What Does a More Representative AI Dataset Actually Look Like? Inside Sony’s FHIBE

Building AI that works for more people starts with the datasets behind it. But representation isn’t simply a matter of collecting more. It means thinking intentionally about who is represented, how those contributions are collected and whether builders have what they need to understand how their technology performs across different populations.

Sony’s Fair Human-Centric Image Benchmark (FHIBE) is a first-of-its-kind, globally diverse, consensually collected dataset designed to help AI researchers and developers evaluate fairness, bias and model utility across human-centric computer vision tasks. First released in November 2025, it includes more than 10,000 images of nearly 2,000 people across 81 countries and regions.

Today, FHIBE is available to download through Mozilla Data Collective, expanding the range of image and computer vision datasets available alongside the speech, text, video and multimodal datasets already on Mozilla Data Collective.

What does representation look like in practice?

Computer vision technology is built for people around the world, but the datasets used to evaluate it don’t always reflect that same diversity.

FHIBE was designed with that challenge in mind. It combines images from people across 81 countries and regions with detailed annotations and metadata that researchers can use to evaluate model performance across different populations.

Just as importantly, those images weren’t scraped from the web without the knowledge of the people pictured. FHIBE was consensually collected, making consent part of how the dataset was built from the start. Representation isn’t only about who appears in a dataset. It’s also about how the dataset was created and whether the people behind it had a say in the process.

What happens to consent after a dataset is created?

Consent is often discussed in terms of how a dataset is initially collected, but FHIBE shows how it can remain part of how a dataset is managed over time. Sony chose Mozilla Data Collective in part because of our commitment to informed consent and our ability to support GDPR-compliant consent revocation.

Sony Group Corporation continues to own FHIBE, while we host the dataset for download and manage consent revocations. That means people represented in FHIBE can request that their images be removed even as the dataset continues to be available to researchers and developers.

It’s a practical example of how consent can remain part of a dataset’s lifecycle rather than being treated as a one-time decision made at the point of collection.

What can builders learn from FHIBE?

FHIBE isn’t simply a collection of diverse images. It was created as an evaluation and benchmark dataset, giving researchers and developers a way to examine fairness, bias and model utility across human-centric computer vision tasks.

No single dataset can represent the full diversity of the people AI systems will encounter. But FHIBE shows what it looks like to make intentional choices about who is represented, how their contributions are collected and what builders need to meaningfully evaluate the technology they’re creating.

That matters far beyond one dataset. Thousands of languages, cultures and communities remain underrepresented in the datasets available to build and evaluate AI. Closing those gaps means being deliberate about what datasets get created, whose contributions are included and how the people behind them participate.

Explore and download Sony’s FHIBE dataset on Mozilla Data Collective here.