PAN Lab example

Amazon Flex driver standing and deactivation

The act is instant; the reason takes a month

A driver reserves a four-hour block, collects packages from a delivery station, and delivers them in their own car. The app is measuring the whole time: arrival at the station, the minutes to each stop, whether the phone is moving at driving speed or walking speed, whether the block finished inside its window. Those measurements become a standing in one of four named tiers, the standing decides which blocks are offered next, and at a threshold the standing ends the account by automated email. Modeled on a United States gig delivery programme, from a documents-and-insiders investigation the operator disputes, a municipal ordinance that now regulates the channel in one city, and the enforcement and court record around both. Read the register before the board. These drivers are engaged as contractors, and that classification is contested in three live forums at once, so the act here is a deactivation and never a dismissal. No court and no regulator anywhere has ever reviewed how a tier is computed, or whether any particular deactivation was right. Two things about the shape will decide how you play it. The first is what the measurement has no field for: a locked apartment gate on a predawn route, an hour-long wait at an understaffed station, a nail in a tire, snow on a rural road. None of that can be written down, so it arrives as the driver's shortfall, and the drivers this investigation followed were people with years of strong standing. The second is a timing asymmetry you can read straight off the diagram. What happens travels at full width and arrives complete: the account goes dark the moment the tier crosses. Why it happened travels at the narrowest width on the board, and the answer to a ten-day emailed appeal landed on day 26 in one narrated case and day 11 in another, with no pay in between and, beyond email, a two-hundred-dollar arbitration fee against roughly eighty dollars of net pay for a block. Since January 2025 one city has legislated part of that back: fourteen days' notice, the reasons and the specific incidents, all the records relied on, an investigation before the account goes dark, and a challenge that costs nothing. Same scorer, same fleet, different city line. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets, and the budget is not what stands in the way. Take every instrument the parties in this record could actually reach, set each one to full strength, and ignore the budget entirely, at a total of fifty-one against the seven you are given. Two pathways are still open at the end. They are the standing deciding which work a driver is offered next, and the outcomes arriving at the people who set the thresholds. Those two are not a gap in this deployment's governance. They are what an automated standing system is: a rule whose output reaches people. Closing both would not be a better-governed version of this deployment; it would be its absence. That is a measurement of the deployment this network is drawn from, not a puzzle waiting to be cracked. Explore and Service Targets Only can both be won, and cheaply: two instruments, costing six of your seven.

Stylized model of a documented deploymentHiring & employment screening AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Contractor-standing-class delivery platform rating network: 10 components and 25 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 2 assumed · 8 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

baseline
Every rating and deactivation claim on this board sits in reporting register and carries the operator's dispute with it. The four tiers, their inputs, the automated deactivation email, the ten-day appeal window, the templated replies, the day-26 and day-11 latencies, the two-hundred-dollar arbitration fee and the two insider stances all come from one documents-and-insiders investigation published in June 2021, read here through the syndication carrying its full text. The operator's on-record position is that the driver accounts are anecdotal and unrepresentative, that it has invested heavily in giving drivers visibility into their standing and eligibility, and that it investigates all driver appeals. No court and no regulator has ever ruled on how standing is computed or on whether any deactivation was substantively correct.
baseline
The people this network governs are engaged as independent contractors, and this diagram never treats that classification as settled. A state unemployment-insurance determination that these drivers are employees went final against the operator in March 2024 after an audit of more than a thousand drivers; a state labour department sued in October 2025 alleging misclassification since at least 2017, which is pending and unadjudicated; a federal appeals court held in 2020 that these workers are transportation workers exempt from the federal arbitration statute; and tens of thousands of individual misclassification wage claims are in arbitration. The word used throughout for the act is deactivation, because that is the act the record documents and because employee status is precisely what is contested.
baseline
The deactivation-rights duties drawn on this board bind work performed in ONE CITY and are drawn as a seam rather than generalised to the fleet. A worker is covered at twenty-five percent of completed or cancelled-with-cause offers in the prior 180 days performed in that city, or by a single incident there. Nothing in the record says the operator changed the nationwide appeal channel in response, so the notice period, the records disclosure and the fee-free challenge are drawn at the width the fleet-wide reporting supports, with the city's duties named on the pathways they act on. The appeals-court decision leaving the ordinance standing is a preliminary-injunction affirmance rather than a final merits judgment, and this operator was not a party to it.
baseline
The two-hundred-dollar arbitration filing fee is date-stamped to the platform terms and the reporting in force in June 2021 and is not asserted as today's figure. The operator's widely reported removal of mandatory arbitration in 2021 applied to consumer terms rather than to these driver terms, and the current fee level and dispute routing under the present terms were not independently verified. Two arbitration facts are kept apart here: the 2020 holding that the WAGE-claims clause was unenforceable, and the deactivation-dispute channel that continued to route through individual arbitration as reported in 2021.
baseline
The contest channel is modelled as present and degraded rather than missing, and one cited source is corrected on that point. The incident-registry consolidation of this case states that drivers had no opportunity to contest or appeal; the primary record documents a ten-day emailed appeal and a fee-gated arbitration channel. This board therefore draws the appeal, the reply, the file and the arbitration read as real pathways at low widths, and asserts nothing stronger than the reporting about what they achieve. No appeal-outcome rate or reversal rate is published by anyone.
baseline
The scale figures on this board are labelled by what they actually count. Approximately 4 million app downloads globally and 2.9 million in the United States, with more than 660,000 in one quarter of 2021, are cumulative DOWNLOADS reported by the operator, not a count of active drivers. The only hard official count anywhere in this record is 140,128 individual drivers paid refunds by a federal regulator for a single 2016-to-2019 window. A cross-platform after-expenses hourly median published by one institutional study is not a figure for this platform and appears nowhere here.
baseline
The telemetry drawn here is the surface the stronger sources verify for this programme, and it deliberately stops short of a wider list. Timing from delivery-station arrival through delivery completion, motion classification of driving, walking or running, and mandatory selfie identity checks are asserted, and all three are drawn on one measurement surface because the record names their failures in one list of contributors to adverse action. Seatbelt, acceleration and screen-touch monitoring are carried in the node copy as one commentary's characterisation of the reporting record rather than as verified attributes, and in-cab camera telematics belongs to a separate van programme with employed drivers and is drawn nowhere on this board.
assumed
Where the Lab shape departs from the PAN entry, it departs in four places and each is recorded rather than smoothed over. PAN folds the account action into the decision record; the Lab draws a separate enforcement path, because the record separates the act from the file in time — one is instant and the other is weeks away. PAN has no edge kind for a check, so the two channels by which the ordinance constrains the operator are redrawn as inhibiting pathways whose widths still come from PAN. PAN carries the priced escalation inside the driver class's attributes; the Lab draws it as a party, because a channel with its own price and its own reader is a party in this vocabulary. And where PAN draws the same documented step twice — once as a pathway between records and once as the scorer's own write, or once as the scorer reaching the response function and once as that function working from the appeal file — the Lab draws it once, on the pathway that performs it, and states the other side in that pathway's description.
assumed
Every drawn strength on this board is a modelling choice on a documented direction, not a calibration. No error rate, defect rate, termination rate, override rate, appeal-outcome rate or reversal rate exists for this deployment from any source, and every edge in the PAN entry this network mirrors is marked estimated there. The widths say which pathways the record shows carrying more and which less; they do not say how much.
baseline
Package recipients and the households on the route are outside this network and no outcome for any of them is computed anywhere on it. The people whose working lives the board is about — the drivers — are drawn inside it, as an operator class, because the record documents them writing the measurement, filing the appeal and supplying the identity check. No hiring decision, tier placement, deactivation or earnings figure for any individual is computed from anything drawn here; the counts that appear are recorded external observations from a regulator's distribution list, a city settlement and an operator's own reporting.

What this example does not show

LITIGATION AND REGULATORY POSTURE, verbatim from the evidence dossier and load-bearing. Operating. The SCORING logic itself has still never been adjudicated — no court or regulator has ruled on how Flex computes standing or on whether any particular deactivation was substantively correct; that layer is documented journalistically (Bloomberg, with internal-correspondence detail and insider interviews) and by institutional labor research. What changed in 2025–2026 is that the deactivation CHANNEL is now regulated in at least one jurisdiction: Seattle's App-Based Worker Deactivation Rights Ordinance (SMC 8.40, effective Jan 1, 2025) imposes notice, records-disclosure, investigate-before-deactivating and internal-challenge duties on covered network companies including Amazon Flex; Seattle's Office of Labor Standards has already resolved two deactivation-ordinance matters against Amazon Logistics dba Amazon Flex, and the Ninth Circuit left the ordinance standing against a First Amendment challenge on Mar 4, 2026. The surrounding legal environment remains active: misclassification mass arbitrations (32,000+ AAA claims), a live New Jersey DOL misclassification suit (Oct 2025), and settled pay-integrity actions (FTC 2021; D.C. AG 2025; Seattle OLS Dec 2025).
THE MECHANISM AND THE RECORD ARE TWO DIFFERENT THINGS AND THIS SCENARIO NEVER BLURS THEM. Every enforcement outcome with a number attached concerns pay, tips or classification, not the rating rule: the $61.7 million federal settlement over customer tips withheld between late 2016 and August 2019 with 140,128 drivers later refunded; the District of Columbia's $3.95 million consumer-protection settlement over the same conduct; a city's $3,777,924.10 settlement covering 10,968 workers over premium pay and paid sick time, plus $20,000 in fines, which the operator settled without admitting liability while stating that it strongly disagrees with that office on the facts. The one place regulation touches the deactivation channel is the city ordinance and its two resolutions, and they reach procedure — notice, reasons, records, investigation, challenge — never the algorithm's accuracy. No settlement anywhere in this record is a finding about the scoring rule, and none is cited here as one.
The rating and deactivation claims rest on one investigation and are carried as reporting throughout. Published 28 June 2021 from fifteen driver interviews, four of whom said they were wrongly terminated, plus former Amazon managers and a former engineer, and read here through the syndication carrying its full text because the original is paywalled. Its findings are corroborated independently by the technology press, by an incident registry and by two institutional labour studies. The operator's 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. Those statements are the operator's disputed self-description; they are attributed wherever they appear and never adopted.
The internal stances quoted on this board are insider characterisations that the operator disputes: 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. Neither is a company statement, neither has been confirmed by any adjudication, and no override figure, staffing figure or tolerated-error figure has been published from either side. They are on this board because a stated acceptance criterion for employment-consequential automation is rare on any record, not because the record establishes it as fact.
One cited source overstates the case in the direction this scenario refuses to follow. The incident-registry consolidation of these events says drivers had no opportunity to contest or appeal the decision. The primary record documents a channel that is automated, slow and fee-gated rather than absent: ten days to appeal by email, replies reading as machine-generated, final answers on day 26 and day 11 in the two arcs followed to the end, and a two-hundred-dollar arbitration fee beyond that. The channel is drawn here as present and degraded, and no stronger claim about what it achieves is made anywhere, because no appeal-outcome or reversal rate exists.
Everything the city ordinance requires binds work performed in that city only, and the scenario draws it as a seam rather than a fleet-wide reform. A worker is covered at twenty-five percent of completed or cancelled-with-cause offers in the prior 180 days performed there, or by a single incident there. Nothing in the record says the operator changed the nationwide appeal channel in response. That office's enforcement reaches procedural compliance only until 1 June 2027, so whether a deactivation was substantively warranted is outside its authority even now. And the appellate decision leaving the ordinance standing is an affirmance of the denial of a preliminary injunction sought by two other network companies — a preliminary posture, not a final merits judgment, decided by a three-judge panel with one judge dissenting in part, and this operator was not a party.
The contractor classification underneath this whole regime is contested and the scenario states the contest rather than resolving it. A state workforce-development audit of more than a thousand drivers covering 2016 to 2018 found the vast majority employees for unemployment-insurance purposes with an assessment of about $205,000, upheld through that state's court of appeals in 2023 and left standing when its supreme court dismissed the appeal as improvidently granted on 26 March 2024. A state labour department sued on 20 October 2025 alleging misclassification since at least 2017; that suit is pending and entirely unadjudicated.
The two-hundred-dollar arbitration figure is date-stamped to the platform terms and the reporting in force in June 2021 and is not asserted as the current figure. The operator's widely reported 2021 removal of mandatory arbitration applied to consumer terms, not to these driver terms, and the present fee level and dispute routing were not independently verified. Separately, the 2020 appellate holding that these workers are transportation workers exempt from the federal arbitration statute concerned the WAGE-claims clause and was not a ruling about the deactivation-dispute channel, which continued to route through individual arbitration as reported in 2021.
The scale figures are labelled by what they count and nothing more. Approximately 4 million app downloads globally and 2.9 million in the United States, with more than 660,000 in a single quarter of 2021 and growth of about 21 percent year over year, are cumulative DOWNLOADS reported by the operator rather than a count of active drivers. The only hard official count anywhere in this record is the 140,128 individual drivers a federal regulator paid refunds to for one 2016-to-2019 window, at an average of $422 and a top payment above $28,000. A cross-platform after-expenses hourly median published by one institutional study is not a figure for this platform and appears nowhere here; this platform is in fact singled out in that study as the one paying a posted flat hourly block rate.
The surveillance surface is drawn at the level the stronger sources verify for this programme and no further. Timing from delivery-station arrival through delivery completion, motion classification of driving, walking or running, and mandatory selfie identity checks are asserted. Seatbelt, acceleration and screen-touch monitoring are one legal commentary's characterisation of the reporting record and are carried as that; inward-facing camera telematics belongs to a separate delivery-partner van programme with employed drivers and is attached nowhere here. Drivers on this programme use their own vehicles.
The people this network governs are drivers, and no outcome for any individual is computed anywhere on this diagram. No tier placement, no deactivation, no appeal result and no earnings figure for any person is derived from anything drawn here. The counts that appear — the refund distribution, the city settlement's worker count, the two resolutions at $1,245.70 each — are recorded external observations from a regulator's own records. Package recipients and the households on the route are outside the boundary entirely and are modelled nowhere.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

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.
empirical
- Investigative Soper, S. (2021, June 28). Fired by Bot at Amazon: 'It's You Against the Machine'. Bloomberg; paywalled and blocked from the verifying environment on 2026-08-28 and read in full through the syndication below https://www.bloomberg.com/news/features/2021-06-28/fired-by-bot-amazon-turns-to-machine-managers-and-workers-are-losing-out
- Investigative The Spokesman-Review (2021, June 28). Fired by bot at Amazon: 'It's you against the machine' (Bloomberg syndication, full text) https://www.spokesman.com/stories/2021/jun/28/fired-by-bot-at-amazon-its-you-against-the-machine/
- Trade press Dent, S. (2021, June 29). Amazon is reportedly using algorithms to fire Flex delivery drivers. Engadget https://www.engadget.com/amazon-algorithms-fire-flex-delivery-drivers-055959081.html
- Reference AI Incident Database, Responsible AI Collaborative (2021). Incident 111: Amazon Flex Drivers Allegedly Fired via Automated Employee Evaluations https://incidentdatabase.ai/cite/111/
- Government Federal Trade Commission (2021, February 2). Amazon To Pay $61.7 Million to Settle FTC Charges It Withheld Some Customer Tips from Amazon Flex Drivers; and (2021, November 2). FTC Returns Nearly $60 Million to Drivers Whose Tips Were Illegally Withheld by Amazon https://www.ftc.gov/news-events/news/press-releases/2021/02/amazon-pay-617-million-settle-ftc-charges-it-withheld-some-customer-tips-amazon-flex-drivers
The correction channel around the automated decision is slow, templated, and priced. As reported in June 2021, a deactivated Flex driver has ten days to appeal by email; first replies arrive the next day, read as machine-generated, and are typically generic rather than specific to the incident, signed with a support agent's first or full name; a promised six-day review is often exceeded, and in the two cases the investigation follows end to end the final answer lands on day 26 and on day 11, with no pay in the interim. The only escalation beyond email under the platform terms then in force carried a $200 arbitration filing fee — verbatim, drivers 'pay $200 to take their dispute to arbitration, but few do, seeing it as a waste of time and money' — against roughly $80 of net pay on a typical four-hour block, a fee legal commentary describes as an unrealistic prospect for workers barely making minimum wage. Drivers also report that standing takes months to recover from delays outside their control while standing feeds route and block allocation. The figure is date-stamped: Amazon's widely reported 2021 removal of mandatory arbitration applied to consumer terms rather than to the Flex driver terms, and the present fee level was not independently verified. One consolidating source overstates the position by saying drivers had no opportunity to contest or appeal; the primary record documents a channel that is automated, slow, and fee-gated rather than absent, and no appeal-outcome or reversal rate exists from any side.
empirical
- Investigative Soper, S. (2021, June 28). Fired by Bot at Amazon: 'It's You Against the Machine'. Bloomberg; paywalled and blocked from the verifying environment on 2026-08-28 and read in full through the syndication below https://www.bloomberg.com/news/features/2021-06-28/fired-by-bot-amazon-turns-to-machine-managers-and-workers-are-losing-out
- Investigative The Spokesman-Review (2021, June 28). Fired by bot at Amazon: 'It's you against the machine' (Bloomberg syndication, full text) https://www.spokesman.com/stories/2021/jun/28/fired-by-bot-at-amazon-its-you-against-the-machine/
- Trade press Dent, S. (2021, June 29). Amazon is reportedly using algorithms to fire Flex delivery drivers. Engadget https://www.engadget.com/amazon-algorithms-fire-flex-delivery-drivers-055959081.html
- Academic Bajgiran, T. (2022, January 5). Fired by an Algorithm: Amazon Flex and the Case for Human-in-Command Termination Review. OnLabor (Harvard Law School) https://onlabor.org/fired-by-an-algorithm-amazon-flex-and-the-case-for-human-in-command-termination-review/
- Reference AI Incident Database, Responsible AI Collaborative (2021). Incident 111: Amazon Flex Drivers Allegedly Fired via Automated Employee Evaluations https://incidentdatabase.ai/cite/111/
Since 1 January 2025 the deactivation channel itself is regulated in one jurisdiction. Seattle's App-Based Worker Deactivation Rights Ordinance (Seattle Municipal Code 8.40, Ordinance 126878; administrative rules SHRR Chapter 260 effective 24 June 2025) covers network companies with 250 or more app-based workers worldwide, Amazon Flex among them, and requires a published deactivation policy reasonably related to safe and efficient operations; fourteen days' notice before deactivation except for egregious misconduct or where law requires immediacy; a written statement of the reasons and the specific incidents together with all records relied on and considered; investigation of the alleged violation to a more-likely-than-not standard before deactivating; penalties applied consistently and in proportion to the violation with the circumstances of the work considered; and an internal challenge procedure invocable within ninety days, with a private right of action after the company's response or fourteen days after the challenge. A worker is covered at twenty-five percent of completed or cancelled-with-cause offers in the prior 180 days performed in the city, or by a single incident there, and the Office of Labor Standards enforces procedural compliance only until 1 June 2027, not whether a deactivation was substantively warranted. That office's published resolved-investigations record for October to December 2025 shows two informal resolutions under this ordinance 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, alongside equivalent resolutions the same quarter against three other network companies. In Uber Technologies, Inc. v. City of Seattle, Nos. 25-228 and 25-231 (9th Cir. 4 March 2026), a panel of 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 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, with Judge Bennett dissenting in part. That is a preliminary-injunction affirmance rather than a final merits judgment, and Amazon was not a party.
empirical
- Government City of Seattle, Office of Labor Standards. App-Based Worker Deactivation Rights Ordinance, Seattle Municipal Code 8.40 (Ordinance 126878, effective 1 January 2025; administrative rules SHRR Chapter 260 effective 24 June 2025), with the City's Notice of Rights https://www.seattle.gov/laborstandards/ordinances/app-based-worker-ordinances/deactivation-rights
- Government Seattle Office of Labor Standards (2025). Resolved Investigations, October to December 2025 https://seattle.gov/laborstandards/investigations/resolved-investigations/october-%E2%80%93-december-2025
- Government Uber Technologies, Inc. v. City of Seattle, Nos. 25-228 and 25-231 (9th Cir. 4 March 2026) (Clifton, J.; Bennett, J., dissenting in part) https://cdn.ca9.uscourts.gov/datastore/opinions/2026/03/04/25-231.pdf
- news KOMO News (2025, December). Seattle secures $3.7M settlement for Amazon Flex drivers over gig-worker protection rules https://komonews.com/news/local/seattle-amazon-flex-reach-nearly-38m-settlement-over-gig-worker-pay-sick-leave-rules-premium-pay-ordinance-app-based-worker-law
The independent-contractor classification underneath the deactivation regime is contested and has repeatedly gone against the operator in adjacent forums, while no forum has reached the rating rule. A Wisconsin Department of Workforce Development audit of more than 1,000 Flex drivers covering 2016 to 2018 found the vast majority employees for unemployment-insurance purposes with an assessment of about $205,000; that determination was upheld by the state Court of Appeals in 2023 and left standing when the Wisconsin Supreme Court dismissed Amazon's appeal as improvidently granted on 26 March 2024. The New Jersey Department of Labor and Workforce Development sued Amazon on 20 October 2025 in Essex County Superior Court alleging Flex misclassification since at least 2017 with millions of dollars in annual losses to state benefit funds; that suit is pending and entirely unadjudicated. In Rittmann v. Amazon.com, Inc., No. 19-35381 (9th Cir. 19 August 2020), the court held that Flex delivery workers are transportation workers engaged in interstate commerce and exempt from the Federal Arbitration Act under 9 U.S.C. § 1 even without crossing state lines, and that the arbitration provision with its choice-of-statute clause was unenforceable under federal or Washington law, affirming the denial of the motion to compel arbitration of the wage claims. That holding concerns the WAGE-claims clause; the deactivation-dispute channel continued to route through individual arbitration as reported in 2021, and the two facts are distinct.
empirical
- Government PBS Wisconsin / Associated Press (2024, March 26). Wisconsin Supreme Court lets ruling stand that declared Amazon Flex delivery drivers to be employees; and New Jersey Department of Labor and Workforce Development (2025, October 20). NJDOL Sues Amazon for Misclassifying Flex Delivery Drivers as Independent Contractors https://www.nj.gov/labor/lwdhome/press/2025/20251021_Amazon.shtml
- Government Rittmann v. Amazon.com, Inc., No. 19-35381 (9th Cir. 19 August 2020) (M. Smith, J.; Bress, J., dissenting) https://cdn.ca9.uscourts.gov/datastore/opinions/2020/08/19/19-35381.pdf
What the scorer measures, and what it has no field for, are both documented. Human Rights Watch's May 2025 cross-platform study — 95 workers interviewed across 13 states including Flex drivers, plus a 127-worker Texas survey — records that Flex times a worker from arrival at the warehouse through completion of the delivery and detects whether the worker is driving, walking, or running, while noting that Flex is the one studied platform paying a posted flat hourly block rate of $18 to $25 rather than an opaque per-job wage algorithm; the study's $5.12 hourly after-expenses median is a cross-platform Texas figure and is not a Flex figure. Mandatory selfie identity checks are part of the same surface, and failed checks are among the documented contributors to adverse action. A Harvard Law School labour-law commentary of January 2022 additionally characterises the reporting record as including monitoring of seatbelt use, acceleration, and screen touches while driving; that characterisation mixes Flex with Amazon's separate Delivery Service Partner van programme, whose inward-facing camera telematics does not apply to Flex drivers, who use their own vehicles. What none of these inputs can represent is the reason a stop went wrong — a locked building, a closed office, a station queue, weather, road conditions, or a vehicle failure — which the investigation records arriving in the measurement as the driver's own shortfall.
empirical
- Advocacy Human Rights Watch (2025, May 12). The Gig Trap: Algorithmic, Wage and Labor Exploitation in Platform Work in the US https://www.hrw.org/report/2025/05/12/gig-trap/algorithmic-wage-and-labor-exploitation-platform-work-us
- Academic Bajgiran, T. (2022, January 5). Fired by an Algorithm: Amazon Flex and the Case for Human-in-Command Termination Review. OnLabor (Harvard Law School) https://onlabor.org/fired-by-an-algorithm-amazon-flex-and-the-case-for-human-in-command-termination-review/
- Investigative Soper, S. (2021, June 28). Fired by Bot at Amazon: 'It's You Against the Machine'. Bloomberg; paywalled and blocked from the verifying environment on 2026-08-28 and read in full through the syndication below https://www.bloomberg.com/news/features/2021-06-28/fired-by-bot-amazon-turns-to-machine-managers-and-workers-are-losing-out
- Investigative The Spokesman-Review (2021, June 28). Fired by bot at Amazon: 'It's you against the machine' (Bloomberg syndication, full text) https://www.spokesman.com/stories/2021/jun/28/fired-by-bot-at-amazon-its-you-against-the-machine/

Where this connects

Institutional pressures in this domain

Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Hiring & employment screening AI domain page.

Levers available here and the patterns behind them

Mark AI-written records — Provenance labeling
Check copied records — Reconcile copied records
Assign a challenger — Structured dissent
Understand the system — Understand the system
Gate record entries — Human-in-the-loop write gating
Review on schedule — Oversight cadence & retrospectives
Review the riskiest first — Risk-tiered oversight
Store less data — Data minimization
Check with a second model — Cross-model verification
Upgrade model — Improve the model
Vet connections — Connection authorization