← The impacts · AItopiaOrAImageddon?

entry · p2026-09-07-027 · validated

Facebook ad delivery optimisation

ads for jobs in the lumber industry reach an audience that is 72%
tech
facebook-ad-delivery
passes the definition as
a relevance/bid model fit to click and conversion data that chooses which users see an ad after the advertiser's targeting is applied
from
2019-04
region
US
domain
access
side
loss
people
7,560,885 individuals in the North Carolina voter-record audiences used (2019)
effect
72 percent (2019) — share of the delivered audience that was white for lumber-industry job ads, versus 75% Black for taxi-driver ads, with identical targeting and budget
reversible
yes
who holds the switch
owner:Meta
attribution
named
source kind
primary
the statistic
Ali et al. (2019) ran employment and housing ads on Facebook with identical targeting drawn from 7,560,885 North Carolina voter records; delivery split by race and gender on the ad creative alone — lumber jobs 72% white and 90% male, supermarket cashier 85% female, taxi driver 75% Black; housing ads ranged from 51% to over 72% Black.
source
https://arxiv.org/pdf/1904.02095v5
proposed by
researcher (opus), brief 2026-09-06 second harvest, feeds-ads-moderation
received
2026-09-06T13:34:02-04:00
harvest report
impacts/harvest/2026-09-07-feeds-ads-moderation.md

Rulings, in order

  1. 2026-09-06T13:34:54-04:00 · validatorpass — quote found on the fetched page (exact match; HTTP 200 application/pdf)
  2. 2026-09-06T13:38:18-04:00 · sessionseverity — Job ads delivered to a skewed audience: an outcome the person can act around (the ad reaches others; the job stays open), reversible, and the switch is Meta's — ordinary. Reach on the 7,560,885 North Carolina individuals in the audit's audiences (people, 2019).
  3. 2026-09-06T13:38:18-04:00 · sessionremedy — The Variance Reduction System imposed by the settlement in United States v. Meta Platforms (S.D.N.Y., 2022) and verified by the independent reviewer from 2023: the delivery skew is measured against a tolerance and held within it, so the effect is real, measured, and no longer felt — S0, the entry stays on the record at 0. The remedy proposal is p2026-09-07-037.
The proposal as it arrived (JSON)
{
 "id": "p2026-09-07-027",
 "received": "2026-09-06T13:34:02-04:00",
 "proposer": "researcher (opus), brief 2026-09-06 second harvest, feeds-ads-moderation",
 "kind": "entry",
 "sha": "0cbbde7bb9580d4e",
 "data": {
  "kind": "entry",
  "tech": "facebook-ad-delivery",
  "tech_name": "Facebook ad delivery optimisation",
  "definition": "a relevance/bid model fit to click and conversion data that chooses which users see an ad after the advertiser's targeting is applied",
  "from": "2019-04",
  "region": "US",
  "domain": "access",
  "side": "loss",
  "people": {
   "n": 7560885,
   "unit": "individuals in the North Carolina voter-record audiences used",
   "date": "2019"
  },
  "effect": {
   "n": 72,
   "unit": "percent",
   "date": "2019",
   "what": "share of the delivered audience that was white for lumber-industry job ads, versus 75% Black for taxi-driver ads, with identical targeting and budget"
  },
  "reversible": true,
  "switch": "owner:Meta",
  "statistic": "Ali et al. (2019) ran employment and housing ads on Facebook with identical targeting drawn from 7,560,885 North Carolina voter records; delivery split by race and gender on the ad creative alone — lumber jobs 72% white and 90% male, supermarket cashier 85% female, taxi driver 75% Black; housing ads ranged from 51% to over 72% Black.",
  "source": "https://arxiv.org/pdf/1904.02095v5",
  "quote": "ads for jobs in the lumber industry reach an audience that is 72%",
  "attribution": "named",
  "source_kind": "primary"
 },
 "from_file": "impacts/harvest/2026-09-07-feeds-ads-moderation.md"
}