Paired Non-Anthropomorphic Robots as Tangible Interfaces
to Alleviate Loneliness: A Co-Design and Evaluation Study
Loneliness increasingly impacts young adults, yet standard robot companions raise ethical concerns about replacing genuine human connection. This research explores an alternative: paired, connected non-anthropomorphic robots serving as tangible interfaces for mediated social touch between distant people. Co-design workshops ($N=10$) with lonely young adults informed the robots’ tactile behaviors. A laboratory user study ($N=23$) then compared a standalone robot against a robot perceived as paired to a distant person’s device. While quantitative acceptance scores (Almere and Godspeed) were similar across conditions, qualitative and behavioral data revealed that the paired connection elicited significantly deeper emotional engagement, more affectionate physical interactions (hugging, caressing), and more expressive facial reactions. Participants reported feeling genuinely understood by another person. These findings demonstrate that paired non-anthropomorphic robots can effectively alleviate loneliness by extending, rather than replacing, human communication, highlighting the critical role of relational framing in robotic design for emotional wellbeing.
Multi to multi object near-neighbour search
Previous work has introduced the idea of polyadic query similarity search as a multi-query generalisation of single query similarity search, enabling retrieval based on the joint characteristics of multiple query arguments. The initial work on polyadic query search has focused solely on searching over a collection of singular objects. In this work we aim to expand this idea to search over a collection of sets of objects and assess its effectiveness. Through the means of practical examples on image embeddings generated from the mirflickr dataset, we demonstrate that polyadic query search over a collection of sets of objects is feasible and somewhat effective.
Event-Camera Visual Speech Recognition: DVS-Based Lip Reading with Deep Learning
Visual Speech Recognition (VSR), is the task of recognizing spoken speech from mouth movements. Dynamic Vision Sensor (DVS) event cameras record the dynamic activity in a scene as sparse streams of independent asynchronous microsecond-timestamped events that indicate the pixel location, timestamp and polarity of a brightness change. Unlike traditional frame-based cameras, DVS cameras are motion-selective and record only these changes, effectively filtering out any static background and providing very high temporal resolution and low latency[AA1.1]. These features are beneficial for capturing the high-frequency articulatory information of lip movements. In this preliminary [AA2.1]work, we propose DVS-LipNet, a hybrid ResNet3D- Bidirectional LSTM architecture that is trained with Connectionist Temporal Classification loss on DVS events simulated from the TCD-TIMIT corpus from 62 speakers. DVS-LipNet obtains a Mean Word Error Rate (MWER) of 3.4% and perfect sentence accuracy on 77.4% of sentences, providing the first baseline sentence-level results for DVS lip reading.
Listening Carefully: Automatic Speech Recognition and the Lombard Effect
The Lombard effect is the phenomenon where individuals subconsciously [AA1.1]change the way they speak (such as volume, mouth movements, and pitch) in the presence of background noise (such as in a conference room or a music venue). This is something that state-of-the art Automatic Speech Recognition (ASR) generally ignores[AA2.1]. Our research investigates [AA3.1]Transformer-based ASR deep learning models for speech recognition, focusing on the performance impact of Lombard speech[AA4.1]. We analyse several state-of-the-art speech recognition models including Wav2Vec 2.0, HuBERT, and WavLM, with Lombard speech databases[AA5.1]. Our results show [AA6.1]that even with state-of-the-art deep learning models , there is still some performance lost to the Lombard effect, even when noise is removed. Future work will build on this b[AA7.1]y assessing performance of Transformer models utilizing visual speech recognition as an additional modality, and the development of a further Lombard robust ASR system utilizing multiple models.
Predicting Breast Cancer Subtype for Personalised Treatment Using Federated Learning
Breast cancer remains one of the most prevalent cancers worldwide and is characterized by significant biological heterogeneity. It comprises multiple molecular subtypes, including Luminal A, Luminal B, HER2-enriched, and triple-negative, each associated with distinct clinical outcomes and therapeutic responses. Accurate identification of these subtypes is critical for guiding treatment selection and estimating patient prognosis. Recent advances in machine learning have enabled data-driven approaches for subtype classification; however, these models often require large, centralized datasets, raising concerns regarding data privacy, security, and institutional data sharing barriers. Federated learning has emerged as a promising solution to these challenges by enabling collaborative model training across multiple institutions without requiring the exchange of sensitive patient data. In this study, we propose a federated learning framework for breast cancer molecular subtype prediction using distributed clinical and genomic datasets. The model is trained across decentralized data sources, allowing institutions to retain data locally while contributing to a shared global model. This approach enhances generalizability, preserves patient privacy, and leverages diverse datasets to improve predictive performance. Our findings suggest that federated learning can achieve robust and accurate subtype classification comparable to traditional centralized models, while addressing key ethical and regulatory constraints. This work highlights the potential of privacy-preserving artificial intelligence in advancing precision oncology and improving breast cancer care.
Harnessing Artificial Intelligence for Drug Toxicity Prediction
Clinical drug development is costly and associated with a high attrition rate. For instance, UK clinical trials can cost up to £1.3 million per study, with an ~90% failure rate. Conventional toxicity testing could be both time-consuming and resource-intensive. This has driven growing interest in artificial intelligence (AI) based approaches for earlier and more efficient in silico prediction of the toxicological risk of drug candidates. Among these, large language models (LLMs) such as Llama and Gemini have increasingly been explored for data generation, knowledge extraction, and predictive support. However, LLM outputs (e.g., toxicity prediction) can vary considerably in response to minor changes in prompt (i.e., question) structure or wording, raising concerns regarding reproducibility, robustness, and sensitivity to prompt engineering. We develop a framework to evaluate the importance of prompt engineering in predicting drug toxicity. The work investigates the extent to which prompt design influences predictive performance for a molecule’s toxic behaviour (e.g., Toxicophores, hydrophobicity). To achieve this, we generate numerous data sets using LLMs, which were subsequently analysed using machine learning algorithms to evaluate the prediction success rate. Results demonstrate that the intrinsic variability of LLM outputs outweigh the effects of prompt fine-tuning alone, and future work in the area should focus on more substantive improvements to the process than optimising the prompts used. This is important in medicine, where unreliable toxicity predictions can lead to poor candidate selection, wasted resources, and the advancement of unsafe compounds through early stages of development.
Context-Intention Interactive Computing: A Personalised Digital Reading Environment
Current AI-assisted reading systems provide generic support through summarisation, explanation, and question answering, but rarely adapt to individual user intent, cognitive state, or reading motivation. As a result, these systems may over-assist users, reduce engagement, or fail to provide support that aligns with a reader’s goals. We introduce a novel approach for visualising user cognitive effort during reading through a dynamic knowledge graph representation, referred to as the Cognitive reading Graph (CRG). The CRG aims to provide real-time insight into the reader’s interaction with the text by highlighting areas of focus, difficulty, and inferred importance. By making the system’s internal reasoning more transparent, this approach supports explainable AI and enables users to better understand how assistance is generated. This research explores whether visualising cognitive effort and adapting support based on user data can help improve reading engagement, comprehension, and allow for self-regulation during reading.
Development of Personality-Orientated Dialogue System for Humanoid Robots
Despite advancements in AI-driven robotics, developing robots with human-level emotional intelligence is still in its infancy. The focus has been on emulating human appearance and behaviour, particularly in humanoid robots. As robots become integrated into daily life, understanding their societal implications is vital. Emotional and social intelligence are crucial for fostering human acceptance and trust, reducing discomfort associated with the human-like appearances of such robots. My research, inspired by Gartner’s multiple intelligence theory, aims to capture sentiments from various intelligence types to shape nuanced personality traits in humanoid robots. This involves exploring methods to influence large language models through prompting, persona steering, and fine-tuning, using a voice-based, offline dialogue system. To achieve this, my research will explore three primary approaches for influencing the behaviour of large language models (LLMs), which serve as the core intelligence component of the dialogue system pipeline: prompt engineering, persona steering, and fine-tuning with a bespoke dataset. This work involves conducting experiments with a range of open-source pre-trained LLMs to evaluate the effectiveness of each approach in shaping model performance, interaction style and consistent personality display. The dialogue system is voice-based and operates entirely offline, ensuring that all speech processing and language generation are performed locally without reliance on cloud services.
Streaming Optimisation: Learning Across Drifting Black-Box Problem Streams
Many real-world optimisation problems occur as streams of related instances. Transfer optimisation can exploit learned similarities between consecutive instances to improve solving speed or solution quality; however, these streams may change over time due to shifts in the underlying problem distribution, i.e. the stream is subject to data-drift. The ability to detect and respond to drift is crucial to optimise performance over the stream. However, there is currently no suitable stream generator on which to evaluate the aforementioned methods. We address this by proposing SCOBench: a fully configurable generator that can generate streams of instances from multiple functions with different drift patterns specified by the user. Drift between consecutive instances can be defined according to a change in three different metrics (landscape features, algorithm probing-trajectories and optimum location). The benchmark provides a principled way to evaluate the adaptability of algorithm selectors and configurators as well as transfer learning techniques.
Multisensor Recognition of Activities of Daily Living with PPIE-Informed Parkinson’s Protocol Design
Pharmacological treatment is commonly used in the management of Parkinson’s disease (PD), yet motor fluctuations related to medication intake remain difficult to monitor objectively in daily life. Current research predominantly evaluates medication effects during standardized walking tasks, which capture gait impairments but provide limited insight into how motor fluctuations impact functional independence. However, people with PD often report that medication-related changes are most evident during Activities of Daily Living (ADLs), such as dressing, personal hygiene, and object manipulation, as reflected in the PDQ-39 quality-of-life questionnaire. This study proposes a multimodal protocol to establish a baseline for automatic ADL recognition in 20 healthy young adults. Movement and interaction data will be collected using six Inertial Measurement Units (IMUs) placed on both wrists, foot and thigh, the lower back, and the chest, complemented by RGB cameras, radar sensing, and RFID tags for real-time tracking of household objects and a medication box. The resulting baseline will be used to train ADL classification models that can later be extended to monitor medication states in PD during everyday functional tasks. In parallel, a Patient and Public Involvement and Engagement (PPIE) activity was conducted with people living with PD (n=6), clinicians (n=3), and carers (n=3). This activity gathered feedback on protocol feasibility, the relevance of selected ADLs, and outcome measures including the PDQ-39. The findings informed protocol refinement, ensuring alignment with real-world needs and enhancing the clinical and practical relevance of the proposed approach.
Towards a Diversity-Aware Deepfake Video Dataset for Improving Model Gerneralizability
A deepfake is a form of synthetic media (e.g., video, image, or audio) generated using artificial intelligence to replace a person’s face or voice with that of another. While deepfake detection methods aim to mitigate potential misuse such as fake news and financial fraud, existing deepfake video benchmarks for training detection models are often created using outdated generation techniques and exhibit limited diversity under real-world conditions. This constrains the generalizability of deepfake detection models, particularly in cross-domain scenarios. To address this limitation, we propose a systematic framework for curating a diversity-aware deepfake video dataset. Specifically, we collect face videos from publicly available benchmarks and analyse attributes including temporal dynamics, video quality, and head pose. We will then sample diverse videos and employ nine state-of-the-art deepfake generation methods, such as SimSwap and InSwapper, to synthesize deepfake videos. The resulting benchmark is expected to support more robust and generalizable deepfake detection models.
Towards an Automated Assessment of The Bias Gap Between GEN AI and Real-World Data
With the growing use of AI in several sectors, Higher education like most industries is facing the increased adoption of AI in all its operational areas including, learning, research, teaching, assessment and feedback. Just as AI aims to improve or mimic human perception, thinking, applications and decisions, it stands to inherits human biases as well, without explicitly inheriting responsibility. This paradigm makes it important to ensure more consideration is given to AI adoption and the ethical implications of such an adoption. This systematic review aims to consolidate two independent categories of works carried out to review Generative AI case studies in higher education. The first category focuses on the frameworks for adopting AI in higher education. It aims to adapt the existing educational technology adoption frameworks to AI, providing key insights and necessary updates to the frameworks. The second categories focus on the ethics of AI in education and aim to highlight the core ethical principles as well as the unique applications to technology used in higher education. The aim of this review is to analyse these case studies and aim to highlight a gap in AI adoption where ethical considerations are not explicitly included in initial considerations for technology adoption nor in the assessment of the use of existing systems. This review analyses AI use cases in higher education and provides details on ethical consideration (if any) and catalogues biases that may exist as a result. It goes further to categorise these biases and provide suggested frameworks for including ethical assessment for generative AI adoption and assessment. The results from this review indicate that over 80% of use cases identified in higher education did not include an ethical or bias component. For those that did highlighted the gaps,They failed to mitigate the biases in thier current implementations of generative AI systems. This review will suggest an area of future work, establishing the baselines for a framework on Bias and ethical assessments of AI solutions as a core part of the system lifecycle
Input Matters: Evaluating Input Structure’s Impact on LLM Summaries of Sports Play-by-Play
A major concern when deploying LLMs in accuracy-critical domains such as sports reporting is that the generated text may not faithfully reflect the input data. We quantify how input structure affects hallucinations and other factual errors in LLM-generated summaries of NBA play-by-play data, across three formats: row-structured, JSON and unstructured. We manually annotated 3,312 factual errors across 180 game summaries produced by two models, Llama-3.1-70B and Qwen2.5-72B. Input structure has a strong effect: JSON input reduces error rates by 69% for Llama and 65% for Qwen compared to unstructured input, while row-structured input reduces errors by 54% for Llama and 51% for Qwen. A two-way repeated-measures ANOVA shows that input structure accounts for over 80% of the variance in error rates, with Tukey HSD post hoc tests confirming statistically significant differences between all input formats.
Conversational AI Without the Cloud: A Lightweight Local Dialogue Pipeline for Non-Commercial Social Robots
A major limitation of current social robots is their dependence on cloud-based dialogue pipelines, which restricts use in settings with limited or unreliable connectivity. We present a lightweight, fully local spoken-dialogue system that runs on consumer-grade hardware and integrates open-source models for speech recognition, dialogue generation, and text-to-speech. The pipeline was deployed on Euclid, a non-commercial humanoid robot, across several public engagement events, enabling extended real-world interaction without internet access. We analyse over 5,000 dialogue turns recorded during these dialogues to characterise system behaviour, user interaction patterns, and challenges arising in noisy, multi-speaker environments. Our observations demonstrate the feasibility of privacy-preserving, on-device conversational robotics while highlighting limitations in turn-taking, response length, and environmental grounding. We outline planned improvements to support more robust and accessible social-robot interaction.
Digital Literacy and Information Behaviour
Among High Blood Pressure Patients Using
Digital Tools for Self-Management
Digital health technologies are increasingly used to support the self-management of long-term conditions such as high blood pressure. However, access to digital tools does not necessarily translate into the ability to use them effectively. This research investigates how digital literacy shapes the ways in which adults with high blood pressure seek, evaluate, and use health information through digital tools and online resources. Grounded in Information Science, the study focuses on human information behaviour and digital literacy rather than clinical outcomes. Using a multi-phase mixed methods design, the research will combine a survey, diary study, and interviews/focus groups to explore digital literacy levels, information behaviours, and perceived preparedness for using digital health technologies. Recruitment will take place through community-based channels across the United Kingdom. The study aims to generate insights into barriers and enablers of digital health engagement, inform the design of more accessible digital health tools, and contribute to theory and practice at the intersection of digital literacy, information behaviour, and health self-management.
AI Assisted Tool for the Design of Immersive Game-Based Learning Applications
Immersive game-based learning (iGBL) offers powerful educational experiences. However, existing design frameworks are largely theoretical, which, although important, fail to address practical aspects to support the design of iGBL. This research addresses this gap by synthesizing 101 design guidelines from literature into a framework and an AI-assisted interactive tool for iGBL design. Analysis showed that guided decisions reduced uncertainty, and improved design process efficiency and clarity, while also shifting the focus from static frameworks to an interactive AI assistant. Future work will evaluate the tool with expert designers and students to assess usability and design efficacy in educational settings. This evaluation will inform how effectively the AI-assisted interactive tool bridges the gap between theoretical guidelines and practical, actionable iGBL design, empowering designers to build effective, pedagogically sound immersive experiences.
Classroom Engagement and Feedback Analysis Using Multimodal Emotion Recognition
Classroom engagement is a multidimensional construct that includes cognitive, emotional, and behavioural aspects. It is widely recognised as a cornerstone of effective learning, as higher levels of engagement are often associated with better academic performance and lower dropout rates. Traditionally, student engagement has been assessed through teacher observations, feedback, and surveys. However, in classrooms with large numbers of students, it can be challenging to monitor every student, particularly for those new to teaching. Our overall goal is to develop a multimodal classroom engagement detection framework that uses facial expressions, body movements, and classroom audio. We present a multimodal classroom engagement dataset collected during live lectures using a fixed camera setup, with privacy and ethical considerations carefully addressed. We then analyse engagement levels using state-of-the-art machine learning models. This dataset provides a valuable resource for future research on AI-driven classroom engagement analysis.
Non-Thermal Plasma Modelling in Novel Multiphase Reactor Systems
Non-thermal plasma (NTP) offers a pathway to energy-efficient chemical synthesis. It can enable thermodynamically demanding reactions, such as ammonia production, to be conducted at near-ambient temperature and pressure conditions. By combining NTP with industrially established fluidised bed reactors results in a novel gas-solid-plasma flow reactor. These reactors have significant potential for scalable and low-carbon process intensification. The modelling of such systems presents computational challenges. The coupled interactions between turbulent gas flow, granular solid mechanics, and plasma physics span length and time scales of various magnitudes. Examples include nanometre-scale surface charge transfer, nanosecond plasma discharge events and metre-scale reactor hydrodynamics. Therefore, this work develops a multi-scale continuum model that accurately captures the physical phenomena seen in these reactors. Preliminary results from our developed spatially homogeneous model provide promising insight into this development, showing that ammonia and precursors are successfully produced from.
Genesis: Genetic Evolution of Deep Neural Networks for Optimal Service Function Chain Embedding
Service Function Chains (SFCs) involve virtualising network functions, such as firewalls, embedding them on servers and linking them virtually using Software Defined Networking to form a chain of Virtualised Network Functions (VNFs). Optimally composing and embedding SFCs on physical networks is an NP-hard optimisation problem consisting of three sub-problems, namely optimal ordering of VNFs in an SFC, optimal embedding of VNFs on servers and optimal embedding of virtual links on physical links. Existing approaches do not optimise all three sub-problems in a scalable manner. We propose using a Genetic Algorithm (GA) to evolve three sine-activated Deep Neural Networks (DNNs) to optimise the three sub-problems. The outputs of the DNNs are used to order the VNFs in SFCs optimally, evolve a Gaussian distribution from which the servers to optimally embed VNFs on are randomly sampled, and predict the heuristic cost of network links so that VNFs can be linked optimally using the A* algorithm. We compare our approach with two other GAs and a greedy algorithm across 48 experiments and show that our approach converges on an optimal solution in all experiments while being the fastest GA approach, whereas the second-best approach converges in only 71% of the experiments.
Roll for Insight: Can TTRPG Mechanics be Leveraged for Persona Creation?
Personas are a useful tool in user experience (UX) design, allowing designers to understand what specific users need a system to do to be successful. However, there is no real consistency in how personas are designed, and what is required to have an effective one. Tabletop Role-Playing Games (TTRPGs) offer structured instructions and materials to aid players in building characters for gameplay. An exploratory study was carried out to establish the transferability of the character creation processes from games including Dungeons & Dragons and Daggerheart, to a persona creation framework as part of a TTRPG-UX environment.
Heritage in the Digital Age: Navigating Emerging Technologies for Small and Medium Museums
Small and medium museums (SMMs) face increasing pressure to modernise within a rapidly changing digital landscape. Emerging technologies, including extended reality (XR) and generative artificial intelligence (AI), can strongly support visitor engagement, education, and access to cultural heritage. However, existing research and implementations often focus on larger institutions, where funding, staffing, and technical capacity better support experimentation and maintenance. For SMMs, these technologies introduce critical challenges relating to constrained resources, expertise, infrastructure, and long-term sustainability, alongside wider concerns such as environmental impact. This PhD research investigates how SMMs can navigate emerging technologies in purposeful, sustainable, and contextually appropriate ways. Through a literature review, sector-wide survey, museum case study, and framework design and evaluation, the project will explore current challenges, priorities, and adoption practices. Ultimately, this project aims to deliver open-source framework enabling SMMs to assess, adopt, create, and sustain appropriate emerging technologies without requiring extensive technical expertise or resources.
Predicting Lakehouse Performance in Clouds: An Empirical Exploration of Query Runtime Variance
Distributed data lakehouses are increasingly used to support analytical query workloads across cloud and private infrastructure. These systems enable flexible data management and scalable query processing, but their performance can vary substantially across repeated executions. This creates challenges for query performance prediction (QPP), where models rely on observed runtimes to estimate future query behaviour. If runtime labels are unstable, prediction accuracy may degrade, limiting the reliability of downstream orchestration tasks such as workload placement, resource provisioning, and low-carbon scheduling. This research investigates runtime variance in distributed lakehouse deployments and examines how deployment conditions influence predictability. It further explores how variance affects query performance prediction models and the implications this has for carbon-aware orchestration. By treating runtime variance as an important signal rather than background noise, this work aims to support more reliable and sustainability-aware decision-making in modern data analytics infrastructure.
“Abuse risks are often inherent to product features”: Exploring AI Vendors Bug Bounty and Responsible Disclosure Policies
As vendors adopt AI technologies, security researchers are working to uncover and fix related vulnerabilities, which is important given AI systems handle sensitive data and critical functions. This process relies on vendors receiving and rewarding AI vulnerability reports. To assess current practices, we analyzed the vulnerability disclosure policies of 264 AI vendors. We employed a mixed-methods approach, combining snapshot and longitudinal qualitative analysis, as well as comparing alignment with 320 AI incidents and 260 academic articles. Our analysis reveals that 36% of AI vendors have no established policy, and only 18% mention AI risks. Data access, authorization, and model extraction vulnerabilities are most consistently declared in-scope. Jailbreaking and hallucination are most commonly declared out-of-scope. We identify three profiles that reflect vendors’ different positions toward AI vulnerabilities: proactive clarification (n = 46), silent (n = 115), and restrictive (n = 103). Our alignment results suggest that vendors may address AI vulnerability disclosure later than academic research and real-world incidents.
Can I Borrow Your Face? Authentic Characters Recreation in Heritage Scenarios
Characters are essential in heritage reconstructions, enriching authentic and dynamic representations of cultural elements in virtual environments. Elite communities and historical celebrities are frequently used, whilst ‘normal people’ from the past are rarely represented, due to a lack of historical and archaeological records. Inspired by this lack of authentic everyday people in heritage reconstructions, this study explores a collaborative design approach which engages ancestors of the Highland Clearances in Scotland, and places them into immersive historical landscapes. We invite these descendants to take part in the reconstruction of their ancestors lives, and ask ’can I borrow your face?’ to guide the design of their narratives and identities.
Physics-Informed Neural Networks for Accelerating Prodrug Delivery System Design
Hydrogel-based implants enable localised administration of prodrug therapeutics, offering significant advantages for targeted cancer treatment where inactive carriers circulate until activation at the tumour site. Despite their clinical promise, characterising the transport properties of these systems typically demands extensive empirical calibration, slowing the design cycle and limiting accessibility for research groups without specialised infrastructure. In our study, we developed an experimentally driven neural network that extracts transport phenomena directly from fluorescence imaging data and is constrained by the governing partial differential equations. Validated against Rhodamine B as a model prodrug, the framework recovers diffusion coefficients in good agreement with literature values. The model will eliminate the need for expensive instrumentation or extensive calibration, instead relying on readily obtainable imaging data from standard laboratory settings. This capability will accelerate the rational design and optimisation of hydrogel implants for clinical applications and support rapid screening of material formulations and transport behaviours.
Habitat: Hardware-Accelerated Binary Translator
Emulators have been essential in supporting legacy systems, future prototypes, and programs compiled for other architectures. However, software-based solutions often suffer from performance and scalability issues. This drastically limits the scope of possible emulation scenarios & raises the entry barrier for such technology. This project aims to explore potential improvements through hardware acceleration with broad support across multiple architectures, while remaining accessible to most existing systems. We achieve this by offloading the binary translation stage from the CPU using a PCIe FPGA dynamic binary translator (DBT). We will also introduce Architecture-Independent Representation (AIR), a unified format that encapsulates instruction set architectures (ISAs) (e.g., RISC-V, ARM). Currently, we have drafted the AIR and software decoder & AIR translator for RISC-V. We also have a DBT that can run statically linked C programs on Linux and perform arithmetic operations using standard I/O based on the translated AIR instructions.
Visualising Uncertainty with Icon Arrays: Communicating the Diagnostic Accuracy of a Cancer Screening Test
Icon arrays are widely used in medical risk communication to translate complex numerical information into simpler visual representations. Uncertainty, however, is rarely disclosed to non-expert audiences despite its pervasiveness in scientific and visualization research alike. This work examines how visualizing uncertainty impacts non-expert audiences’ perspectives by incorporating visualization techniques that encode uncertainty into icon arrays. We conducted a survey-based experiment where participants were randomly assigned to one of two uncertainty visualization stimuli groups (color bar or gradient), then assessed comprehension and subjective evaluations of trust and confidence. Participants in the color bar group showed significantly higher accuracy of understanding, while confidence and trust remained consistently moderate across participants.