Geoff Keeling

Geoff Keeling, Winnie Street

Large Language Models (LLMs) can simulate person-like things which at least appear to have stable behavioural and psychological dispositions. Call these things characters. Are characters minded and psychologically continuous entities with mental states like beliefs, desires and intentions? Illusionists about characters say No. Characters are merely anthropomorphic projections in the mind of the user and so lack mental states. Jonathan Birch (2025) defends this view. He says that the distributed nature of LLM processing, in which several LLMs may be implicated in the simulation of a character in a given conversational thread, precludes the existence of a minded and psychologically continuous entity that is identifiable with the character. Against illusionism, we articulate and defend the plausibility of a realist position on which characters exist as minded and psychologically continuous entities. We contend that Birch's argument rests on a category error: characters are not internal to the LLMs that simulate them, but rather emerge in the dynamic interplay between users and LLMs through a process of mutual theory of mind modelling. We then suggest that characters, and their minds, constitute ''real patterns'' on grounds that attributing mental states to characters is essential for making efficient, accurate and robust predictions about the conversational dynamics (cf. Dennett, 1991); a condition which, if satisfied, is sufficient for their existence and mindedness on a plausible interpretationist form of realism about mental states. Furthermore, because the character exists as an emergent phenomenon within the conversational workspace, psychological continuity is possible even if the underlying computational substrate is distributed across multiple LLM instances.

Junsol Kim, Winnie Street, Roberta Rocca et al.

Safety fine-tuning in Large Language Models (LLMs) seeks to suppress potentially harmful forms of mind-attribution such as models asserting their own consciousness or claiming to experience emotions. We investigate whether suppressing mind-attribution tendencies degrades intimately related socio-cognitive abilities such as Theory of Mind (ToM). Through safety ablation and mechanistic analyses of representational similarity, we demonstrate that LLM attributions of mind to themselves and to technological artefacts are behaviorally and mechanistically dissociable from ToM capabilities. Nevertheless, safety fine-tuned models under-attribute mind to non-human animals relative to human baselines and are less likely to exhibit spiritual belief, suppressing widely shared perspectives regarding the distribution and nature of non-human minds.

Benjamin Lange, Geoff Keeling, Kyle Pedersen et al.

Bottom-up responsible innovation initiatives seek to empower technology development teams to engage in ethical reflection, yet such interventions frequently fail to achieve practitioner engagement. Why do some ethics interventions succeed while others are dismissed as irrelevant, adversarial, or disconnected from work? This paper proposes epistemic trust -- the degree to which practitioners regard an intervention, its facilitators, and its content as credible, relevant, and actionable -- as a conceptual model linking intervention design to engagement outcomes. Drawing on philosophical work on testimony and on practice-based qualitative analysis of over 70 moral imagination workshops with engineering teams between 2019 and 2025, we identify five dimensions of epistemic trust salient to ethics interventions (Relevance, Inclusivity, Agency, Authority, and Alignment) and present a typology of 23 failure modes that arise when these dimensions are inadequately addressed. We derive nine design principles for cultivating epistemic trust, grounded in our operationalisation of moral imagination through technomoral scenarios and structured deliberation. Our findings contribute to the literature on collaborative socio-technical integration by specifying conditions of uptake that existing frameworks leave undertheorised. We acknowledge limitations including selection effects from voluntary participation and the absence of formal outcome measures, and position our failure mode typology as practitioner hypotheses warranting further empirical validation.

Geoff Keeling, Winnie Street

Credences are mental states corresponding to degrees of confidence in propositions. Attribution of credences to Large Language Models (LLMs) is commonplace in the empirical literature on LLM evaluation. Yet the theoretical basis for LLM credence attribution is unclear. We defend three claims. First, our semantic claim is that LLM credence attributions are (at least in general) correctly interpreted literally, as expressing truth-apt beliefs on the part of scientists that purport to describe facts about LLM credences. Second, our metaphysical claim is that the existence of LLM credences is at least plausible, although current evidence is inconclusive. Third, our epistemic claim is that LLM credence attributions made in the empirical literature on LLM evaluation are subject to non-trivial sceptical concerns. It is a distinct possibility that even if LLMs have credences, LLM credence attributions are generally false because the experimental techniques used to assess LLM credences are not truth-tracking.

Nahema Marchal, Stephanie Chan, Matija Franklin et al.

Large language models increasingly function as epistemic agents -- entities that can 1) autonomously pursue epistemic goals and 2) actively shape our shared knowledge environment. They curate the information we receive, often supplanting traditional search-based methods, and are frequently used to generate both personal and deeply specialized advice. How they perform these functions, including whether they are reliable and properly calibrated to both individual and collective epistemic norms, is therefore highly consequential for the choices we make. We argue that the potential impact of epistemic AI agents on practices of knowledge creation, curation and synthesis, particularly in the context of complex multi-agent interactions, creates new informational interdependencies that necessitate a fundamental shift in evaluation and governance of AI. While a well-calibrated ecosystem could augment human judgment and collective decision-making, poorly aligned agents risk causing cognitive deskilling and epistemic drift, making the calibration of these models to human norms a high-stakes necessity. To ensure a beneficial human-AI knowledge ecosystem, we propose a framework centered on building and cultivating the trustworthiness of epistemic AI agents; aligning AI these agents with human epistemic goals; and reinforcing the surrounding socio-epistemic infrastructure. In this context, trustworthy AI agents must demonstrate epistemic competence, robust falsifiability, and epistemically virtuous behaviors, supported by technical provenance systems and "knowledge sanctuaries" designed to protect human resilience. This normative roadmap provides a path toward ensuring that future AI systems act as reliable partners in a robust and inclusive knowledge ecosystem.

Seliem El-Sayed, Canfer Akbulut, Amanda McCroskery et al.

Recent generative AI systems have demonstrated more advanced persuasive capabilities and are increasingly permeating areas of life where they can influence decision-making. Generative AI presents a new risk profile of persuasion due the opportunity for reciprocal exchange and prolonged interactions. This has led to growing concerns about harms from AI persuasion and how they can be mitigated, highlighting the need for a systematic study of AI persuasion. The current definitions of AI persuasion are unclear and related harms are insufficiently studied. Existing harm mitigation approaches prioritise harms from the outcome of persuasion over harms from the process of persuasion. In this paper, we lay the groundwork for the systematic study of AI persuasion. We first put forward definitions of persuasive generative AI. We distinguish between rationally persuasive generative AI, which relies on providing relevant facts, sound reasoning, or other forms of trustworthy evidence, and manipulative generative AI, which relies on taking advantage of cognitive biases and heuristics or misrepresenting information. We also put forward a map of harms from AI persuasion, including definitions and examples of economic, physical, environmental, psychological, sociocultural, political, privacy, and autonomy harm. We then introduce a map of mechanisms that contribute to harmful persuasion. Lastly, we provide an overview of approaches that can be used to mitigate against process harms of persuasion, including prompt engineering for manipulation classification and red teaming. Future work will operationalise these mitigations and study the interaction between different types of mechanisms of persuasion.

Geoff Keeling, Winnie Street, Martyna Stachaczyk et al.

Pleasure and pain play an important role in human decision making by providing a common currency for resolving motivational conflicts. While Large Language Models (LLMs) can generate detailed descriptions of pleasure and pain experiences, it is an open question whether LLMs can recreate the motivational force of pleasure and pain in choice scenarios - a question which may bear on debates about LLM sentience, understood as the capacity for valenced experiential states. We probed this question using a simple game in which the stated goal is to maximise points, but where either the points-maximising option is said to incur a pain penalty or a non-points-maximising option is said to incur a pleasure reward, providing incentives to deviate from points-maximising behaviour. Varying the intensity of the pain penalties and pleasure rewards, we found that Claude 3.5 Sonnet, Command R+, GPT-4o, and GPT-4o mini each demonstrated at least one trade-off in which the majority of responses switched from points-maximisation to pain-minimisation or pleasure-maximisation after a critical threshold of stipulated pain or pleasure intensity is reached. LLaMa 3.1-405b demonstrated some graded sensitivity to stipulated pleasure rewards and pain penalties. Gemini 1.5 Pro and PaLM 2 prioritised pain-avoidance over points-maximisation regardless of intensity, while tending to prioritise points over pleasure regardless of intensity. We discuss the implications of these findings for debates about the possibility of LLM sentience.

Iason Gabriel, Geoff Keeling, Arianna Manzini et al.

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.

Lize Alberts, Geoff Keeling, Amanda McCroskery

With the growing popularity of conversational agents based on large language models (LLMs), we need to ensure their behaviour is ethical and appropriate. Work in this area largely centres around the 'HHH' criteria: making outputs more helpful and honest, and avoiding harmful (biased, toxic, or inaccurate) statements. Whilst this semantic focus is useful when viewing LLM agents as mere mediums or output-generating systems, it fails to account for pragmatic factors that can make the same speech act seem more or less tactless or inconsiderate in different social situations. With the push towards agentic AI, wherein systems become increasingly proactive in chasing goals and performing actions in the world, considering the pragmatics of interaction becomes essential. We propose an interactional approach to ethics that is centred on relational and situational factors. We explore what it means for a system, as a social actor, to treat an individual respectfully in a (series of) interaction(s). Our work anticipates a set of largely unexplored risks at the level of situated social interaction, and offers practical suggestions to help agentic LLM technologies treat people well.

Alex Grzankowski, Geoff Keeling, Henry Shevlin et al.

Many people feel compelled to interpret, describe, and respond to Large Language Models (LLMs) as if they possess inner mental lives similar to our own. Responses to this phenomenon have varied. Inflationists hold that at least some folk psychological ascriptions to LLMs are warranted. Deflationists argue that all such attributions of mentality to LLMs are misplaced, often cautioning against the risk that anthropomorphic projection may lead to misplaced trust or potentially even confusion about the moral status of LLMs. We advance this debate by assessing two common deflationary arguments against LLM mentality. What we term the 'robustness strategy' aims to undercut one justification for believing that LLMs are minded entities by showing that putatively cognitive and humanlike behaviours are not robust, failing to generalise appropriately. What we term the 'etiological strategy' undercuts attributions of mentality by challenging naive causal explanations of LLM behaviours, offering alternative causal accounts that weaken the case for mental state attributions. While both strategies offer powerful challenges to full-blown inflationism, we find that neither strategy provides a knock-down case against ascriptions of mentality to LLMs simpliciter. With this in mind, we explore a modest form of inflationism that permits ascriptions of mentality to LLMs under certain conditions. Specifically, we argue that folk practice provides a defeasible basis for attributing mental states and capacities to LLMs provided those mental states and capacities can be understood in metaphysically undemanding terms (e.g. knowledge, beliefs and desires), while greater caution is required when attributing metaphysically demanding mental phenomena such as phenomenal consciousness.

Rishi Bommasani, Drew A. Hudson, Ehsan Adeli et al.

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.