Hutiri argued that the AI industry has spent years acknowledging these concerns while continuing to rely heavily on scraped datasets assembled without meaningful consent. The result is a credibility gap. Systems meant to evaluate fairness often depend on data collected through the same practices they are supposed to scrutinize.

Against that backdrop, Hutiri introduced the Fair Human-centric Image Benchmark, or FHIBE, a globally diverse, consensually collected dataset for diagnosing bias in computer vision systems. Built from more than 10,000 images of nearly 2,000 subjects across 81 countries, FHIBE is designed to show that ethical data collection is not merely an aspiration but an operational discipline.

The details matter. Hutiri described the work required to secure explicit informed consent, support consent revocation, and provide fair compensation. In doing so, she framed ethical data not as a compliance exercise, but as infrastructure. If evaluation benchmarks are to measure harm, they cannot themselves be built on exploitative labor or unauthorized collection.

The session then widened to the generative AI economy, where creators face a more existential challenge. Current protections, including opt-out mechanisms and tools such as robots.txt, remain inadequate for an ecosystem in which creative work can be absorbed, transformed, and monetized at scale.

Hutiri’s central argument was not that ethical AI data is easy, or fully solved. It was that the alternative is untenable. If AI systems are to be trusted, the industry must move beyond abstract commitments and build mechanisms for consent, attribution, compensation, and revocation into the data supply chain itself.