Fifteen live deployments and five closed cases.
Credit, insurance, payments, hiring, gig work, housing, clinical care, health coverage, policing, immigration enforcement and public benefits. These are the decisions regulators in Colorado, California and the European Union have already singled out as consequential. Fifteen live deployments across four areas, and five closed cases kept as benchmarks because their records are complete. Our scores are first readings from public documents. Every company receives its full file and its reply is published alongside. The index is evidence for the project's other work and the step that makes collective action possible, the term, the cases, the campaign and the tools, and it is built to a strict standard because evidence has to survive checking; how it is used is set out in the methodology.
Live deployments, ranked by the person's score.
How to read this table. The score is the affected person's average across the steps we could document, on a scale of 0 to 100. Steps with no public record are left out of the average and counted in the ND column. "Standing blocked" means the person scored 0 at one of three steps where a 0 ends everything: being told, being given reasons, or reaching a person with authority. These are first readings: an AI-assisted reading verified by a human coder, with a second AI pass and a human review pass completed and further independent human passes, at least two per edition, added as the team grows; the log of disagreements and the reconciled scores will be published in October. Palantir's work for ICE is described but not ranked, because we do not score enforcement targeting numerically.
| System | Person's score | Weakest step (score) | ND | Flag |
|---|---|---|---|---|
| Seattle deactivations (Uber, DoorDash, Instacart under city law) | 83 | exit (1) | 0 | |
| Natural Cycles | 82 | reaching a person before the decision (2) | 1 | |
| Upstart | 70 | data collection (2) | 3 | |
| Lemonade | 68 | data collection (2) | 1 | |
| UnitedHealth nH Predict | 47 | data collection (0) | 0 | |
| Cigna PxDx | 44 | reaching a person before the decision (0) | 0 | standing blocked |
| PayPal | 39 | reaching a person before the decision (0) | 1 | standing blocked |
| Aidoc (patient) | 25 | notification (0) | 4 | standing blocked |
| SafeRent | 25 | reaching a person before the decision (0) | 0 | standing blocked |
| Workday | 21 | reaching a person before the decision (0) | 1 | standing blocked |
| CAF, France | 19 | assessment (0) | 0 | |
| Flock Safety | 14 | reaching a person before the decision (0) | 1 | standing blocked |
| Evolv in schools | 11 | reaching a person before the decision (0) | 1 | standing blocked |
| Clearview AI, Vermont | 3 | data collection (0) | 0 |
The full step-by-step scores for every role are in the data file (CSV, CC BY 4.0). Reasoning for each number is on the methodology page.
Upstart
Live deploymentAI credit model → bank → borrowerAdverse action law forces specific reasons; SHAP-generated denial reasons published; no documented human appeal.
Provisional gap 30 · ND on three stepsLemonade
Live deploymentAI Jim → insurer → claimant96% of first notices of loss and 55% of claims automated end to end per 10-K; escalation to humans is system-initiated.
Provisional gap 32PayPal
Live deploymentRisk models → PayPal → account holderFunds frozen by algorithm before any contact; appeals through a resolution center; holds up to 180 days.
Provisional gap 36Workday
Live deploymentAI screening → employer → applicantCourt treats the screening tool as the employer's agent; applicants rarely know AI rejected them; bias-testing data shielded as privileged.
Provisional gap 54SafeRent
Live deploymentScreening score → landlord → applicantThe landlord's own words: "We do not accept appeals and cannot override." Authority on paper, none in practice.
Provisional gap 25 · both sides lowAidoc
Live deploymentTriage AI → hospital → radiologist and patientFDA label: notification only, radiologist retains authority. Patient is not told AI was involved.
Provisional gap 75Natural Cycles
Live deploymentFertility algorithm → direct userSix FDA clearances; GDPR rights for all users; consent-based data flows. The person chose the AI.
Single role · score 3.3 of 4UnitedHealth nH Predict
Live deploymentCare prediction → insurer → memberFormal appeal ladder reaches an independent reviewer; the finding is friction and time, not absence.
Provisional gap 28Flock Safety
Live deploymentPlate recognition → police → driverHuman verification of hotlist hits required by policy; the driver gets no notice, explanation or appeal at any step.
Provisional gap 86Palantir for ICE
Live deploymentTargeting system → ICE → person targetedSole-source contracts to prioritize removals and score addresses; first contact is enforcement. Analytical review, not a numerical score.
Analytical profile, not rankedMoney
AI assesses the person: credit, insurance, claims, fraud, payments. Next: Zest AI, Affirm, Klarna, Root, Stripe.
Work and housing
AI gates access to income and shelter: hiring, gig deactivation, tenant screening, rent setting. Next: HireVue, Amazon Flex, DoorDash, RealPage.
Body
AI decides about health and reproduction: clinical triage, coverage, femtech, wearables. Next: Viz.ai, Epic, Cigna, Flo, Oura.
Surveillance
AI observes, targets or sanctions people who never signed anything. Next: Clearview AI, Axon, CBP targeting, benefits systems. Military and targeting systems reviewed analytically only.
Reproductive health apps
These systems sit inside the Body cluster, and they get their own edition for one reason: the person chose the app, and the record it keeps about cycles, pregnancy and fertility can be used by someone else in a decision the person never agreed to. Since Dobbs, prosecutors and litigants have sought digital records to reconstruct a pregnancy, and the record most people worried about was the one in their period tracker. The recourse standing of a user here runs beyond the app to whoever can obtain the data later, which makes this the bridge between Body, State and Records. Natural Cycles anchors it in the first edition; Flo, Clue, Oura and Apple Cycle Tracking follow.
The name on the card is the system. The words under the name say who decided, who deployed it, and who the system decided about. The paragraph underneath is what the public record shows. The number is the gap between what the company's staff can do and what the affected person can do, on a scale of 100. Every profile also names the affected person's weakest step, because a strong appeal is worth little to someone who was never even told a decision was made. A high gap means the person has little say. A low gap with both sides scored low means nobody, not even the company, had a human option. Every number links to the reasoning on the methodology page.
Eight decisions that, to our knowledge, no index has scored from the person's side.
Current as of September 2026. Five are European, because the same kind of decision under the EU's data protection law reads differently from the same decision without it. Two are closed cases, kept as benchmarks. One is the only system in the index where the law gives the person more options than it gives the platform.
Evolv in schools
Live deploymentAI weapons scanner → school → studentThe FTC found the company misrepresented what its scanners detect; one missed a seven-inch knife used in a 2022 stabbing. Recourse ran to the buyer: schools could cancel contracts. The student searched in front of classmates got nothing.
Gap 64 · student ND on one stepClearview AI, Vermont
Live deploymentFace database → police users → residentThree state lawsuits ended in December 2025 with a dismissal for lack of jurisdiction: the company has no presence in Vermont, so its residents' faces are a "random" connection. The person cannot even obtain a forum.
Gap 72CAF, France
Live deploymentRisk score → family benefits agency → recipientThirteen million households scored every month. Low income, unemployment and disability benefits raise the score. The 2026 model's code is public, its training data is not; 25 organizations are before the Conseil d'État.
Gap 56Seattle deactivations
Live deploymentPlatform decision under city law → workerFourteen days' notice, the records behind the decision, a human review and a challenge procedure, by ordinance. The city helped more than 30 workers get reactivated in its first fourteen months. A federal appeals court upheld the law in March 2026.
Gap 17 · highest person score in the indexCigna PxDx
Live deploymentProcedure-to-diagnosis algorithm → insurer → memberThree hundred thousand denials in two months of 2022 and an average of 1.2 seconds of physician review each, according to internal records reported by ProPublica; the figures are allegations in a pending class action, not findings of a court. Part of the action has been allowed to proceed.
Both sides lowWhere the person's score is close to the operator's and both are low, the record shows that nobody at the institution had a human option either. Deliveroo, Cigna and Michigan follow this pattern.
Five closed cases, scored as benchmarks.
These systems no longer run in the form that was scored, or the scored period has ended. They are in the index because the public record about them is unusually complete: audits, court findings and regulators' decisions describe what a person could and could not do at each step. They are reported separately from live deployments and are not averaged with them.
Uber
Historic period 2018 to 2022, scored from the regulator's findings; current practice under reviewFraud and rating models → Uber → driverEuropean regulators fined Uber 825 million euros for deactivating drivers with no person involved, 2018 to 2022. Nobody had a human option: operator and driver both score near zero for that period.
Both sides low · gap near zeroMichigan MiDAS
Closed case, 2013 to 2015Fraud determination → state agency → claimantFully automated fraud findings 2013 to 2015 with no person checking; the state later admitted it and paid 20 million dollars to settle. The appeal existed on paper only.
Both sides zero · historic benchmarkDUO, Netherlands
Closed case, decisions reversed 2024Risk profile → student finance agency → studentInvestigators chose whom to visit from a profile trained on their own past choices; nearly 25,000 home visits. The government apologized in 2024, the data authority called the algorithm discriminatory, and decisions were reversed with restitution.
Gap 34 · both sides lowRotterdam welfare model
Closed case, model paused 2021Risk score, 315 inputs → city → recipientThe only system whose model file, training data and code journalists obtained. Being a woman, a parent, young or not fluent in Dutch raised the score. The city paused it in 2021; hundreds had lost benefits.
Gap 71Deliveroo Italy
Closed case, 2020 ranking modelReputation ranking → platform → riderA Bologna court found the algorithm blind to the reason for an absence: a strike, a sick child and a no-show scored the same. Nobody at the company could see the reason either. First European ruling of its kind.
Both sides near zeroWhat we read, and how much weight it carries.
The chain is fixed: question, claim, evidence, source, assessment, score. Anyone can retrace it. Sources are ranked in five tiers and the tier limits what a source can prove.
Terms of service, privacy policies, application and customer agreements, SEC and regulatory filings, court records, public contracts, adverse-action procedures, regulator decisions, deployment policies and transparency portals
Supports any scoreInterfaces, help centres, appeal workflows, manuals, screenshots, documented processes
Supports any scoreBlogs, press releases, marketing, executive statements
Context only · never alone a high scoreAcademic studies, regulatory investigations and enforcement, consumer complaint databases, NGO and journalistic investigations, independent testing
Can lower a score or confirm absenceResearcher interpretation where the documents are silent
Never supports a positive scoreRight of reply
Every company receives its full dossier before publication with a ten-business-day window. The assessment does not wait for the reply and needs nothing from the company. The reply is published with the profile and can raise the score if it supplies Tier A or B evidence.
Two coders per case
Every pilot case is scored independently by two researchers. Disagreements are recorded and published with the methodology paper.
Versioned and firewalled
Profiles carry a version history. Corrections are logged publicly. No paid service can change a published score; the funding firewall is published.
Other indexes in this field.
The Observatory catalogs existing ratings, benchmarks and scales with their object, unit of analysis, method and how they relate to the Human Option framework. Each entry links to the original. We describe and map; we do not republish anyone's scores.
| Rating | Object | Method | Relation to the Human Option |
|---|---|---|---|
| FLI AI Safety Index | Nine frontier AI developers, 37 indicators | Expert panel grading of published policies | Developer layer · complementary |
| HumanAgencyBench | LLM assistants, six agency-support behaviours | Automated behavioural benchmark | Model behavior toward user |
| Stanford Human Agency Scale | Work tasks, H1 to H5 involvement | Worker and expert surveys | Worker layer |
| MIT AI Agent Index | Deployed agentic systems | Documentation audit | System documentation, no affected person |
| MIT Human Agency Platform · Agency Index | Technologies across domains | Design-principles framework | Principles, not scores |
| Ranking Digital Rights | Tech and telecom companies | Public commitments scored | Company commitments |
| World Benchmarking Alliance | 2,000 companies, human rights and social | Public evidence benchmark | Format precedent |
| Mozilla Privacy Not Included | Consumer products and apps | Policy review and researcher testing | Privacy layer, overlaps Data power |
| The Human Option | One human role in one documented AI-mediated decision | Decision Path Audit on public evidence, two coders, right of reply | Affected person layer |
Entries are added on request and on discovery. Submit a rating we missed and it appears with attribution.
Support the project.
We are funding the first two editions of the index, the methodology paper, the tools and the Not Final campaign. Foundations, newsrooms, researchers and companies willing to open their decision paths are welcome.
or email info@humanoption.org