Human Judgment Fails Where AI Safety Depends on It

Dataset of stories from 140 webpages documenting Missing and Murdered Indigenous Relatives (MMIR) cases. Designed to resist epistemic erasure by amplifying Indigenous voices and providing researchers with community-centered narratives about systemic, technological, and institutional barriers affecting Indigenous peoples across Turtle Island.

Wi-Fi backscatter-based silent speech interface supporting open-vocabulary sentence recognition via lexicon-guided subword decoding. Achieves 85.61% word prediction accuracy and 36.87% WER on continuous sentence recognition. Uses frequency-shifted backscatter tag with cluster-based self-supervision and Transformer encoder-decoder with beam search for variable-length sequences.

Visual-illusion-inspired privacy preservation system for mobile data visualizations. Automatically generates privacy-preserving visualizations by stacking decoy visualizations over originals, adjusting visual channels (shape, position, tilt, size, color, spatial frequency). Models human perception at different viewing distances to protect against shoulder surfing attacks.

Interactive system supporting mobile UI design through example-driven workflow. Powered by multimodal retrieval-augmented generation (MMRAG) model enabling iterative search, selection, and adaptation of examples at global (whole interface) and local (component) levels. Includes source transparency cues (ratings, download counts, developer information) to foster trust.

Fine-grained content-based visual permission system for AR glasses using group-based control mechanism. Detects objects using YOLO with pre-classified schema grouping by privacy sensitivity, category, or spatial proximity. Achieves mAP50 of 0.6704 with 14.0 ms latency and 1.7% battery consumption increase per hour.

Empirical dataset establishing base measurements for visual target localization tasks in visualizations with up to hundreds of objects. Provides comparisons between visual variables (hue, size, etc.) and their robustness across target location and visual arrangement dimensions. Available at https://osf.io/z68ak/overview with all supplemental materials.

Interactive system enabling programming instructors to track students' AI usage in code-writing processes. Allows instructors to create personalized assessments, detect AI use conflicting with pedagogical goals, and provide timely interventions within the workflow of monitoring coding histories. Enables accurate detection of when students require intervention.

Systematic survey of 171 visualization and humanities (VIS*H) design studies analyzing evaluation workflows and rigor. Documents recurring evaluation flaws, monomethod approaches, and quality patterns. Provides evidence for triangulation strategies and validation criteria for humanities visualizations reconciling analytical utility with interpretive depth.

Annotated dataset from audit of 123 major companies' privacy policies identifying 282 potential instances of jurisdiction-siloed disclosure across 77 companies (62.6%). Conservative estimate finds 138 instances across 54 companies (44%) where substantive data practices appear only in regional compliance sections. Released with methodology to enable replication.

Labeled dataset from online survey with 74 participants (51 parents, 23 teens) responding to simulated cybergrooming scenarios. Documents four types of vulnerable responses and four types of protective strategies with stage-based taxonomy showing progression as risk escalates. Provides teen-centered understanding for educational programs and sociotechnical interventions.