Soňa Mokrá, Nicholas Fernando, Boxi Wu, Rachel Foley, Susannah Young, Iason Gabriel, William Isaac, John Mellor, Demis Hassabis, Koray Kavukcuoglu, Lisa Anne Hendricks, Geoffrey Irving

Abstract: We present Sparrow, an information-seeking dialogue agent trained to be more helpful, correct, and harmless compared to prompted language model baselines. We use reinforcement learning from human feedback to train our models with two new additions to help human raters judge agent behaviour. \r\n\r\nFirst, to make our agent more helpful and harmless, we break down the requirements for good dialogue into natural language rules the agent should follow, and ask raters about each rule separately. We demonstrate that this breakdown enables us to collect more targeted human judgements of agent behaviour and allows for more efficient rule-conditional reward models. \r\n\r\nSecond, our agent provides evidence from sources supporting factual claims when collecting preference judgements over model statements. For factual questions, evidence provided by Sparrow supports the sampled response 78% of the time. Sparrow is preferred more often than baselines while being more resilient to adversarial probing by humans, violating our rules only 8% of the time when probed. \r\n\r\nFinally, we conduct extensive analyses showing that though our model learns to follow our rules it can exhibit distributional biases.

Author(s): Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Elsen, Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Iason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, Geoffrey Irving

Abstract: Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.

Author(s): Iason Gabriel, Arianna Manzini, Geoff Keeling, Lisa Anne Hendricks et al

Abstract: This paper focuses on the demandingness of morality in an age where spending on luxury goods and extreme poverty continue to exist side by side. If morality grants the wealthy permissions, then what do they allow? If there are limits on what morality may demand of us, then how much does it permit? For a view Henry Shue has termed 'yuppie ethics', the answer to both questions is a great deal. It holds that rich people are morally permitted to spend large amounts of money on themselves, even when this means leaving those living in extreme poverty unaided. Against this view, I demonstrate that personal permissions are limited in certain ways: their strength must be continuous with the reasons put forward to explain their presence inside morality to begin with. Typically, these reasons include non-alienation and the preservation of personal integrity. However, when personal costs do not result in alienation or violate integrity, they are things that morality can routinely demand of us. Yuppie ethics therefore runs afoul of what I call the ‘continuity constraint’.

Author(s): I Gabriel

Abstract: Effective altruism is a philosophy and a social movement that aims to revolutionise the way in which we do philanthropy. It encourages individuals to do as much good as possible, typically by contributing money to the best-performing aid and development organizations. Surprisingly, this approach has met with considerable resistance among aid practitioners. They argue that effective altruism is insensitive to justice insofar as it overlooks the value of equality, urgency and rights. They also hold that the movement suffers from methodological bias, which means that it takes materialistic, individualistic and instrumental approach to doing good. Finally, they maintain that effective altruists hold false empirical beliefs about the world, and that they reach mistaken conclusions about how best to act for that reason. This paper weighs the force of each objection in turn, and looks at responses to the challenge they pose. 

Author(s): H Lazenby, I Gabriel

Abstract: Award-winning paper (OUP Best of Philosophy, 2018) that offers an account of the information condition on morally valid consent in the context of sexual relations. The account is grounded in rights. It holds that a person has a sufficient amount of information to give morally valid consent if, and only if, she has all the information to which she has a claim-right. A person has a claim-right to a piece of information if, and only if: a. it concerns a deal-breaker for her; b. it does not concern something that her partner has a strong interest in protecting from scrutiny, sufficient to generate a privilege-right; c.i. her partner is aware of the information to which her deal-breaker applies, or c.ii. her partner ought to be held responsible for the fact that he is not aware of the information to which her deal-breaker applies; and d. she has not waived or forfeited her claim-right.