The Quiet Part Out Loud: AI Isn’t Neutral, So Let’s Stop Pretending
The lights come up on an off-Broadway black-box theater. A young UX designer named Maneesh is demoing a slick mapping interface to his fictional employer, Athena. The audience chuckles—until a post-show talkback reveals that Athena is Palantir in thin disguise and the “college side-project” on screen is now an ICE raid-planning engine. Gasps ripple through the cheap seats. If theatergoers can spot the surveillance twist in ninety minutes, why does the ad-tech industry still insist its algorithms are ethically agnostic?
When the Playbook Becomes the Play
Matthew Libby’s Data isn’t subtle, and that’s the point. Maneesh’s arc—from idealistic coder to reluctant enabler—mirrors the lived reality of many engineers who once thought they were building ad-targeting widgets, not deportation pipelines. During the talkback, Libby hammered the takeaway: “Well-intentioned work can still enable harm.” Swap “college predictive model” for “look-alike seed pool” and the story lands inside every major DSP. The same probabilistic scoring that decides who sees a sneaker ad can, with a few pivots, decide who gets a knock on the door at 5 a.m.
From Bid Stream to Border Patrol
ICE’s shopping list reads like a programmatic spec sheet: mobile ad IDs, lat/long within five meters, IP-derived household data, cross-device graphs. 404 Media obtained 2022 RFPs where the agency asked for “ad-tech-style data enrichment” to “enhance operational planning.” The winning stack reportedly ingests bid-stream lat/long into a Palantir-built mapping layer that scores likely undocumented status. Same data points used to sell you sneakers—now used to sell deportations. The industry consolidation we keep applauding for “efficiency” has quietly removed the friction that once slowed these pivots from commerce to control.
Privacy Sandbox ≠ Ethics Sandbox
Chrome’s third-party cookie funeral marches on, pushing buyers into Google’s privacy-safe cohorts. But “privacy-safe” only means identity-diffused, not bias-free. When a cohort ID correlates 92 % with lower-income Hispanic ZIPs, the new toolchain can reproduce red-lining at machine-learning velocity. As one buyer told AdExchanger last month: “Algorithms may not create bias, but they scale it at warp speed.” In the aggregated world of Privacy Sandbox, impression-level audit trails vanish; bias incidents become statistically invisible, leaving buyers with plausible deniability and zero incentive to fix the model.
Consolidation = Amplification
Remember when twenty DSPs meant a biased algorithm could only burn through a fraction of open-web spend? Those days are gone. Five global buying platforms now touch roughly 90 % of U.S. programmatic budgets. When one of them ships a model that under-values Black-owned media or over-targets predatory lending offers, the blast radius is the entire internet. Smaller players historically acted as bias circuit-breakers; their disappearance removes the last speed bump between a bad model and a billion impressions.
Measurement in the Dark
Post-cookie measurement relies on aggregated conversion APIs that deliberately obscure user-level data. That’s great for privacy, terrible for accountability. Without impression-level logs, how do you prove a housing-ad campaign excluded protected classes? You can’t—you can only infer it from skewed cohort composition, and even that data is locked behind Google’s noised reporting. The industry’s answer, effectively, is “trust us.” The same industry that still claims, with a straight face, that AI is neutral.
Stop Hiding Behind the Black Box
The play Data ends with Maneesh deleting his repository, but the real world doesn’t offer that clean exit. Palantir contracts stretch for years; ICE keeps expanding its data appetite; DSPs keep merging. The only honest path forward is to drop the pretense of neutrality. Every model encodes human choices—what data to ingest, what success metric to optimize, what threshold to call a “good” user. Those choices are never bias-free, and the bigger the platform, the larger the moral multiplier.
So the next time a vendor slides a deck across the table promising “algorithmic neutrality,” push back. Ask for the audit trail, the fairness report, the third-party civil-rights review. Because AI isn’t a magic wand; it’s a powerful and sometimes dangerous tool, and the curtain is already falling on the era when we could pretend otherwise.
💡 Deep Dive: Don’t miss our Ultimate Industry Guide for advanced strategies.
