Ministry of AI · Dispatch from 2047
The End of the Job Interview: My Last Résumé
Editorial note. The Ministry of AI is a work of disciplined foresight: it describes the year 2047 in the present tense, and treats our own era as history. The institutions are imagined. The economics, the evidence and the historical parallels are real and sourced.
The last résumé I wrote took me eleven hours, and the system that read it had been trained on documents I had written.
That is not a complaint about the machine. The machine did what it was built to do, and I know exactly how it was built, because through the 2020s I was one of the people building things like it. I had spent a decade producing exactly the kind of artefact — plans, analyses, forecasts, the confident memo before the quarter — that eventually became training material. My last application was, in a small and precise way, a document arguing that I should be paid for work a compression of my own past output could now do faster. It was rejected in under a minute, which was the only honest part of the exchange.
My youngest was six then. He is twelve now, and last month he asked me what an interview was for, in the tone you would use about a customs form for a country that no longer exists. I gave him the modern answer: it is a conversation about whether you want to do a piece of work with some people. He said that sounded fine. I did not tell him the other thing an interview used to be, which is that it decided whether you ate.
What the interview was actually for
The official story was matching. An employer had a task, a person had a capability, and the interview existed to pair them accurately. That story described maybe a third of what was happening in the room.
The rest was rationing. Income was distributed almost exclusively through employment, employment was scarce relative to the people needing it, and so a queue formed. The interview was the gate on that queue. It did not merely assess whether you could do the work; it decided whether you would be permitted to earn at all this year. The Ministry now calls this instrument the permission interview — a screen whose real function was allocating access to survival, wearing the costume of a skills assessment.
Once you name it that way, the strangest features of pre-transition hiring stop being strange. Why would a firm run five rounds for a job it had already decided about? Why would applicants perform enthusiasm they did not feel, for work they did not admire, to people who would forget them by dinner? Because the stake was never the task. The stake was the rent.
The admission cost
Every market has entry costs, and most of them are recorded somewhere. This one was not. So the Ministry gave it a name and a column: the admission cost — the total unpaid hours a person spends buying permission to earn.
It includes the applications, the tailored versions, the portfolio piece produced for free, the unpaid technical exercise, the travel, the waiting, the third conversation with a person who has no authority. None of it appeared in any national account. It was work, performed to a specification, for an organisation, at the organisation’s convenience, and it was uncompensated by definition. A labour market that measured everything measured this at zero.
The distribution is what makes it a justice question rather than an efficiency one. The admission cost falls hardest on whoever is screened out most often, which means the people with the least income paid the largest bill. My mother, at fifty-six, after nineteen years reading insurance claim files, wrote forty-one applications in her first year of unemployment. She received two conversations. She had, by the Ministry’s later reckoning, worked something close to a month unpaid for the privilege of being told no by systems partly trained on her own closed cases.
The evidence was already in, and it was damning
None of this required 2047 hindsight. The pre-Ministry research literature had established, decades before the transition, that the screen was noisy in ways that had nothing to do with skill.
In 2003, two economists sent fictitious résumés to help-wanted advertisements in Boston and Chicago, varying only the name at the top. Their results reported that white-sounding names received 50 percent more callbacks for interviews than African American-sounding names. Worse, the returns to merit were unequal: a higher-quality résumé produced about 30 percent more callbacks for white names and a far smaller increase for Black ones. Employers who advertised themselves as equal opportunity employers discriminated as much as everyone else.
Eighteen years later the finding was reproduced at industrial scale. A 2021 correspondence experiment sent more than 83,000 applications with randomised characteristics to jobs posted by 108 of the largest US employers. Distinctively Black names reduced the probability of employer contact by 2.1 percentage points on average — and the average was the least interesting number. Discrimination was highly concentrated: firms in the top quintile were responsible for nearly half of all lost contacts, and 23 individual companies could be identified as discriminating with statistical confidence.
Then there was time itself. A 2012 experiment sending résumés to real job postings in 100 US cities found that the likelihood of a callback fell significantly with the length of a person’s unemployment spell, most of the decline occurring within the first eight months. The applicant did not change. The screen simply read the gap as a signal. This is the mechanism by which a labour market converts one piece of bad luck into a permanent condition, and it operated silently, in aggregate, on millions of people who were told the problem was their attitude.
Even the human judgment inside the room did poorly against a dull test. A study of job testing across 15 firms hiring low-skilled service workers found that managers who hired against the test’s recommendation ended up with worse hires on average — evidence that discretion was more often bias than superior private information. The interview was defended as the place where the humans caught what the numbers missed. Sometimes it was. On average, it was the place where the numbers caught what the humans got wrong.
And the regulators knew. The pre-transition US guidance on employment tests and selection procedures already required that any screen with a disparate impact be shown to be job-related and consistent with business necessity. The rule was right. Enforcement was complaint-driven, and almost nobody who was quietly not called back knew they had been wronged.
What changed, and what did not
The interview did not end because someone banned it. It ended because the Automation Dividend removed the thing it was really rationing.
| The permission interview (pre-2030s) | The fit conversation (2047) | |
|---|---|---|
| What is at stake | Rent, healthcare, standing | Suitability for one piece of work |
| Who bears the cost | Applicant, unpaid, uncounted | Both parties, inside paid time |
| Effect of rejection | Income loss, spell length penalty | The work does not happen |
| Duration signal | Time unemployed read as defect | Not recorded, not readable |
| Record created | Private, adversarial, unappealable | Contribution Record entry, read-only |
| Right to leave | Nominal | Real, because the floor is paid |
The fit conversation is what remains: short, mutual, and consequence-free on the downside. If it leads to work, the work appears in the person’s Contribution Record — the Ministry’s account of chosen contribution, which is never a condition of the dividend and cannot be used to reduce it. If it leads nowhere, nothing is written anywhere. There is no spell to measure, because there is no spell.
This is the part I did not expect when we were arguing about the accounting in the mid-2020s. We thought we were solving an income problem. What we actually did, almost as a side effect, was end a specific kind of humiliation: the performance of eagerness by an adult who cannot afford to be honest. My mother describes the difference precisely. She is seventy-three, she sits unpaid on Tuesdays with families arguing with the health service, and she says the difference between that and her old job is that when she disagrees with someone now, she is still disagreeing an hour later.
The cost we did not remove
Here is what the dividend did not fix.
It did not end discrimination. The same concentrated preference for familiar names, familiar schools and familiar faces survived intact into the fit conversation; the large firms identified in that 2021 evidence still exist, still choose, and still choose badly. We took away the survival stake attached to being passed over. We did not take away being passed over, and the two are not the same wound.
It did not end ranking. People still want to be chosen, and the desire is not irrational; it is how we know our judgment is worth something to somebody. The Ministry has instruments for yield, for displacement, for allocation and for contribution. It has none for the feeling of writing eleven hours of argument for your own worth and receiving one minute of silence. Capacity is not contribution, and work that must be done to survive is not chosen work — but neither line reaches the man refreshing an inbox at two in the morning.
What we did was narrower than the slogans and larger than it sounds. We stopped charging people, in unpaid hours and swallowed dignity, for the right to be considered. The rest of it — being wanted, being good, being picked — is still out there, unmetered, doing what it always did.
FAQ
Why did the job interview end? Not because a machine replaced the interviewer. It ended because the interview’s main function stopped being needed. A hiring screen decided who was allowed to earn, and once the Automation Dividend made income independent of being selected, the screen no longer rationed survival. What remains is a conversation about fit for a specific task, which people can leave without consequence.
What is the admission cost? The Ministry’s term for the unpaid hours a labour market charged a person for permission to earn: applications, tailored résumés, portfolio work, unpaid tests, travel, waiting. It was never invoiced, never appeared in any national account, and fell hardest on people who were screened out most often — so the people with the least income paid the largest admission cost.
Was hiring ever actually accurate? Less than it believed. A 2003 field experiment found identical résumés with white-sounding names received 50 percent more callbacks, and that improving a résumé raised callbacks 30 percent for white applicants but barely moved them for Black applicants. A 2012 experiment found callback probability fell steeply with time unemployed, mostly within the first eight months, independent of the applicant’s skills.
What replaced the interview? The fit conversation: a short, mutual, low-stakes discussion about a specific piece of work, recorded in the Contribution Record if it leads anywhere and in nothing at all if it does not. It ranks people for a task. It does not rank them for existence, because the floor is already paid.
Does the Contribution Record work like a résumé? No, and the difference is the whole design. A résumé was an argument for being allowed to live decently. The Contribution Record is a read-only account of what a person chose to do, never a condition of the dividend and never usable to reduce it. A record that gates money gets gamed; a record that gates nothing stays roughly honest.
Where does this design still fail? Ending the interview did not end discrimination — the same 2021 evidence that identified biased employers showed bias concentrated in a small set of large firms, and those firms still exist. Preference for the familiar survived the abolition of the screen. The Ministry removed the survival stake from being rejected. It did not remove rejection, and it has no instrument for what rejection does to a person.