Ground truth is a process, not a dataset

Mozilla is flipping the script on how AI gets trained. Instead of feeding models one massive dataset and calling it a day, they're building ground truth as an ongoing process — something that evolves as data gets used, questioned, and corrected.

The idea is simple but rarely practiced: datasets rot. They get stale, biased, or just wrong. Treating ground truth as a living system means catching errors, updating labels, and keeping the data honest over time. It's the difference between building a house once and maintaining it.

Why this matters for us: if AI keeps eating stale data, it keeps serving us stale answers — especially about the communities and languages we rely on.