Bias - Harm annotation dataset

This dataset contains of a CSV file of aggregated annotations derived from a vignette-based study examining how different types of AI bias relate to various categories of AI harm across two domains: disease diagnosis and hiring. Each vignette describes a specific scenario involving an AI system. For each scenario, the study considers: • a domain-specific context anchor• a bias type (e.g., representation, measurement, algorithmic, evaluation, deployment),• a stakeholder affected (e.g., individual applicant, applicant group, individual patient, patient group),• a set of harm categories , based on an adaptation of Shelby et al.’s taxonomy of allocative, quality-of-service, interpersonal, and representational harms in MCQ format. Human participants selected the harms they believed were most likely to occur in each scenario (SUMS). A large language model (LLM) received the same scenarios and selected harms as well. This dataset aggregates both sources of annotations. All data are fully anonymised. The dataset does not include raw survey responses or any personally identifiable information. Study design (brief) Human participants (e.g., university students) read short vignette-based scenarios describing AI systems in hiring and disease diagnosis domain. For each scenario and stakeholder, participants selected one or more harms they believed were most likely to occur under specific bias. A large language model was given the same scenarios and asked to select harms under comparable instructions. The dataset includes only aggregated data: human vote counts and LLM selections. No raw questionnaire data are included (IDs, full questions, etc). Ethics and privacy No direct or indirect identifiers are contained in the dataset.Because individual-level responses are not included, re-identification is not expected to be feasible.Users remain responsible for ensuring ethical and legally compliant reuse.