AI IS MEASURED BY WHAT IT CAN DO. WE MEASURE WHETHER YOU STILL HAVE A HUMAN OPTION.
An index of recourse standing

When AI decides about you, is there a human option?

The Human Option is an independent project that gives a person recourse when AI makes decisions that affect their lives: denies a loan, rejects a claim, refuses a job, won't let them rent, reads a scan, cuts a benefit, or flags a car to the police. We score real AI-based decisions from public records, then give people a letter they can send to fight back and a guide to their rights. We also campaign for three changes in AI law under the name Not Final. And we give companies that use AI a guide to improve their practices. The index is the measurement at the center of that work, not the whole of it.

INDEPENDENT · BUILT FROM PUBLIC RECORDS · OPEN METHOD · NO FUNDING FROM COMPANIES WE ASSESS

From the record · SafeRent, Louis v. SafeRent Solutions
"We do not accept appeals and cannot override the outcome of the Tenant Screening."
A property manager wrote this to Mary Louis after a screening score refused her rental application. The company that used the score could not change it, and neither could she. Read the story
What we found so far

We scored 20 decisions that AI makes about people: loans, claims, jobs, rentals, benefits, diagnoses, surveillance alerts.

In four of them, nobody could change the decision. Not the person it was about, and not the company that used the system. Three are closed cases with a full public record; one, Cigna, is a court case still running, scored on records reported by ProPublica.

In one place, Seattle, a city law fixed it: notice, records, a human review and an appeal, and more than 30 workers got their accounts back.

ALL 20, WITH THE REASONING →

The project's five components

01THE TERM
Recourse standingthe legal property we measure

Recourse standing is the position from which a person can actually seek and obtain reconsideration of an automated decision about them: to know it was made, to learn its reasons, to correct the record it rests on, to reach a person with authority to change it, and to decline the automated path without disproportionate loss.

The law in both Europe and the United States already names this person. The EU AI Act calls them the "affected person" and gives them a right to explanation; the EU's GDPR gives them a right to contest; American credit, housing and insurance law give them notice, reasons and dispute rights; Colorado adds correction and human review in 2027. What these laws do not do is measure whether these rights exist in practice in a specific bank, hospital, platform or agency. Recourse standing provides that measurement: it records the standing a person in fact holds at each step of a documented decision, as distinct from the right the law grants on paper. It is not a property of a model. It belongs to one human relationship inside one documented deployment. The same clinical AI can leave a radiologist with full authority but a patient with none. We score both. The Human Option is the plain-language name for the same thing, and the name of our index. Recourse standing is applied through a three-part test, symmetry, silence and practical effect, set out in our methodology.

In plain words

A bank, an insurer, an employer or a city uses software to decide something about you. The law says you should be told, given the reasons, allowed to fix mistakes and able to reach a person who can change the answer. We check, from documents anyone can read, whether you actually can. If you can, the system has a human option. If the company can and you cannot, we publish the difference.

The criterion is simple: at every step of the decision, does this person have a human option, and what does it cost to use it?
02WHAT IS DIFFERENT

Rights exist on paper. But whether they work in a specific bank, hospital or platform has not been measured.

Existing measures look at the developer, the chatbot user, the worker or the company. The law, in turn, gives the affected person rights to notice, explanation, correction and review. The Human Option checks, deployment by deployment and from public records, whether those rights are exercisable by the borrower, the patient, the claimant, the driver and the applicant, including people who never used the AI and signed nothing.

MeasureWhat it scoresWhose powerScores deployed decisions
Frontier lab safety indexesPolicies and safety practices of model developersThe developerNo
Assistant agency benchmarksHow a chatbot behaves toward its user in conversationThe model userNo
Task automation scalesHow much human involvement a work task needs or workers wantThe workerNo
Cognitive and behavioural studiesHow AI use changes judgment, confidence and dependenceThe user, in experimentsNo
Corporate accountability benchmarksCompany-level human rights and transparency commitmentsThe companyPartial
The Human OptionPractical power at each step of a documented AI-mediated decision, with separate profiles for operator and affected personThe affected personYes · public evidence

We do not claim to be the first to name the affected person, or to define contestability: that work belongs to the law and to the scholars we cite in the methodology. We claim a measurement that, to our knowledge, nobody has published, and a method anyone can retrace.

THIS WEEK
REGULATORS AND COURTS
03HOW IT WORKS

The Decision Path Audit follows the person through the decision, step by step.

We do not check whether a policy exists. We reconstruct the path a real decision takes and ask, at each of eight steps, what practical power does the affected person have here? Each step is scored 0 to 4 against public documents.

STEP01

Data collection

Which data about the person enters the system, and whether they can see it.

STEP02

AI assessment

Whether the person is told that AI participates and what role it plays.

STEP03

Institutional decision

Who has final authority, and whether a human with that authority can override the output.

STEP04

Notification

Whether the person learns that a decision was made about them at all.

STEP05

Explanation

Whether the reasons are specific enough to act on, not a generic notice.

STEP06

Correction

Whether wrong data or a wrong assumption can be fixed before it counts.

STEP07

Human review and appeal

How deep the recourse goes, from automated reconsideration to an independent body.

STEP08

Exit

Whether the person can refuse the AI-mediated path, and what refusing costs.

4 = REAL POWER, DOCUMENTEDSeven powers are scored across the path: Understand, Choose, Influence, Challenge, Control, Data, Exit. Capability retention is a separate research track.0 = SHOWN TO BE ABSENT
RECOURSE DEPTH

How deep the appeal goes.

Each path is placed on a five-point scale of what the appeal can actually change.

0No meaningful recourse
1Automated reconsideration
2Human review without decision authority
3Human review with authority to change the outcome
4Independent or external contestability
04THE HUMAN OPTION GAP

The same decision, scored from both sides.

A company can have excellent governance, audit logs and human oversight, and all of that power can belong to the institution. The Human Option Gap is the difference between the operator's power and the affected person's power around the same decision. Below: the format of a profile for an AI lending deployment. Illustrative structure, not a published score.

StepLender · operatorBorrower · affected person

Knows AI was usedUNDERSTAND

4
3

Sees the specific reasonsUNDERSTAND

4
ND

Adds context before the decisionINFLUENCE

4
1

Corrects the data usedDATA

4
3

Reaches a human with authorityCONTROL

4
1

Appeals and changes the outcomeCHALLENGE

4
2

Walks away without losing accessEXIT

4
1
54Human Option Gap · points held by the institution, not the person
88Operator score
34Affected person score · provisional
NDWeakest step for the borrower · reaching a person with authority · not shown in public records

ND means the public record shows nothing either way; it is scored separately from zero. Every profile also records how serious the decision is and how unequal the parties are, so that a loan refusal is never weighed like a film recommendation.

05THE FIRST EDITION

20 decisions people cannot avoid.

Credit, insurance, payments, hiring, gig work, housing, clinical care, health coverage, policing, immigration enforcement and public benefits. Four of the 20:

Work and housing · 004

Workday

AI screening → employer → applicant

Court treats the screening tool as the employer's agent; applicants rarely know AI rejected them; bias-testing data shielded as privileged.

Provisional gap 54
Work and housing · 005

Uber

Fraud and rating models → Uber → driver

European 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 zero
Body · 007

Aidoc

Triage AI → hospital → radiologist and patient

FDA label: notification only, radiologist retains authority. Patient is not told AI was involved.

Provisional gap 75
Surveillance · 010

Flock Safety

Plate recognition → police → driver

Human verification of hotlist hits required by policy; the driver gets no notice, explanation or appeal at any step.

Provisional gap 86
06THE PEOPLE BEHIND THE NUMBERS

The people behind the numbers.

HOUSING · SAFERENT
"We do not accept appeals and cannot override the outcome of the Tenant Screening."

Mary Louis had a housing voucher and a long rental history. A score said no, and the management company wrote back that nobody could change it. She found a place two hundred dollars a month more expensive, in a worse area. In 2024 a court approved a settlement under which the score is no longer shown for anyone using a voucher.

Louis v. SafeRent, D. Mass.
GIG WORK · UBER
"From one moment to the next, they no longer had any income."

Between 2018 and 2022, when Uber's software suspected fraud or saw a rating fall below a threshold, it switched drivers off. No warning, no person. More than 170 French drivers took their case to a human rights league, then to the regulator. Six years later Uber was fined 825 million euros. The words are the regulator's.

CNIL, August 2026
07WHY THESE DECISIONS

Why these decisions.

The index follows one person through the decisions a machine now makes about them: money, work, housing, health, dealings with the state, and the records that follow them from one to the next. The question at each of these points is the same, and so is the scale.

Four rules determined the list

The law already calls the decision consequential. Credit, insurance, employment, housing, health care, essential services, law enforcement and migration are the categories that Colorado's automated decision law, California's rules and the EU AI Act name explicitly. Where the law has drawn a line, a person has a foothold, and we can measure whether it holds.

The decision can be reconstructed from public records. Each of these areas produces documents an outsider can read: adverse action notices, securities filings, device labels, transparency portals, court records, regulator decisions. We do not score what cannot be checked.

Each area tests a different relationship between the machine and the person. A credit model judges the applicant directly. A screening tool decides whether the applicant gets past the door. A clinical tool advises a professional about a patient who may never know. A plate reader watches a driver who signed nothing. A background check remembers a person whether or not the record is right. A method that produces comparable scores across all of these can be trusted on any one of them.

Exit is expensive. People cannot opt out of credit, employment, housing, health care or public roads. A human option matters where leaving is not a real choice; a film recommendation does not qualify.

Five areas, one path

Money

Credit, insurance, payments. The machine decides whether you get funds, a payout or your own balance back. Governed by adverse action law, state insurance codes and user agreements.

Work and housing

Hiring, gig work, renting. The machine gates income and shelter. Governed by anti-discrimination law, fair housing law, consumer reporting law and, in Europe, the data protection rules that fined Uber.

Body

Clinical triage, coverage decisions, reproductive health. The machine advises a professional or an insurer about you. Governed by device regulation, medical records rights and Medicare appeal rules.

Surveillance

Plate readers, enforcement targeting, benefit fraud systems. The machine observes or sanctions people who signed nothing. Governed by constitutional due process, public records law and little else.

Records

Background checks, data brokers, court and eviction records, pretrial risk scores. The machine keeps a file on you and other machines read it. Governed by consumer reporting law and the new data broker deletion rights. This is the fifth area and the second edition: when the file is wrong, what can you do?

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.

Write to us →

or email info@humanoption.org