Amazon Is Deploying AI to Spy on Its Workers and Bust Unions Before They Form

A Truthout piece by Jonathan Rosenblum, a union organizer, part-time Amazon delivery driver and activist in residence at Arizona State University’s Center for Work and Democracy. His material is a set of internal Amazon documents disclosed in proceedings before British Columbia’s Labour Relations Board, after Unifor won certification at YVR2, the warehouse near Vancouver, and the board found that Amazon had interfered in the union drive.

The documents describe Atlas, a framework that scores Amazon sites by how likely they are to organize. It takes in hundreds of data points: training completion, safety incidents, tenure and turnover, local wage levels, warehouse temperatures, and how close a site is to others that are organizing. Workers’ own words are inputs too. Posts on the internal Voice of the Associate board, answers to daily surveys and threads on Reddit are mined for keywords and sentiment, so managers can see who is driving a discussion. Organizing itself is logged as “employee relations events”, from card signings to walkouts, each with a severity score. The output is a labour score in five tiers, from Tier 5 (no immediate risk) to Tier 1 (activity imminent), tracked month by month. In 2024, 16 warehouses were at Tier 1, among them LDJ5 on Staten Island, near the JFK8 warehouse that unionized in 2022. Sites at Tiers 1 and 2 get Amazon’s rapid-response team. The documents also track organizing in the English Midlands, Melbourne, Poland, Germany and Italy. Amazon did not respond to Truthout. Rosenblum frames Atlas as Bentham’s panopticon rebuilt with software, and the fight to organize Amazon as one over freedom and democracy as much as pay.

The labour-board facts hold: Unifor won certification in July 2025, and on 4 August 2026 the board ordered a first contract settled by binding arbitration. Benjamin Y. Fong had reported on the same documents in Jacobin two weeks earlier, on 16 September. His account describes the score as a “Labor Continuity Score” from 100 to 900 and defines the tiers a little differently, with Tier 1 as sites where cards are already being signed. On the headline’s two claims, the “AI” is clearest in the analysis of worker feedback, which Fong says is run through machine-learning models; the scoring itself, as described, could be a weighted scorecard. And “before they form” describes Atlas’s purpose. Neither piece shows whether Tier 1 sites were in fact targeted, or whether drives there failed.