The Bristol AI Summer School (BAIS) 2026

The Interactive AI and the Practice-Oriented AI CDTs at the University of Bristol will hold an in-person Summer School between the 8 and 10 September 2026. For three days the fundamentals and the latest progress in key areas of AI will be discussed by a range of experts from both industry and academia. This year, the Summer School will broadly focus on Human-Centred AI.

T‌his year's event will take place at Engineers' House in Bristol. 

Programme (Subject to change):

 

Tuesday 8 September, 2026

 

Talk title: Guiding Sociotechnical Systems towards Value-Norm Equilibrium

Abstract: 

Values and norms are complementary constructs that undergird prosocial behaviour in sociotechnical systems (STSs). Whereas values are intrinsic motivators for prosocial behaviour, norms are extrinsic motivators for meeting mutual expectations. An STS is in equilibrium when the values of its member actors and the norms that govern it align with each other. Such an equilibrium is not permanent, as actors join or leave the STS and their values and norms evolve. 
In general, an STS must be guided towards equilibrium by systematically refining its norm specifications and influencing the values of its member actors. This tutorial introduces value-norm equilibrium from first principles and discusses how we can detect misalignment in an STS, as well as how an STS can be guided towards and maintained in equilibrium. We examine the key computational challenges of representing values and norms, measuring their alignment, identifying tipping points, managing adaptation timescales, revising norms, and facilitating value reflection.

Bio:

Nirav is a Senior Lecturer in Artificial Intelligence in the School of Computer Science at the University of Bristol. His research interests are in AI, intelligent agents, and multiagent systems with an emphasis on ethics, cybersecurity, and privacy.

 

This session will be followed by 11:15 - 12:00 - Poster session from CDT students (IAI and Pr0AI)

 

Talk title: Citizen Centric AI

Abstract:

AI can help us address important societal challenges, from switching to cleaner transportation to achieving net zero through electrification. However, to realise these benefits, we need AI systems that are citizen-centric and that can be trusted by citizen end users. This can only happen if AI systems act in the interest of users, while also remaining robust to strategic behaviour when allocating scarce resources (like energy or transportation capacity). In this talk, I will present recent work on how we tackle these challenges, by first learning the preferences of users through data-efficient methods, and then designing mechanisms that manage strategic behaviour through appropriate incentives.

Bio:

Seb Stein is a Professor of AI and Multi-Agent Systems at the University of Southampton, where he leads the Citizen-Centric AI Systems team (https://ccais.ac.uk/). His work focuses on designing AI systems that support sustainability, especially in smart energy and transportation. Within this area, he has worked on preference learning, human-AI collaboration and mechanism design, and has published 150+ papers. Seb is also on the management board of the UKRI AI Centre for Doctoral Training on AI for Sustainability (https://sustai.info/) and Co-Lead of the EPSRC FEVER Programme Grant on off-grid electric vehicle charging (https://www.fever-ev.ac.uk/).

Talk title: Sensorimotor Regularities as Alignment between Humans and Large Language Models

Abstract: 

Large language models produce fluent, human-sounding text, but fluency is not evidence that they represent concepts as humans do. Human conceptual representation draws on two sources: linguistic distributional knowledge and sensorimotor knowledge. The latter consists of regularities abstracted from repeated bodily experience, known in cognitive linguistics as image schemas. Disembodied language models have access only to the first. This talk presents a framework for measuring the consequences. Sensorimotor regularities are operationalised into three metrics that can be elicited from any linguistic agent, human or artificial, allowing people and models to be compared on the same scale: the distribution of image schemas, the associations formed between schemas and abstract concepts, and the co-occurrences between schema pairs. Applied across contemporary language models and human participants, the framework reveals a consistent pattern. Models reproduce the human inventory of conceptual building blocks closely, but diverge systematically in how those blocks are mapped onto concepts and in which ones they combine. Further, I show that augmenting a model with targeted sensorimotor priors derived from its own diagnosed misalignments yields outputs that human evaluators rate as significantly clearer, more contextually contingent, and more human-like. I will close on what conceptual-level alignment means for designing intuitive human-AI interaction.

Bio:

Jingyi Li is an HCI researcher working on human-centred AI and human–AI interaction. She is a Senior Research Associate in the School of Computer Science at the University of Bristol, where she designs generative AI systems with a focus on legibility, controllability, and human agency, and studies how people understand and appropriate such systems in practice. Her research grounds generative AI and Mixed Reality systems in sensorimotor regularities, with the aim of keeping interactive systems aligned with how people perceive and act. She completed her PhD in Engineering at the University of Cambridge in the Intelligent Interactive Systems Group, where her work brought image schemas from cognitive linguistics into the design and evaluation of immersive systems and generative AI systems.

 

Talk title: AI impact on Sustainability 

Abstract:

AI is increasingly presented both as a tool for addressing sustainability challenges and as a growing sustainability problem in its own right. Yet assessments of AI sustainability often focus on the AI component itself: its energy consumption, carbon emissions, computational efficiency, or ability to optimise a particular task. This talk shows that AI impact on sustainability is a property of the socio-technical system in which AI is deployed, not of the AI component alone. Whether an apparently efficient AI system produces a positive sustainability outcome depends on how it changes human behaviour, work practices, resource use, organisational processes, infrastructure, and the distribution of costs and benefits.

Bio:

Ruzanna Chitchyan is Professor of Software Engineering for Sustainability at the University of Bristol, where her research focuses on the role of digital technologies and software systems in sustainability transitions. Her work takes a socio-technical perspective, examining how technologies interact with people, organisations, infrastructure, policy and wider social systems to shape sustainability outcomes. Her current research spans sustainable software and AI, smart energy and heat pumps, local energy systems, and digital systems supporting the circular economy.

Talk title:: LEMON: Local Explanations via Modality-aware OptimisatioN

Abstract:
Multimodal models are ubiquitous, yet existing explainability methods are often single-modal, architecture-dependent, or too computationally expensive to run at scale. We introduce LEMON (Local Explanations via Modality-aware OptimisatioN), a model-agnostic framework for local explanations of multimodal predictions. LEMON fits a single modality-aware surrogate with group-structured sparsity to produce unified explanations that disentangle modality-level contributions and feature-level attributions. The approach treats the predictor as a black box and is computationally efficient, maintaining competitive deletion-based faithfulness at substantially reduced query cost. We evaluate LEMON on vision–language question answering and a clinical prediction task with image, text, and tabular inputs, comparing against representative multimodal baselines. Across backbones, LEMON achieves competitive deletion-based faithfulness while reducing black-box evaluations by up to 67x and runtime by up to 8x compared to strong multimodal baselines. 

Wednesday 9 September, 2026

 

 

 

 

Title: How Much Reasoning Can Prediction Buy? Context-Directed Extrapolation and the Limits of Generalisation in LLMs

Abstract: 

Large language models have demonstrated exceptional performance across diverse tasks for which they were not explicitly trained, including those that require complex reasoning abilities in people. However, despite their impressive capabilities, LLMs tend to fail unexpectedly even on tasks that young children can solve with relative ease. How do we reconcile these conflicting capabilities and limitations of state-of-the-art LLMs? To answer this question, I will draw on a series of papers (ACL, TMLR) and argue that both follow from one mechanism. What we call reasoning is “context-directed extrapolation”: extrapolation from training priors, directed by the prompt. This gives one account of capability and failure alike. More importantly, from this perspective, it becomes possible to predict, and therefore begin to mitigate, specific failure modes in even the most advanced “thinking” models (EMNLP, book chapters). I will close by exploring the open problem this leaves, characterising the boundary of generalisation itself and how it moves as models scale.

Bio:

Harish Tayyar Madabushi is a Senior Lecturer (Associate Professor) in Artificial Intelligence at the University of Bath. His research focuses on the fundamental mechanisms that underpin the performance and functioning of large language models such as ChatGPT. His work was included in the discussion paper on the Capabilities and Risks of Frontier AI that informed discussions at the UK AI Safety Summit held at Bletchley Park. His research on the constructional information encoded in language models has been influential in bringing together the fields of construction grammar and pre-trained language models. His work also includes collaborative industrial research aimed at correcting biases in speech-to-text systems widely used across the UK. Before starting his PhD in automated question answering at the University of Birmingham, he founded and led a social media data analytics company based in Singapore.

 

 

Title: Current research practices in NeuroAI do not support many of the strong claims of ANN-Human Alignment

Abstract: 

Artificial neural networks (ANNs) developed in computer science are successful in a range of vision, language, and reasoning tasks.  They can also predict behavioural responses and brain activations of humans better than alternative models. This has led to the common claim that ANNs are the best models of biological intelligence. However, most prediction studies are correlational, and accordingly, do not support causal conclusions.  Furthermore, researchers are incentivized to identify ANN-human similarities, as reviewers and editors are more likely to publish studies that report similarities rather than differences.  Accordingly, researchers rarely carry out “severe” tests of their claims that are more likely to falsify their conclusions (if indeed the conclusions are false).  I show when the relevant experiments are carried out, ANNs do a poor job in explaining human intelligence.  The field of NeuroAI needs to change its methods to better characterize ANN-human alignment and build better models of minds.

 

Talk title: TBC

Bio:

Amid Ayobi is a Lecturer (Assistant Professor) at University College London and a member of the UCL Interaction Centre. His research team investigates conversational AI systems and applies human-centred design approaches to support the lived experiences and self-care needs of people living with chronic health conditions.
 
He received his doctorate in Human-Computer Interaction from UCL, where he focused on self-tracking in multiple sclerosis care. As a postdoctoral researcher at the University of Bristol, he investigated the digital health needs of young adults with diabetes. He developed his HCI expertise through multidisciplinary R&D roles at IBM, SAP, and Microsoft.

Title: AI Trust and Security workshop

 

 

Thursday 10 September, 2026 - student-centred day for IAI and PrOAI CDT students

An Introduction to Responsible Research and Innovation