Keynote Speakers

Workshop Organizers

Accepted Papers

Posters:

1. From Techne Equalisation to Phronesis Premium: Generative AI and Seniority Bias in Ageing Europe

   • Dora Moscato

   • Contrary to the prevailing view that older workers are disadvantaged by technological change, this paper argues that generative AI may advantage them at the expense of younger ones. The mechanism operates through two channels. First, generative AI equalises procedural competence (techne), enabling lower-performing workers to reach the quality of top performers—an executional prosthesis that recovers skills depreciated with age. Second, as execution becomes cheap, the scarce factor shifts to contextual judgment and responsible decision-making (phronesis), capabilities that correlate with reflective experience rather than technical currency. An event study on Eurostat EU-LFS data (2008–2025) shows that the youth employment share in high-AI-adoption sectors diverges sharply downward after ChatGPT’s diffusion, following fifteen years of no statistically significant differential trend, while the senior share expands. The seniority bias carries significant ethical implications: the devaluation of entry-level work erodes the pipeline through which future judgment is formed, while the advantage itself is structurally transitory, shaped by the design choices of AI producers, the organisational strategies of adopting firms, and the institutional responses of regulators. The paper examines the EU AI Act’s human oversight requirement as institutional demand for phronesis and argues that its effectiveness depends on whether oversight remains substantive or degrades into formal compliance.

2. Hybrid LSTM Framework For Vessel-Type-Aware AIS Anomaly Detection

   • Raj Singh, Moirangthem Suchitra Devi, Sandeep Kumar, Gwyneth A Chullai

   • Automatic Identification System (AIS) data plays an important role in large-scale maritime traffic monitoring and anomaly detection. Deep learning models, particularly Long Short-Term Memory (LSTM) networks, are commonly used to model vessel trajectories due to their ability to capture temporal dependencies. However, many existing approaches rely on unified global models trained across different vessel categories, often assuming similar motion patterns. In practice, cargo ships, tankers, and fishing vessels show distinct behaviors, which can affect detection performance. In this work, we examine how vessel-type differences influence AIS-based trajectory anomaly detection using a controlled experimental setup. We compare three modeling strategies under the same architectural settings: (i) a global LSTM model with full parameter sharing, (ii) vessel-type-specific models with full specialization, and (iii) a hybrid shared-backbone model that learns common temporal patterns while keeping vessel-type specific output layers. Experiments on real-world AIS data with controlled anomaly injection show that incorporating vessel-type information improves detection performance. The hybrid model reduces false alarm rates (by about 2–3% compared to the global model) and also improves precision and recall. It further reduces detection latency across all vessel types which allows anomalies to be detected earlier. While fully vessel-type-specific models achieve the best overall performance, the hybrid model offers a practical balance between generalization and specialization, without increasing model complexity. Overall, these results suggest that accounting for vessel-type differences is important for building more reliable AIS anomaly detection systems, and that hybrid modeling provides an effective and scalable solution.

3. A Three-Layer Threat Assessment Pipeline for Human-Robot Interaction Safety

   • Sanaullah Sanaullah, Han Byung-Kil, Alexandra Dmitrienko, Hirotada Honda, Thorsten Jungeblut, Dong Il Park

   • Ensuring safety in human-robot interaction (HRI) environments requires the ability to detect not only physical threats but also subtle indicators of psychological distress — a challenge that existing single-modality systems fail to address comprehensively. This paper presents MindGuard, a three-layer threat assessment pipeline designed for real-time safety monitoring in HRI contexts, uniquely combining mental health risk detection with physical threat recognition within a unified framework. The first layer employs keyword pattern matching and a fine-tuned BERT classifier to identify linguistic indicators of self-harm, suicidal ideation, and violent intent from natural language input. The second layer leverages YOLO-based object detection and a Vision-Language Model (VLM) for continuous scene analysis, identifying physical threat objects and hazardous environmental conditions through live camera feeds. The third layer fuses all signals — textual, semantic, and visual — through a large language model reasoning agent that produces a structured risk decision across five escalation levels: SAFE, MONITOR, RESPOND, ALERT, and EMERGENCY. The system operates in real time on consumer GPU hardware using FP16 inference, achieving low-latency detection suitable for deployment on mobile robotic platforms. Experimental evaluation demonstrates that the multi-modal fusion approach reduces false negatives compared to text-only or vision-only baselines, with the cross-modal signal agreement improving overall classification confidence. MindGuard addresses a critical gap in HRI safety systems by treating mental health crises and physical dangers as equally urgent threat categories, offering a deployable, extensible architecture for socially aware autonomous robots.

Oral Presentations:

1. Mitigating Demographic Bias Decay in Machine Learning Models

   • Skanda Sunil, Chen Zhao

   • Real-world machine learning systems operate under temporal distribution shifts, where fairness degrades over time, causing bias decay. Current methods and approaches consist of either static fairness-aware learning or inline learning methods, which adapt to evolving data but fail to incorporate fairness considerations. As a result, these methods fail to maintain proper fairness under distributional change. To bridge this issue, we propose a novel framework, Temporal Threshold Optimization (TTO). The key idea of TTO is to optimize predictive performance and fairness across time. We evaluated TTO based on real-world datasets, which consisted of financial complaints, police incidence, and news based market data, each exhibiting variability and degradation in fairness over time. TTO constantly has achieved competitive accuracy while maintaining low and stable disparity levels across these shifts. All in all, these findings highlight the limitations of static fairness assumptions and highlight the importance of TTO for reliable machine learning systems to reduce demographic bias.

2. Conformal Multiverse Analysis: Auditing Uncertainty-Aware Prediction Pipelines

   • Julia Broden, Jan Simson, Christoph Kern

   • We introduce Conformal Multiverse Analysis (CMA), an auditing framework which integrates Conformal Prediction with Multiverse Analysis. CMA enables a comprehensive assessment of how uncertainty of prediction models propagates across different modeling choices and affects group-conditional coverage of different groups and minorities. Pairing Multiverse Analysis with Conformal Prediction provides a novel procedure for identifying disparities and critical design decisions, particularly in contexts where uncertainty-aware decision policies are aimed to be deployed. We illustrate CMA by comparing 900 modeling pipelines for an algorithmic profiling task and show that (1) design decisions can affect uncertainty outcomes for subgroups and (2) uncertainty reduction and group-specific coverage can be misaligned. Our work reinforces the need and value of integrating uncertainty quantification into fairness-aware model design and evaluation.

3. SignCV: A Sign-Consensus Vector Probe for Model-Dependent Multi-Attribute Bias Geometry

   • Kyungmin Kwon, SeungHun Han, HaYoung Oh 

   • Direction-based bias mitigation methods commonly assume that bias-associated parameter directions for different demographic attributes can be merged into a single shared edit. We show this assumption holds to sharply different degrees across LLMs matched in scale (7B) and release window. Under an identical extraction pipeline on BBQ, Mistral-7B-Instruct-v0.3 and Llama-2-7B-hf exhibit nearly identical within-attribute update stability (within-axis survivor rates of ≈64.0% vs. ≈67.3%), yet their cross-attribute alignment differs by nearly an order of magnitude: only 7.05% of coordinates survive race–gender–SES sign intersection on Mistral, versus 53.48% on Llama—a 7.6× gap. We surface this contrast with SignCV, a sign-consensus vector probe that extracts per-attribute LoRA update directions, retains coordinates with stable within-axis sign agreement, and measures how many survive cross-axis sign intersection; the cross-axis survivor rate quantifies, before any edit is applied, whether a unified multi-attribute projection is structurally plausible for a given checkpoint. Downstream projection edits at 𝑘 = 1.0 yield modest, model-dependent outcomes rather than uniform improvement: on Mistral, Race bias decreases (0.0307→0.0245) while SES and Gender move adversely; on Llama, SES and Gender decrease slightly while Race drifts marginally upward. Evaluation PPL decreases across all three single axes for both models, but we do not interpret this as a fairness gain. Ablations show Full SignCV is not uniformly best on aggregate metrics; its value is diagnostic. We therefore position SignCV as a pre-edit geometry check that reveals when—and on which models—a shared multi-attribute projection direction is structurally plausible.

4. Curation Cards: Structuring Decisions Made During Human-LLM Collaborations for Data Curation

   • Shreya Chappidi, Andra Krauze, Jat Singh

   • Early-stage decision-making during the algorithm development lifecycle, including problem formulation and data labelling, critically shape downstream processes including algorithmic performance and user adoption. Human-LLM collaborations during data mining bring multiple new dynamics to this early-stage decision-making as users can more explicitly shape data annotation and labeling practices. First, users may not understand how design choices made during data curation could alter or shape their intended problem formulations. Second, the non-deterministic nature of LLMs means that small (and often implicit) choices during data curation such as prompt style or phrasing can meaningfully influence the distribution of resulting data labels. As a result, there can be limited visibility into how user choices shape LLM-curated data, which may result in misaligned datasets that do not effectively support the original problem intent, create unreliable or biased data labels, and add complexity during audits of human-LLM data workflows. To support early-stage decision-making on how to design LLM-supported data curation workflows, we propose curation cards—a structured documentation approach that captures purpose/goals of data curation, data input and model specifications, labeling schema and concept definitions, system prompts, reference sources, prompt text and non-semantic details, pipeline design, and error analysis. These elements reflect key early-stage design choices influencing data curation outcomes, where small variations in aspects such as prompt design or schema definition can cascade into systematic shifts that meaningfully influence data. In all, curation cards seek to support AI practitioners by (1) explicitly considering critical factors that will shape dataset creation and curation, (2) clarifying user choices and decisions made during LLM-supported curation, and (3) improving reviewability and auditability of curation processes. 

5. When Equity Looks Like Concentration: A Placebo Protocol for Fairness Claims in Recommender Systems

   • Vairaaj Bindal

   • Subgroup-level fairness claims in ML often get confused with descriptive concentration. An aggregate effect can be large in a sub-group because the subgroup is large, even when the subgroup-defining feature carries no targeting information at all. We propose a stratified label-shuffle placebo that tells the two apart, and we apply it to a real recommender for smallholder fertilizer and crop-switch advice trained on 8,744 plots from the LSMS-ISA Ethiopia panel. The recommender maximizes Expected Realized Benefit (predicted yield gain weighted by predicted adoption probability) and beats an accuracy-only baseline by 50 kg/ha on 800 enumeration- area-disjoint test plots, with the largest absolute lift on the no education stratum (+53 kg/ha, 𝑛 = 543/800). On its face that looks like equity targeting. The placebo says otherwise: per-subgroup means fall inside the null under random education-label assignment (𝑝 ∈ [0.14, 0.87]). The aggregate clears Rosenbaum Γ★ ≥ 5, EA cluster permutation 𝑝 < 5 × 10−4, and a SAFI-anchored partial identification bound at +26 kg/ha worst case. The per-subgroup gradient is just descriptive concentration. We argue the placebo should be a default in any subgroup-stratified fairness report.

6. A Task-Agnostic Metric for Adversarial Vulnerability in Vision-Language Models

   • Maisha Binte Rashid, Pablo Rivas, Chen Zhao

   • Vision-language models (VLMs) are highly vulnerable to adversarial perturbations, yet there is no unified metric for measuring their robustness across different tasks and architectures. In this paper, we propose a task-agnostic Vulnerability Score (VS) that quantifies robustness degradation under both structured noise and adversarial perturbations. The metric combines noise-induced and FGSM-based performance degradation into a single interpretable score. We evaluate VS on CLIP, BLIP, and LLaVA using five benchmark datasets spanning classification, captioning, visual question answering, and image-text retrieval tasks. Results show that VS consistently reflects adversarial susceptibility across models and tasks. Further validation with PGD and Carlini & Wagner (C&W) attacks demonstrates strong inverse correlations between VS and adversarial performance, confirming the reliability of the proposed metric. Analysis with Jacobian-based sensitivity measures shows that VS captures broader aspects of practical robustness beyond local gradient sensitivity.

7. Hyperbolic Graph Learning for Hierarchical Cognitive Diagnosis in Intelligent Education Systems

   • Han Xiang, Haodong Qian, Shengpeng Wang, Lingyun Liu, Dongran Yu, Xingcheng Fu 

   • Cognitive diagnosis estimates student and group knowledge mastery from response logs, yet educational graphs contain implicit and heterogeneous hierarchies over groups, students, exercises, and concepts. Existing Euclidean or single-space models distort deep hierarchy or conflate student-behavior and knowledge-semantic structures. We propose Hyperbolic Graph Learning for Hierarchical Cognitive Diagnosis (HGCD), which builds a heterogeneous educational graph, learns two adaptive-curvature hyperbolic spaces for student and knowledge hierarchies, and fuses them through a geometry-consistent common tangent space. Experiments on four real datasets show consistent gains for individual- and group-level diagnosis. 

Workshop Reviewers

• Ram Prasad Nethi (Amazon Web Services)

• Venkata Ratna Kumar Bonagiri (Macys)

• Ankur Gupta (LinkedIn)

• Shatrughna Upadhyay (Intuit Inc)

• Ankur Bhatnagar (Macys)

• Shahazad Qurashi (Jazan University)

• Dong Li (Baylor University)

• Jinyang Li (Amazon)

• Zhao Zhu (Meta Platforms)

• Xinyu Wu (Baylor University)

• Denglin Jiang (Bloomberg)

• Jiaxin Liu (Meta Platforms)

• Boyang Li (University of Notre Dame)

• Zhuosheng Liu (UC Davis)

• Xiaohui Chen (Baylor University)

Previous Workshops

• The 1st Workshop on Ethical Artificial Intelligence: Methods and Applications, KDD 2022, Washington D.C., USA

• The 2nd Workshop on Ethical Artificial Intelligence: Methods and Applications, KDD 2023, Long Beach, California, USA

• The 3rd Workshop on Ethical Artificial Intelligence: Methods and Applications, KDD 2024, Barcelona, Spain

• The 4th Workshop on Ethical Artificial Intelligence: Methods and Applications, KDD 2025, Toronto, Ontario, Canada