Sony Sets a New Standard for Ethical AI

Sony Sets New Benchmark for Ethical AI with Fair Human-Centric Image Benchmark

Sony AI has launched a groundbreaking dataset called the Fair Human-Centric Image Benchmark (FHIBE), designed to test the fairness and bias of artificial intelligence (AI) models. This dataset is being hailed as the “first publicly available, globally diverse, consent-based human image dataset for evaluating bias across various computer vision tasks.” It aims to determine how equitably AI models treat individuals, revealing that no existing dataset from any company fully meets the necessary benchmarks.

FHIBE addresses ethical issues and bias challenges in the AI industry, featuring images of nearly 2,000 volunteers from over 80 countries, all sharing their likenesses with explicit consent. This contrasts sharply with the common and often contentious practice of scraping vast amounts of web data without permission. Participants in the FHIBE project can also remove their images at any time, ensuring their autonomy.

The research affirmed existing biases in today’s AI models, noting that some produced lower accuracy for individuals identified by “she/her/hers” pronouns. FHIBE pointed out that hairstyle variability significantly influenced results, highlighting factors previously overlooked in AI assessments. Disturbingly, the dataset also found that AI models reinforced stereotypes when asked neutral questions about occupations, disproportionately skewing representations against certain pronoun and ancestry groups.

Sony AI believes that FHIBE demonstrates the feasibility of ethical and fair data collection. The dataset is now publicly available and will undergo regular updates, with research findings published in Nature. This initiative sets a new benchmark for ethical AI, and its broader implications could lead to fairer, more responsible AI technologies.

Join the conversation about Sony’s innovative approach to ethical AI. How do you think datasets like FHIBE will shape the future of artificial intelligence?