Research Papers

Audits Under Resource, Data, and Access Constraints: Scaling Laws For Less Discriminatory Alternatives

Sarah H. Cen, Salil Goyal, Zaynah Javed +3

AI audits play a critical role in AI accountability and safety. One branch of the law for which AI audits are particularly salient is anti-discrimination law. Several areas of anti-discrimination law implicate the "less discriminatory alternative" (LDA) requirement, in which a protocol (e.g., model) is defensible if no less discriminatory protocol that achieves comparable performance can be found with a reasonable amount of effort. Notably, the burden of proving an LDA exists typically falls on the claimant (the party alleging discrimination). This creates a significant hurdle in AI cases, as the claimant would seemingly need to train a less discriminatory yet high-performing model, a task requiring resources and expertise beyond most litigants. Moreover, developers often shield information about and access to their model and training data as trade secrets, making it difficult to reproduce a similar model from scratch. In this work, we present a procedure enabling claimants to determine if an LDA exists, even when they have limited compute, data, information, and model access. We focus on the setting in which fairness is given by demographic parity and performance by binary cross-entropy loss. As our main result, we provide a novel closed-form upper bound for the loss-fairness Pareto frontier (PF). We show how the claimant can use it to fit a PF in the "low-resource regime," then extrapolate the PF that applies to the (large) model being contested, all without training a single large model. The expression thus serves as a scaling law for loss-fairness PFs. To use this scaling law, the claimant would require a small subsample of the train/test data. Then, the claimant can fit the context-specific PF by training as few as 7 (small) models. We stress test our main result in simulations, finding that our scaling law holds even when the exact conditions of our theory do not.

AI-generated health misinformation poses unprecedented threats to patient safety and healthcare system trust globally. This white paper presents an explainable AI framework developed through the EPSRC INDICATE project to combat medical misinformation while enhancing evidence-based healthcare delivery. Our systematic review of 17 studies reveals the urgent need for transparent AI systems in healthcare. The proposed solution demonstrates 95% recall in clinical evidence retrieval and integrates novel trustworthiness classifiers achieving 76% F1 score in detecting biomedical misinformation. Results show that explainable AI can transform traditional 6-month expert review processes into real-time, automated evidence synthesis while maintaining clinical rigor. This approach offers a critical intervention to preserve healthcare integrity in the AI era.

Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by atypical brain maturation. However, the adaptation of transfer learning paradigms in machine learning for ASD research remains notably limited. In this study, we propose a computer-aided diagnostic framework with two modules. This chapter presents a two-module framework combining deep learning and explainable AI for ASD diagnosis. The first module leverages a deep learning model fine-tuned through cross-domain transfer learning for ASD classification. The second module focuses on interpreting the model decisions and identifying critical brain regions. To achieve this, we employed three explainable AI (XAI) techniques: saliency mapping, Gradient-weighted Class Activation Mapping, and SHapley Additive exPlanations (SHAP) analysis. This framework demonstrates that cross-domain transfer learning can effectively address data scarcity in ASD research. In addition, by applying three established explainability techniques, the approach reveals how the model makes diagnostic decisions and identifies brain regions most associated with ASD. These findings were compared against established neurobiological evidence, highlighting strong alignment and reinforcing the clinical relevance of the proposed approach.

As AI rapidly advances, the security risks posed by AI are becoming increasingly severe, especially in critical scenarios, including those posing existential risks. If AI becomes uncontrollable, manipulated, or actively evades safety mechanisms, it could trigger systemic disasters. Existing AI safety approaches-such as model enhancement, value alignment, and human intervention-suffer from fundamental, in-principle limitations when facing AI with extreme motivations and unlimited intelligence, and cannot guarantee security. To address this challenge, we propose a Governable AI (GAI) framework that shifts from traditional internal constraints to externally enforced structural compliance based on cryptographic mechanisms that are computationally infeasible to break, even for future AI, under the defined threat model and well-established cryptographic assumptions.The GAI framework is composed of a simple yet reliable, fully deterministic, powerful, flexible, and general-purpose rule enforcement module (REM); governance rules; and a governable secure super-platform (GSSP) that offers end-to-end protection against compromise or subversion by AI. The decoupling of the governance rules and the technical platform further enables a feasible and generalizable technical pathway for the safety governance of AI. REM enforces the bottom line defined by governance rules, while GSSP ensures non-bypassability, tamper-resistance, and unforgeability to eliminate all identified attack vectors. This paper also presents a rigorous formal proof of the security properties of this mechanism and demonstrates its effectiveness through a prototype implementation evaluated in representative high-stakes scenarios.

Time-series forecasting models (TSFM) have evolved from classical statistical methods to sophisticated foundation models, yet understanding why and when these models succeed or fail remains challenging. Despite this known limitation, time series forecasting models are increasingly used to generate information that informs real-world actions with equally real consequences. Understanding the complexity, performance variability, and opaque nature of these models then becomes a valuable endeavor to combat serious concerns about how users should interact with and rely on these models' outputs. This work addresses these concerns by combining traditional explainable AI (XAI) methods with Rating Driven Explanations (RDE) to assess TSFM performance and interpretability across diverse domains and use cases. We evaluate four distinct model architectures: ARIMA, Gradient Boosting, Chronos (time-series specific foundation model), Llama (general-purpose; both fine-tuned and base models) on four heterogeneous datasets spanning finance, energy, transportation, and automotive sales domains. In doing so, we demonstrate that feature-engineered models (e.g., Gradient Boosting) consistently outperform foundation models (e.g., Chronos) in volatile or sparse domains (e.g., power, car parts) while providing more interpretable explanations, whereas foundation models excel only in stable or trend-driven contexts (e.g., finance).

Explainable Artificial Intelligence (XAI) has become an essential component of medical imaging research, aiming to increase transparency and clinical trust in deep learning models. This study investigates brain hemorrhage diagnosis with a focus on explainability through Class Activation Mapping (CAM) techniques. A pipeline was developed to extract pixellevel segmentation and detection annotations from classification models using nine state-of-the-art CAM algorithms, applied across multiple network stages, and quantitatively evaluated on the Hemorica dataset, which uniquely provides both slice-level labels and high-quality segmentation masks. Metrics including Dice, IoU, and pixel-wise overlap were employed to benchmark CAM variants. Results show that the strongest localization performance occurred at stage 5 of EfficientNetV2S, with HiResCAM yielding the highest bounding-box alignment and AblationCAM achieving the best pixel-level Dice (0.57) and IoU (0.40), representing strong accuracy given that models were trained solely for classification without segmentation supervision. To the best of current knowledge, this is among the f irst works to quantitatively compare CAM methods for brain hemorrhage detection, establishing a reproducible benchmark and underscoring the potential of XAI-driven pipelines for clinically meaningful AI-assisted diagnosis.

As Vision-Language Models (VLMs) achieve widespread deployment across diverse cultural contexts, ensuring their cultural competence becomes critical for responsible AI systems. While prior work has evaluated cultural awareness in text-only models and VLM object recognition tasks, no research has systematically assessed how VLMs adapt outputs when cultural identity cues are embedded in both textual prompts and visual inputs during generative tasks. We present the first comprehensive evaluation of VLM cultural competence through multimodal story generation, developing a novel multimodal framework that perturbs cultural identity and evaluates 5 contemporary VLMs on a downstream task: story generation. Our analysis reveals significant cultural adaptation capabilities, with rich culturally-specific vocabulary spanning names, familial terms, and geographic markers. However, we uncover concerning limitations: cultural competence varies dramatically across architectures, some models exhibit inverse cultural alignment, and automated metrics show architectural bias contradicting human assessments. Cross-modal evaluation shows that culturally distinct outputs are indeed detectable through visual-semantic similarity (28.7% within-nationality vs. 0.2% cross-nationality recall), yet visual-cultural understanding remains limited. In essence, we establish the promise and challenges of cultural competence in multimodal AI. We publicly release our codebase and data: https://github.com/ArkaMukherjee0/mmCultural

In user-centric design, persona development plays a vital role in understanding user behaviour, capturing needs, segmenting audiences, and guiding design decisions. However, the growing complexity of user interactions calls for a more contextualized approach to ensure designs align with real user needs. While earlier studies have advanced persona classification by modelling user behaviour, capturing contextual information, especially by integrating textual and tabular data, remains a key challenge. These models also often lack explainability, leaving their predictions difficult to interpret or justify. To address these limitations, we present ExBigBang (Explainable BigBang), a hybrid text-tabular approach that uses transformer-based architectures to model rich contextual features for persona classification. ExBigBang incorporates metadata, domain knowledge, and user profiling to embed deeper context into predictions. Through a cyclical process of user profiling and classification, our approach dynamically updates to reflect evolving user behaviours. Experiments on a benchmark persona classification dataset demonstrate the robustness of our model. An ablation study confirms the benefits of combining text and tabular data, while Explainable AI techniques shed light on the rationale behind the model's predictions.

Probability Density from Latent Diffusion Models for Out-of-Distribution Detection

Joonas Järve, Karl Kaspar Haavel, Meelis Kull

Despite rapid advances in AI, safety remains the main bottleneck to deploying machine-learning systems. A critical safety component is out-of-distribution detection: given an input, decide whether it comes from the same distribution as the training data. In generative models, the most natural OOD score is the data likelihood. Actually, under the assumption of uniformly distributed OOD data, the likelihood is even the optimal OOD detector, as we show in this work. However, earlier work reported that likelihood often fails in practice, raising doubts about its usefulness. We explore whether, in practice, the representation space also suffers from the inability to learn good density estimation for OOD detection, or if it is merely a problem of the pixel space typically used in generative models. To test this, we trained a Variational Diffusion Model not on images, but on the representation space of a pre-trained ResNet-18 to assess the performance of our likelihood-based detector in comparison to state-of-the-art methods from the OpenOOD suite.

This paper proposes a Trans-XFed architecture that combines federated learning with explainable AI techniques for supply chain credit assessment. The proposed model aims to address several key challenges, including privacy, information silos, class imbalance, non-identically and independently distributed (Non-IID) data, and model interpretability in supply chain credit assessment. We introduce a performance-based client selection strategy (PBCS) to tackle class imbalance and Non-IID problems. This strategy achieves faster convergence by selecting clients with higher local F1 scores. The FedProx architecture, enhanced with homomorphic encryption, is used as the core model, and further incorporates a transformer encoder. The transformer encoder block provides insights into the learned features. Additionally, we employ the integrated gradient explainable AI technique to offer insights into decision-making. We demonstrate the effectiveness of Trans-XFed through experimental evaluations on real-world supply chain datasets. The obtained results show its ability to deliver accurate credit assessments compared to several baselines, while maintaining transparency and privacy.

This paper is a collaborative piece between two worlds of expertise in the field of data visualization: accessibility and bias. In particular, the rise of generative models playing a role in accessibility is a worrying trend for data visualization. These models are increasingly used to help author visualizations as well as generate descriptions of existing visualizations for people who are blind, low vision, or use assistive technologies such as screen readers. Sighted human-to-human bias has already been established as an area of concern for theory, research, and design in data visualization. But what happens when someone is unable to verify the model output or adequately interrogate algorithmic bias, such as a context where a blind person asks a model to describe a chart for them? In such scenarios, trust from the user is not earned, rather reliance is compelled by the model-to-human relationship. In this work, we explored the dangers of AI-generated descriptions for accessibility, playing a game of telephone between models, observing bias production in model interpretation, and re-interpretation of a data visualization. We unpack ways that model failure in visualization is especially problematic for users with visual impairments, and suggest directions forward for three distinct readers of this piece: technologists who build model-assisted interfaces for end users, users with disabilities leveraging models for their own purposes, and researchers concerned with bias, accessibility, or visualization.

We introduce fCrit, a dialogue-based AI system designed to critique furniture design with a focus on explainability. Grounded in reflective learning and formal analysis, fCrit employs a multi-agent architecture informed by a structured design knowledge base. We argue that explainability in the arts should not only make AI reasoning transparent but also adapt to the ways users think and talk about their designs. We demonstrate how fCrit supports this process by tailoring explanations to users' design language and cognitive framing. This work contributes to Human-Centered Explainable AI (HCXAI) in creative practice, advancing domain-specific methods for situated, dialogic, and visually grounded AI support.