Domain Atlas / Hiring & employment screening AI
The account goes dark at nine; the reason arrives on day twenty-six
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 2 assumed · 8 published baseline.
Amazon Flex, a first-party last-mile delivery programme launched in September 2015, rates its United States contract driver fleet into four standing tiers — Fantastic, Great, Fair and At Risk — computed from arrival punctuality at delivery stations, completion of routes inside the reserved block window, compliance with customer special requests, and delivery-quality signals. A Bloomberg investigation published on 28 June 2021, read here through the syndication carrying its full text, reports that algorithms scan incoming performance data and decide which drivers get more routes and which are deactivated, that human feedback is rare, and that terminations arrive by automated email; it interviewed fifteen drivers, four of whom said they were wrongly terminated, together with former Amazon managers and a former engineer, and documents deactivations following circumstances the input set cannot represent — locked apartment gates on predawn routes, malfunctioning lockers, hour-long waits at understaffed stations, a nail in a tire, snowbound rural roads and failed selfie identity checks. Former insiders told the investigation that the programme's benefits far outweigh the collateral damage and that the company decided it was cheaper to trust the algorithms than to pay people to investigate mistaken firings. Amazon disputes the characterisation: its spokesperson called the driver accounts anecdotal and unrepresentative and said the company has invested heavily in technology and resources to provide drivers visibility into their standing and eligibility to continue delivering, and investigates all driver appeals. No court or regulator has ever ruled on how standing is computed or on whether any particular deactivation was substantively correct, and no scoring internals, error rate, termination rate or reversal rate has been published by anyone. Scale figures are the operator's own and count downloads rather than active drivers — approximately 4 million globally and 2.9 million in the United States, with more than 660,000 in one quarter of 2021, up about 21 percent year over year; the only hard official count in this record is the 140,128 individual drivers a federal regulator paid tip refunds to for a single 2016-to-2019 window.[5]
What happened
A driver opens an app, reserves a four-hour block, drives to a delivery station, waits for a cart of packages, and delivers them in their own car. From the moment they arrive at the station the app is measuring: where the phone is and when, whether it is moving at driving speed or walking speed or running, whether each stop was made, whether the block finished inside its window, whether a customer's special request was followed. Before the shift there is a selfie identity check to pass.
Those measurements become a standing. Amazon Flex, launched in September 2015, places drivers in one of four named tiers — Fantastic, Great, Fair and At Risk — and the tier decides which blocks a driver is offered next. At a threshold it ends the account. A Bloomberg investigation published on 28 June 2021 reported the mechanism from documents and insiders: algorithms scan incoming performance data for patterns and decide which drivers get more routes and which are deactivated, human feedback is rare, and the termination arrives as an automated email. Amazon disputes the characterisation. Its spokesperson called the driver accounts anecdotal and unrepresentative and said the company has invested heavily in technology and resources to give drivers visibility into their standing and eligibility to continue delivering, and that it investigates all driver appeals. Both positions belong in the record and both are carried here.
The investigation's specific finding is not that the arithmetic miscomputes. It is that the input set is blind to cause. The drivers it followed lost their accounts after a locked apartment gate on a predawn route, a malfunctioning locker, an hour-long wait at an understaffed station, a nail in a tire, a snowbound rural road, a selfie check that failed. There is no field anywhere in the measurement where the reason for a lost hour could be written down, so the hour arrives as the driver's shortfall. Stephen Normandin, 63, had delivered in Phoenix for close to four years and had been asked to help train others. Neddra Lira, 42, was deactivated from a Great standing in Arlington, Texas, and later had her car repossessed. Ryan Cope, 29, lost his account after driving snowbound roads outside Denver, and gave the investigation its title: it is you against the machine, so you do not even try. Fifteen drivers were interviewed; four said they were wrongly terminated. Former Amazon managers and a former engineer were interviewed too, and two of the things they said are the reason this case is on the atlas at all: that the programme was a great success whose benefits far outweigh the collateral damage, and that the company decided it was cheaper to trust the algorithms than to pay people to investigate mistaken firings. A stated tolerance for wrong outcomes, attached to employment-consequential automation, is rare on any public record. It is an insider characterisation, the operator disputes it, and no figure of any kind stands behind it.
Now look at the channel that is supposed to catch the mistake. A deactivated driver has ten days to appeal, by email. The first reply arrives the next day, reads as machine-generated, and is typically generic rather than specific to the incident; it is signed with a support agent's first or full name. A six-day review is promised and often exceeded. In the two cases the investigation follows to the end, the final answer lands on day 26 and on day 11. There is no pay in the interim. The one escalation the terms then provided beyond email was individual arbitration with a $200 filing fee, against roughly $80 of net pay on a typical four-hour block. The investigation's sentence on that is the whole mechanism in one line: drivers pay $200 to take their dispute to arbitration, but few do, seeing it as a waste of time and money. A Harvard Law School labour-law commentary called the fee an unrealistic prospect for workers barely making minimum wage and described the ten-day window as an appeal to a bot. Drivers also report that standing takes months to recover from a delay they could not avoid, while standing keeps gating the offer of work. One consolidating registry summarises this case as drivers having no opportunity to contest at all; that overstates it, and the honest description is a channel that exists and is automated, slow and priced.
Everything else in this record is adjacent, and the adjacency is the point. The Federal Trade Commission settled with Amazon on 2 February 2021 for $61,710,583 — the exact amount withheld — over a variable-base-pay scheme that used customer tips to cover promised base pay between late 2016 and August 2019 while drivers and customers were told all tips passed through; internal emails called it a huge public-relations risk and a reputation tinderbox, the Commission voted 4-0, and in November 2021 it paid 140,128 drivers an average of $422, with a top payment above $28,000. That is the only hard official count of affected drivers anywhere in this record. The District of Columbia pursued the same conduct under its consumer-protection statute, noting when it sued in December 2022 that the federal resolution had produced no civil penalties, and settled on 7 February 2025 for $3.95 million. Wisconsin's workforce-development department audited more than a thousand Flex drivers and found the vast majority employees for unemployment-insurance purposes; that determination was upheld in 2023 and left standing when the state supreme court dismissed Amazon's appeal on 26 March 2024. New Jersey's labour department sued in October 2025 over the same classification. In 2020 the Ninth Circuit held in Rittmann that Flex drivers are transportation workers exempt from the Federal Arbitration Act, so their wage claims could go to court — a holding about the wage-claims clause, not about the deactivation channel, which continued to route through arbitration as reported in 2021. Every one of those outcomes concerns pay, tips or classification. Not one of them touches how a tier is computed.
What changed is smaller and more precise than a lawsuit. On 1 January 2025 Seattle's App-Based Worker Deactivation Rights Ordinance took effect, and it regulates the deactivation channel itself. A covered network company — 250 or more app-based workers worldwide, which includes Amazon Flex — must publish a deactivation policy reasonably related to safe and efficient operations. It must give fourteen days' notice before most deactivations, excepting egregious misconduct or where the law requires immediacy. The notice must state the reasons and the specific incidents and hand over all records relied on and considered. The company must investigate the alleged violation and show it more likely than not before deactivating, apply penalties consistently and in proportion to the violation, and consider the circumstances of the work. And the worker gets an internal challenge procedure they may invoke within ninety days, with a private right of action afterwards. In the City's published resolved-investigations record for October to December 2025, its Office of Labor Standards resolved two deactivation matters against Amazon Logistics, Inc. doing business as Amazon Flex, each returning $1,245.70 to one worker and requiring Amazon Flex to restart that worker's deactivation process — with equivalent resolutions the same quarter against three other network companies, so this is ordinary enforcement rather than a landmark. The same quarter records a much larger and quite separate settlement: $3,777,924.10 plus $20,000 in fines, covering 10,968 Amazon Flex workers, over gig-worker premium pay and paid sick and safe time. Amazon settled without admitting liability and said it strongly disagrees with the office on the facts of the matter.
On 4 March 2026 the ordinance survived its first constitutional test. In Uber Technologies, Inc. v. City of Seattle, a Ninth Circuit panel of Judges Graber, Clifton and Bennett affirmed the denial of a preliminary injunction sought by Uber and Instacart, holding that the ordinance regulates nonexpressive conduct — the unwarranted deactivation of worker accounts — that any burden on speech is incidental, that any compelled disclosure would in any event be commercial speech surviving Zauderer review, and that "reasonably related to safe and efficient operations" is not unconstitutionally vague. Judge Bennett dissented in part, agreeing on vagueness but viewing the policy-disclosure requirement as compelled speech. It is a preliminary-injunction affirmance rather than a final merits judgment, and Amazon was not a party.
So the record now contains something rare: one automated decision running under two correction regimes at the same time. A driver whose work sits inside one city gets fourteen days, the reasons, the specific incidents, the records, an investigation held to a preponderance standard, and a challenge that costs nothing. A driver two counties away gets an automated email and a $200 door. The scorer is identical. Only the price of contesting it, and the duty to explain it, differ by geography. And the layer nobody has reached is the same in both places: the ordinance's own enforcement is confined to procedural compliance until 1 June 2027, so even the one authority that touches this channel is barred from asking whether the deactivation was right.
The sociotechnical reading
Read this case for the timing, not for the score.
The measurement writes fast. Every stop, every minute, every scan updates a standing, and the standing gates the next block. The correction is slow: ten days to appeal, a promised six-day review often exceeded, answers on day 26 and day 11, no pay in between. Put those two clocks side by side and the deployment's behaviour follows without needing any claim about the algorithm's accuracy at all. A fast automatic write against a slow human correction produces a system whose errors accumulate faster than they can be found — and whose errors, if the investigation is right about their cause, are systematically of one type: circumstances outside the driver's control, recorded as the driver's shortfall, because the input set has no field for a locked gate.
The second structure is the price on the correction link. The Lab draws contest channels as pathways, and a pathway's width is how much actually flows along it, not whether it exists on paper. Here the escalation beyond email carried a $200 fee against roughly $80 of net pay per block, and the reporting says plainly that few drivers used it. That is a channel drawn narrow not because the operator refused it but because its price sets its exercise rate — and an unexercised correction channel discovers nothing. Former insiders describe that as a costed choice rather than an accident; the operator says the opposite. What is not in dispute is that no appeal-outcome rate and no reversal rate has ever been published by anybody, so nobody outside can say what the channel achieves.
The third structure is classification as governance stripping. Because the drivers are engaged as contractors, none of the ordinary scaffolding around losing a job attaches to the event: no unemployment machinery, no wrongful-termination standard, no anti-discrimination process. Institutional labour research enumerates exactly what the classification removes. And the same classification is losing, repeatedly, in adjacent forums — a state unemployment-insurance determination final against the operator, a state suit pending, tens of thousands of arbitrations outstanding, and a federal appeals court holding that these workers are exempt from the arbitration statute. So the ground under the whole regime is contested in real time while the regime runs.
The fourth structure is the one that makes this case unusual in this atlas: a jurisdictional seam in the correction link. Since January 2025 the same scorer has run against two different procedures depending on where the work was performed. In one city the operator must give notice before the act, state the reasons and the specific incidents, hand over the records relied on, investigate to a preponderance standard first, and provide a free challenge. Everywhere else the reporting record still describes an automated email and a fee. That is close to a natural experiment in what a priced correction link does versus a mandated one — and it comes with a warning attached, because the regulated layer is procedure, not the decision rule. The instrument the legislature reached for was disclosure and process. It did not cap a deactivation rate, require any standard of accuracy, or say anything at all about how a tier is computed, and it deferred substantive review until mid-2027.
Which leaves the finding that organises the rest. Every authority that has ever acted on this deployment has acted on something adjacent to the score: the tips, the pay, the sick time, the employment classification, the arbitration clause, the notice, the records, the challenge. Not one has looked at the rule. The scoring internals have never been published. No error rate, termination rate or reversal rate exists. No independent evaluation of this deployment has ever been made. On the diagram that appears as a single pathway drawn at zero — an outside read of the rule — and it is drawn at zero on a stated absence rather than an unexamined one, which is a different and much stronger claim.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.