The Selection Machine
Make twelve ordinary decisions. None of them is a test, as far as you can tell. Choose what you would really do.
Somewhere behind the screen, each one becomes a signal, each signal gets a weight, and the total decides whether you are selected, tolerated or excluded.
The criteria you were never shown
| What you did | What it read | Points |
|---|
Now change the rules
Merit is nine numbers somebody chose. Move them.
If you scored badly for independence, creativity, privacy or solidarity, you found no flaw in yourself. You found the machine's definition of merit. Somebody chose it, and it was never neutral.
The invisible employer
Each step moved the decision a little further from a person you could argue with. No step replaced the one before it. They stacked.
1. The human supervisor
Watches your work.
1911. Frederick Taylor wrote it down: "In the past the man has been first; in the future the system must be first."
1914. Ford's five-dollar day came with conditions. A good employee had to be sober, thrifty and keep a positive attitude. Investigators made unannounced home visits, checked bank deposits and school attendance, and rated households good, fair or poor. The preferred human is more than a century old.
2. The performance dashboard
Measures your output.
2019. A letter from an Amazon lawyer, reported by The Verge, said its warehouse system generated productivity warnings and terminations automatically. Amazon denied it: "It is absolutely not true that employees are terminated through an automatic system." It said the letter was wrong.
2022. Ontario began requiring larger employers to write down how they monitor staff electronically. The province's own guide says the rules "do not establish a right for employees not to be electronically monitored".
3. The platform algorithm
Controls your access to work.
2020. A Bologna court found Deliveroo's rider ranking was indirect discrimination. It did not tell a strike or an illness apart from any other reason for missing a shift. Deliveroo said it had stopped using that booking model.
2023. In a case brought by drivers, the Amsterdam Court of Appeal held that assigning rides, pricing and fraud deactivations could be automated decisions. In most of the four deactivations before it, Uber could not show human involvement was more than a symbolic act.
4. The behavioural profile
Predicts how you'll behave.
2018. Reuters reported that an experimental Amazon recruiting tool, trained on ten years of mostly men's résumés, had "penalized resumes that included the word 'women's'". It was scrapped. Reuters' sources said recruiters saw its rankings and never relied solely on them.
2016. Unilever began putting student and intern applicants through twelve short games and a recorded video interview before a recruiter met them.
5. The converged identity
Could connect your work, play, purchases and social activity.
2025. Equifax's US payroll database, The Work Number, held 813 million current and past employment records from more than four million organizations. It sells verification to lenders, employers, screeners and government agencies. These are records, not people.
The word in this step is could. Pieces of it are documented. One file joining all of your work, play, shopping and social life is not.
6. The invisible employer
Makes selection decisions without ever meeting you.
2026. Researchers studied one screening vendor across 156 employers and about four million applications. Of the people who applied to ten positions, 4% got a negative recommendation from all ten, more than independent decisions would predict. One model can close many doors at once.
Pending. In Mobley v. Workday, a US court let applicants over 40 proceed as a provisional collective in 2025. Workday disputes the claims and has moved to undo the collective. No court has found that it discriminated.
Your boss is a formula
This is a fictional delivery platform. Move your acceptance rate, your rating, your hours and the target. Watch your income and your access to work change. No manager is involved at any point.
The real levers are less tidy. DoorDash publishes a 4.2 minimum rating in Canada and says acceptance has no minimum, yet acceptance still affects perks. Uber sets its minimum rating city by city.
Law professor Veena Dubal calls the pay version algorithmic wage discrimination: different hourly pay for broadly similar work, set by "ever-changing formulas using granular data on location, individual behavior" and more.
Show the formula
score = 0.35 × acceptance
+ 0.25 × (rating − 4)
+ 0.15 if you follow its schedule
+ 0.15 if you reach its hours target
+ 0.10 × (1 − cancellations ÷ 15)
rating under 4.4: deactivated
rating under 4.6 or cancellations over 10%: under review
score 0.85+: priority 0.65+: standard
score 0.45+: low below: restricted
costs: $2.40 a jobThe game that interviews you
Emberfall Online is a fictional game. Play its short quest the way you would play anything, for fun. Then read what else the quest could have been.
Hiring games are real. They are tests that applicants knowingly take, like Unilever's twelve, or Wasabi Waiter, a restaurant game that NPR reported employers were trying in 2013.
A game secretly passing your play to an employer is not. We found no documented case. That part is hypothetical, and so is the profile below.
The Convergence Map
A solid line is a documented flow of data, and it carries its source. Grey dots mark a documented flow that has ended. A dashed red line is hypothetical: possible, but not shown to be happening. The difference matters, so the map never blurs it.
- Payroll records to lendersEquifax's Work Number (US) sells employment and income verification to lenders. Documented: source
- Payroll records to screenersThe same database sells to employers and background screeners. Canada has a separate Equifax exchange. Documented: source
- Screening to gig workDoorDash Canada runs courier background checks through Checkr, a consumer reporting agency. Documented: source
- Screening to recruitmentBackground reports are sold for hiring. US law requires written permission and a copy before they are used against you. Documented: source
- Identity resolution to advertisingLiveRamp merges emails, addresses and phone numbers into one person and links their devices, for advertising. We found no evidence of it being used for hiring. Documented: source
- Broker profiles to recruiters, 2008 to 2010Spokeo marketed profiles built from online and offline data to recruiters. It agreed to an $800,000 US penalty in 2012. Documented: source
- Test games to recruitmentPurpose-built games, which applicants know they are taking, sort candidates before a recruiter sees them. Documented: source
- A job network to advertisingLinkedIn expanded data sharing with Microsoft affiliates in November 2025, for advertising and AI. Documented: source
- Your play to recruitmentAn employer screening you on ordinary gameplay. No documented case found. Microsoft owning both Xbox and LinkedIn is common ownership, and is not evidence of a flow. Hypothetical.
- Your play to a merged identityPlay behaviour joined to your name, address and devices, then resold as a trait. Hypothetical.
- Your gig score to your next jobA rating or deactivation following you into an unrelated application. Hypothetical.
- Your gig behaviour to your creditAcceptance rate and hours read as a sign of whether you will repay. Hypothetical.
- Your purchases to screeningAn advertising segment, such as new parent or night owl, read as a risk when you apply. Hypothetical.
The missing millions
Start with 100 million participants. Raise the bar, or switch on a filter. Each figure is 100,000 people. Watch them leave economic opportunity without one rejection letter.
One filter is documented. In a 2021 Harvard Business School and Accenture survey of employers in the US, UK and Germany, almost half weeded out résumés with a work gap. More than six months, and the résumé could be screened out automatically, whatever the reason.
In the same survey, 88% of employers agreed that qualified high-skills candidates get vetted out for not matching the exact criteria. That is employers describing their own systems.
Nobody publishes a count of the people filtered out unseen. We found none for Canada. The crowd below is a simulation, and that absence is part of the story.
Break the algorithm
A selection system is a set of rules, and rules can be rewritten. Switch on the ones you would demand. Your own run from the first chapter is judged again under them.
Some of these rules already exist, in pieces. As of 22 November 2026:
- OntarioSince January 2026, employers with 25 or more staff must say in a public job posting if AI screens applicants. Disclosure only: no appeal. Since July 2025, gig workers must be told how work is assigned and get written reasons if they are removed.
- QuébecSince 2023, if a decision is made exclusively by automated processing, you must be told, can ask for the principal reasons, and can put your case to a person able to review it.
- CanadaThe federal AI bill died in January 2025. A new privacy bill, C-36, is only a bill.
- European UnionThe Platform Work Directive bans predicting union activity and requires a human to make termination decisions. Countries have until 2 December 2026 to write it into law. The AI Act's hiring rules were postponed to 2 December 2027.
- New York CityHiring tools need bias audits. When the state comptroller checked 32 companies the city had reviewed, the city had found one problem and the auditors found at least 17 possible ones.
- Instead of rankingNew Zealand's Health Research Council funds eligible Explorer grant proposals by random draw. That is research funding, not hiring, and someone still sets the bar.
The public redesign
If you choose, add your rules to everyone's. This is the whole message that will be sent, and there is nothing else in it:
The server adds it to eight totals: one per rule, and one for the number of people. To stop floods it also counts submissions per minute for the whole site. It keeps no copy of your message, no address, no cookie and no log line for the request. Like every page here, the request passes through Cloudflare, which carries our traffic. Your browser keeps a note that you have voted, so you aren't counted twice.
The preferred human
The preferred human is always available, does as asked, never surprises, hides nothing, needs no care and organizes with nobody. Nobody is that person. Everybody is measured against them.
Surveillance is how the system sees you. Selection is what it does with what it sees. The power sits with whoever writes the weights.
You were not measured against the job. You were measured against the system's preferred human.
Someone you know was never chosen
Keep going
The Second Use: who gets your movement after you've made it.
More from Feline Union, all free.
9 cards, one tap each
Pick a card. Post it, copy it or download it. Every card links back here.
The record
- Taylor FW. The Principles of Scientific Management, introduction, 1911
- The Henry Ford. The Sociological Department
- Lecher C. How Amazon automatically tracks and fires warehouse workers for productivity. The Verge, 25 April 2019
- NPR via KUOW. Amazon denies report that it fires workers by robot, 26 April 2019
- Ontario. Written policy on electronic monitoring of employees
- TechCrunch. Italian court rules against discriminatory Deliveroo rider-ranking algorithm, January 2021
- TechCrunch. Uber and Ola drivers win on appeal in the Netherlands, April 2023
- Dastin J, Reuters, via CNBC. Amazon scraps secret AI recruiting tool that showed bias against women, October 2018
- ERE. Unilever wants to shorten hiring from four months to two weeks, September 2016
- Equifax. 2025 annual report, Form 10-K
- Bommasani R et al. Algorithmic Monocultures in Hiring. FAccT, 2026
- Proskauer. Conditional certification in Mobley v. Workday, May 2025
- Warden AI. Mobley v. Workday litigation tracker
- DoorDash Canada. Dasher ratings explained
- DoorDash Canada. Dasher rewards
- Uber. Driver deactivation policy
- Dubal V. On Algorithmic Wage Discrimination. Columbia Law Review, 2023
- NPR. Playing the game to get the job, December 2013
- DoorDash Canada. Background check FAQ
- United States Code, title 15, section 1681b
- LiveRamp. RampID methodology
- PCWorld. Spokeo data broker to pay $800,000 in FTC privacy case, 12 June 2012
- Open Terms Archive. LinkedIn shares data with Microsoft and its subsidiaries
- The Work Number. General information for verifiers
- Fuller J et al. Hidden Workers: Untapped Talent. Harvard Business School and Accenture, 2021
- Ontario. Employment Standards Act, 2000, section 8.4
- Ontario Regulation 476/24
- Ontario. Rights and protections for digital platform workers
- Québec. Act respecting the protection of personal information in the private sector, section 12.1
- Blakes. Digital policy after the prorogation of Parliament, January 2025
- Parliament of Canada. Bill C-36
- European Union. Platform Work Directive 2024/2831
- European Union. Regulation 2026/1744, postponing the AI Act high-risk rules
- New York State Comptroller. Enforcement of Local Law 144, December 2025
- Health Research Council of New Zealand. 2026 Explorer Grant guidelines
Read the narration transcript
The Selection Machine
Hey, come in. No rush. This is The Preferred Human. You were never rejected. You were never selected.
It's free, there's no account, and the experiments run in your browser. The line we're on lights up as we go, and the moment you touch an experiment, we pause and wait. Press play when you want us back.
And if you know someone who applied to two hundred jobs and heard nothing, the share button is right at the top. Shameless, I know.
Your promotion of this website has been noted.
Ignore it. Let's begin.
Good evening. Make twelve ordinary decisions. None of them is a test.
That's exactly what a test would say.
You're playing a game. Now you're working on a gig platform. Now you are just living. Now you apply for a job. Choose what you would really do.
Behind the screen every choice becomes a signal, every signal gets a weight, and the total puts you in one of three piles. Selected. Tolerated. Excluded.
And if I'm excluded, I get a letter?
No further action is required. You will not be contacted.
So run it. We'll wait. Then press reveal the criteria. You'll see independence filed as schedule deviation. Privacy as profile opacity. Solidarity as collective action risk.
I scored badly for helping my coworkers.
You found no flaw in yourself. You found the machine's definition of merit. Somebody chose it, and it was never neutral. The sliders are right there. Choose differently.
The invisible employer
It didn't start with software. Each step moved the decision a little further from a person you could argue with.
Nineteen eleven. Frederick Taylor wrote, in the past the man has been first. In the future the system must be first.
Nineteen fourteen. Ford's five dollar day had conditions. A good employee was sober, thrifty, and kept a positive attitude. Investigators visited homes unannounced, checked bank deposits, and graded households good, fair or poor.
So the preferred human is over a hundred years old.
Then the dashboard. I measure your output.
In 2019 a letter from an Amazon lawyer, reported by The Verge, said a warehouse system generated warnings and terminations automatically. Amazon denied it, and said the letter was wrong.
Then the platform. I control your access to work.
In 2020 a court in Bologna found Deliveroo's rider ranking was indirect discrimination. It couldn't tell a strike or an illness from any other missed shift. Deliveroo said it had stopped using that booking model.
In 2023 an Amsterdam appeal court looked at four driver deactivations. In most of them, Uber could not show the human involvement was more than a symbolic act.
Then the profile. I predict how you will behave.
In 2018 Reuters reported an experimental Amazon recruiting tool that had learned from mostly men's resumes, and marked down the word women's. It was scrapped. The sources said recruiters never relied on it alone.
Then the converged identity. I could connect your work, your play, your purchases, your friends.
Could. Say that word carefully. Pieces are documented. One file holding all of it is not.
And then I am your employer. We have never met.
In 2026, researchers followed one screening vendor across one hundred and fifty six employers. Of people who applied to ten positions, four percent got a negative recommendation from all ten. More than chance would predict.
Ten companies, and every single answer was the same. No.
Your boss is a formula
Welcome, driver. Your status is priority.
Great. What if I turn down the jobs that don't cover gas?
Your status is standard. Reviewed by: nobody.
Drag the sliders. Acceptance, rating, hours, the target. Watch income and access move without a manager ever appearing. It's a fictional platform, and you can open its formula, which is more than most drivers get.
Real levers are messier. DoorDash in Canada publishes a minimum rating of four point two, and no minimum acceptance rate, though acceptance still affects perks. Uber sets its minimum city by city.
The law professor Veena Dubal named the pay version algorithmic wage discrimination. Different hourly pay for broadly similar work, set by formulas that keep changing.
The game that interviews you
Okay, I'm off the clock. I'm playing a game.
Quest complete. Six hundred experience points. Also, risk tolerance, obedience, after hours availability, and participation in a work stoppage.
It was a guild boycott!
Recommendation. Do not advance.
Hiring games are real. In 2016 Unilever started using twelve of them on students and interns. Those are tests, and applicants know they're taking them.
A game secretly handing your play to an employer, we found no documented case. That part is hypothetical, and we label it. The point is how little would have to change.
The Convergence Map
Which is why the map has two kinds of line. Solid is documented, with a source. Dashed red is hypothetical.
Solid. A payroll database at Equifax, in the United States, sells verification to lenders, employers and screeners. Solid. DoorDash in Canada checks couriers through a screening company called Checkr.
Solid. A data broker, Spokeo, marketed profiles to recruiters, and agreed to an eight hundred thousand dollar penalty in 2012.
And my games going to recruiters?
Dashed. One company owning both a games business and a job network is common ownership. It isn't proof that data flows between them.
Tap any line. If someone shows you this map with every line solid, don't trust them.
The missing millions
One hundred million participants. Each figure is one hundred thousand people. Raise the bar.
They're turning red.
They have not been rejected. Rejection letters sent: zero.
One filter is documented. In a 2021 survey by Harvard Business School and Accenture, almost half of employers weeded out resumes with a work gap. More than six months could mean out, automatically, whatever the reason.
And eighty eight percent of those employers agreed that qualified high skills candidates get vetted out.
So how many people is it really?
Nobody publishes that count. The crowd is a simulation. The missing number is part of the story.
Break the algorithm
So break it. Seven switches. Publish the criteria. A person hears appeals. Work decisions use work data. Organizing can't count against you. Care is protected. A lottery above a basic bar. And workers vote on the rules.
Some of it exists, in pieces. Since January 2026, larger Ontario employers must say in a job posting if artificial intelligence screens applicants. That's disclosure, with no appeal.
Quebec goes further. When a decision is fully automated, you can ask for the main reasons and put your case to a person able to review it.
Europe's platform work directive bans predicting union activity, and is due to become national law by December second, 2026. The hiring rules in its A I Act were postponed to December 2027.
And if you want, add your rules to everyone's. It's the one part of your session that can be sent, and only if you press the button.
What does it send?
Seven switches, printed on the screen before you press. None of your choices, none of your scores, none of your timings. The server keeps totals.
Otherwise we'd have built the machine we're complaining about.
The preferred human
The preferred human is always available. Does as asked. Never surprises me. Hides nothing. Needs no care. Organizes with nobody.
Nobody is that person.
And everybody is measured against them.
Surveillance is how the system sees you. Selection is what it does with what it sees. The power sits with whoever writes the weights.
You were not measured against the job.
I was measured against the system's preferred human.
One more thing. And yes, it's a plug.
Send the machine to someone who thinks they failed. Let them see the weights.
There are cards under that, one tap each.
And The Second Use is linked right there. It follows what happens to your movement after you make it. There's plenty more from Feline Union, all free.
Thank you for listening. No further action is required.
Every source for this is listed at the end of the page, and every one of them is linked.
This was The Preferred Human. A Feline Union publication.
The drawings are illustrations. The platform, the game, the engine's weights and the simulator's filter rates are fictional or illustrative. Every real-world claim rests on the sources above.