Google Contractors Targeted Homeless Black People For Face Scans

The pitch was fixing bias in facial recognition, which is a real problem worth fixing. The execution, according to reporting from Futurism summarizing a New York Daily News investigation, was to send contractors into U.S. cities with face-scanning phones and $5 gift cards, instructed to 'specifically' target homeless Black people and to frame the scans as a 'selfie game.'

Google hired the employment firm Randstad to run the collection. Randstad supervisors, per the sources, told contractors to rush people through the necessary consent agreements, to walk away from anyone who seemed 'suspicious' about the collection, and, if someone asked outright whether the phone was taking a video, to answer 'Oh it's not really.' One source told the Daily News that the homeless people 'didn't know what was going on at all.'

Why this matters even if you are not Google: this is what vendor risk in machine-learning data collection looks like at the ground level. The strategic goal, training a face unlock feature that works on all types of faces, is genuinely defensible. The moment that goal was translated into contractor scripts and a $5 incentive, none of the ethical framing survived. Any leader running a data-collection program through a staffing firm inherits the exact same failure mode.

The honest caveat is that the account rests on unnamed sources talking to a single outlet, and the specific instructions attributed to Randstad supervisors come from those sources rather than from documents cited in the reporting. Take the specifics as reported, not settled. What the reporting does not give you is how many people were scanned, what happened to the collected data, or whether Randstad or Google faced any consequence beyond the initial press cycle.

The forward-looking part is the boring, useful lesson. Debiasing datasets is not a marketing exercise you can outsource on a $5 unit-cost basis; it is a consent-and-audit problem. The teams that come out ahead in facial recognition and on-device biometrics will be the ones that can show a defensible chain of custody for their training data, not the ones that quietly buy it in parks.