A5030262115- Ethics and Social Impacts of AI
- Explainable Artificial Intelligence (XAI)
- Psychology of Moral and Emotional Judgment
- Artificial Intelligence in Healthcare and Education
- Autonomous Vehicle Technology and Safety
- Law, AI, and Intellectual Property
- Decision-Making and Behavioral Economics
- Dementia and Cognitive Impairment Research
- AI-based Problem Solving and Planning
- Multi-Agent Systems and Negotiation
- Human-Automation Interaction and Safety
- Traffic and Road Safety
- Blockchain Technology Applications and Security
- Behavioral Health and Interventions
- Neuroethics, Human Enhancement, Biomedical Innovations
- Reinforcement Learning in Robotics
- Free Will and Agency
- Death Anxiety and Social Exclusion
- Context-Aware Activity Recognition Systems
- Climate Change Communication and Perception
- Bayesian Modeling and Causal Inference
- Social Robot Interaction and HRI
- Technology Use by Older Adults
- Legal and Policy Issues
- Adversarial Robustness in Machine Learning
University College London
2020-2024
Cambridge Cognition (United Kingdom)
2023
Public Health England
2022
This paper focuses on the opportunities and ethical societal risks posed by advanced AI assistants. We define assistants as artificial agents with natural language interfaces, whose function is to plan execute sequences of actions behalf a user, across one or more domains, in line user's expectations. The starts considering technology itself, providing an overview assistants, their technical foundations potential range applications. It then explores questions around value alignment,...
The rapid development of advanced AI agents and the imminent deployment many instances these will give rise to multi-agent systems unprecedented complexity. These pose novel under-explored risks. In this report, we provide a structured taxonomy risks by identifying three key failure modes (miscoordination, conflict, collusion) based on agents' incentives, as well seven risk factors (information asymmetries, network effects, selection pressures, destabilising dynamics, commitment problems,...
Automated vehicles (AVs) have made huge strides toward large-scale deployment. Despite this progress, AVs continue to make mistakes, some resulting in death. Although mistakes are avoidable, others hard avoid even by highly skilled drivers. As these shape attitudes AVs, we need understand whether people differentiate between them. We ask the following two questions. When an AV makes a mistake, does perceived difficulty or novelty of situation predict blame attributed it? How that attribution...
Abstract The European Union (EU) is in the middle of comprehensively regulating artificial intelligence (AI) through an effort known as AI Act. Within vast spectrum issues under Act’s aegis, treatment technologies classified general purpose systems (GPAIS) merits special consideration. Particularly, existing proposals to define GPAIS do not provide sufficient guidance distinguish these from those designed perform specific tasks, denominated fixed-purpose. Thus, our working paper has three...
The European Union Artificial Intelligence (AI) Act proposes to ban AI systems that ”manipulate persons through subliminal techniques or exploit the fragility of vulnerable individuals, and could potentially harm manipulated individual third person”. This article takes perspective cognitive psychology analyze understand what algorithmic manipulation consists of, who individuals may be, is considered as harm. Subliminal are expanded with concepts from behavioral science study preference...
Recent empirical work on people's attributions of responsibility toward artificial autonomous agents (such as Artificial Intelligence or robots) has delivered mixed findings. The conflicting results reflect differences in context, the roles AI and human agents, domain application. In this article, we outline a causal framework attribution which integrates these It outlines nine factors that influence - causality, role, knowledge, objective foreseeability, capability, intent, desire,...
This paper argues that market governance mechanisms should be considered a key approach in the of artificial intelligence (AI), alongside traditional regulatory frameworks. While current approaches have predominantly focused on regulation, we contend market-based offer effective incentives for responsible AI development. We examine four emerging vectors governance: insurance, auditing, procurement, and due diligence, demonstrating how these can affirm relationship between risk financial...
This paper argues that model-free reinforcement learning (RL) agents, while lacking explicit planning mechanisms, exhibit behaviours can be analogised to System 1 ("thinking fast") processes in human cognition. Unlike model-based RL which operate akin 2 slow") reasoning by leveraging internal representations for planning, agents react environmental stimuli without anticipatory modelling. We propose a novel framework linking the dichotomy of and distinction between RL. framing challenges...
Recent work showed Best-of-N (BoN) jailbreaking using repeated use of random augmentations (such as capitalization, punctuation, etc) is effective against all major large language models (LLMs). We have found that $100\%$ the BoN paper's successful jailbreaks (confidence interval $[99.65\%, 100.00\%]$) and $99.8\%$ in our replication $[99.28\%, 99.98\%]$) were blocked with Defense Against The Dark Prompts (DATDP) method. DATDP algorithm works by repeatedly utilizing an evaluation LLM to...
Abstract The dominant practice of AI alignment assumes (1) that preferences are an adequate representation human values, (2) rationality can be understood in terms maximizing the satisfaction preferences, and (3) systems should aligned with one or more humans to ensure they behave safely accordance our values. Whether implicitly followed explicitly endorsed, these commitments constitute what we term a preferentist approach alignment. In this paper, characterize challenge approach, describing...
The European Union (EU) is in the middle of comprehensively regulating artificial intelligence (AI) through an effort known as AI Act. Within vast spectrum issues under Act's aegis, treatment technologies classified general purpose systems (GPAIS) merits special consideration. Particularly, existing proposals to define GPAIS do not provide sufficient guidance distinguish these from those designed perform specific tasks, denominated fixed-purpose. Thus, our working paper has three objectives....
Recent generative AI systems have demonstrated more advanced persuasive capabilities and are increasingly permeating areas of life where they can influence decision-making. Generative presents a new risk profile persuasion due the opportunity for reciprocal exchange prolonged interactions. This has led to growing concerns about harms from how be mitigated, highlighting need systematic study persuasion. The current definitions unclear related insufficiently studied. Existing harm mitigation...
This study investigates how people assign blame to autonomous vehicles (AVs) when involved in an accident. Our experiment (N = 2647) revealed that placed more on AVs than human drivers accident details were unspecified. To examine whether assess major classes of blame-relevant information differently for and humans, we developed a causal model introduced novel concept prevention effort, which emerged as crucial factor or judgement alongside intentionality. Finally, addressed the “many hands”...
As artificial intelligence becomes more powerful and a ubiquitous presence in daily life, it is imperative to understand manage the impact of AI systems on our lives decisions. Modern ML often change user behavior (e.g. personalized recommender learn preferences deliver recommendations that online behavior). An externality preference change. This article argues for establishment multidisciplinary endeavor focused understanding how preference: Preference Science. We operationalize incorporate...
Globally around 50 million people are currently living with dementia, and there nearly 10 new cases every year. The decline of memory and, it, lack self-confidence continuous confusion have a devastating effect on this disease. Dementia patients even struggle to accomplish mundane chores require assistance for daily social connectedness. Over the past decade, we seen remarkable growth in wearable technologies manage our health wellbeing improve awareness However, ask why wearables not...
The third driverless car competition of the DARPA Grand Challenge (Urban Challenge) in 2007 saw six autonomous vehicle teams finishing event successfully. Since then, Automated Vehicles (AVs) made huge strides towards deployment on a large scale. Despite all this progress, AVs continue to make mistakes, some which have resulted deaths passengers and pedestrians. These crashes received wide coverage media drew parallel bleak picture public’s lack enthusiasm for technology. However, not...
Abstract Background Sending a social norms feedback letter to general practitioners who are high prescribers of antibiotics has been shown reduce antibiotic prescribing. The 2017-9 Quality Premium for primary care in England sets target broad-spectrum prescribing, which should be at or below 10% total We tested norm that targeted prescribing and the addition chart text-only overall Methods conducted three 2-armed randomised controlled trials, on different groups practices: Trial A compared...
In a recent article K ̈obis, Bonnefon, and Rahwan (2021) propose framework to identify four different primary roles in which Artificial Intelligence (AI) cause unethical or corrupt human behaviour; namely - role model, delegate, partner, advisor. this we two further AI as boss counterparty. We argue that the exerts coercive power over its employees whilst perceptual abilities of an counterparty provide opportunity for humans behave differently towards them than they would with analogues....
The European Union's Artificial Intelligence Act aims to regulate manipulative and harmful uses of AI, but lacks precise definitions for key concepts. This paper provides technical recommendations improve the Act's conceptual clarity enforceability. We review psychological models define "personality traits," arguing should protect full "psychometric profiles." urge expanding "behavior" include "preferences" since preferences causally influence are influenced by behavior. Clear provided...
Given rapid progress toward advanced AI and risks from frontier systems (advanced pushing the boundaries of capabilities frontier), creation implementation governance regulatory schemes deserves prioritization substantial investment. However, status quo is untenable and, frankly, dangerous. A gap has permitted labs to conduct research, development, deployment activities with minimal oversight. In response, system evaluations have been proposed as a way assessing development systems. Yet,...
The dominant practice of AI alignment assumes (1) that preferences are an adequate representation human values, (2) rationality can be understood in terms maximizing the satisfaction preferences, and (3) systems should aligned with one or more humans to ensure they behave safely accordance our values. Whether implicitly followed explicitly endorsed, these commitments constitute what we term a preferentist approach alignment. In this paper, characterize challenge approach, describing...