Sony has a new benchmark for ethical AI (www.engadget.com)

Sony AI today released the Fair Human-Centric Image Benchmark (FHIBE, “Phoebe”), which it calls the first publicly available, globally diverse, consent-based human image dataset for evaluating bias across a wide range of computer-vision tasks. FHIBE contains images of nearly 2,000 volunteers from 80+ countries who explicitly consented to inclusion and can withdraw at any time. Each image is richly annotated — demographic and physical attributes, environmental context and camera settings — enabling controlled, fine-grained evaluation of fairness. Sony also reports that no existing dataset from other organizations fully met its benchmark standards and published the work in Nature. Technically, FHIBE both confirmed known failure modes and revealed new drivers of bias. Tests showed lower accuracy for people using “she/her/hers” pronouns and highlighted hairstyle variability as a previously underappreciated factor; models also produced stereotyped or toxic outputs when asked about occupations or crimes, disproportionately affecting people of African or Asian ancestry, darker skin tones and certain pronoun groups. By offering consented, diverse, and richly labeled images, FHIBE provides a practical template for diagnosing, measuring and mitigating bias in model development, evaluation and policy — a usable step toward more ethical data practices and targeted fairness interventions.