The Signal Collapse

The Rung That Is Being Removed

Three different measures, each against its own baseline (peers, January 2023, 2019). They are not additive and not ranked against each other.

Five Findings

The Cliff: How Fast Entry-Level Work Is Actually Disappearing

For most of the past three years, AI's effect on entry-level hiring has been debated mostly as a prediction — an executive warning, a forecast, a fear about what might happen next. Across the second half of 2025, several datasets, built from different methods, converged on the same finding.

It is no longer a prediction. Entry-level hiring has already contracted, and the contraction appears concentrated where AI substitutes for junior work rather than assists it.

Three different measures, each against its own baseline (peers, January 2023, 2019). They are not additive and not ranked against each other.

The Most Rigorous Dataset Is Stanford’s

The most rigorously designed of these is the Stanford figure. Drawing on high-frequency ADP payroll records covering millions of U.S. workers, the researchers compared employment trends for workers in occupations most exposed to generative AI against workers in less-exposed occupations — and, within the same exposed occupations, compared younger workers (22–25) against older, more experienced colleagues.

Even after controlling for firm-level shocks — ruling out the possibility that a handful of struggling companies were driving the whole result — early-career workers in AI-exposed roles such as software development, customer service, and accounting now sit 19% below where they would be had they kept pace with their less-exposed peers, while more experienced workers in the identical roles show no comparable gap.

That divergence has widened over the year: the authors’ August 2026 update puts it at 19% on payroll data running through June 2026, against 15% on the same measure at the July 2025 data vintage.

What the Design Isolates

This design matters because it isolates something anecdote cannot: the decline does not look like general economic softness spread evenly across the workforce. It appears concentrated where two conditions overlap — the worker is young, and the task is one AI can substitute for rather than assist.

The authors are careful about what this is: early descriptive indicators, not causal estimates.

They note the pattern attenuates when controlling for education and is more pronounced in their payroll sample than in national survey benchmarks. We report it on the same terms.

Two Other Providers Describe the Same Shape

Two independent data providers, using entirely different methods, describe the same shape from different angles. Revelio Labs, which tracks live job postings rather than payroll outcomes, finds U.S. entry-level postings that typically require at least some college education down roughly 35% relative to January 2023, in the company’s August 2025 analysis — with highly AI-exposed entry-level roles specifically falling even faster than the market average.

And in the sector historically most associated with graduate opportunity, SignalFire’s 2025 report, tracking hiring at the top 15 technology companies by market capitalisation, finds new-graduate hiring down by more than half since 2019.

This Is Not a Story About a Shrinking Labour Market

None of this means employment is collapsing broadly. The wider U.S. labour market absorbed 1.2 million employer-announced job cuts in 2025 — up 58% year-on-year, the highest annual total since the pandemic year of 2020 — but AI is one cited driver among several, alongside corporate restructuring, government downsizing, and tariff-related cost pressure.

The sharper, more specific signal sits with the entry-level data above: this is not a broad recession story. It is a story about where, inside a functioning labour market, one particular rung of the ladder is being removed.

By 1 April 2026 the establishment had fallen to approximately 188,000, a reduction of more than 8,000 posts, with a further 2% cut confirmed for each of the 2026/27 and 2027/28 financial years and a cumulative target of more than 10,000 posts by the end of the current government term.

The detail that matters for this report is not the total but the grade mix: of the roughly 4,987 posts deleted in the 2026/27 estimates, the Civil Service Bureau puts approximately 57% at junior grades and 40% at middle grades — leaving about 3% at senior level.

The driver here is fiscal consolidation, not AI. But the shape of the cut is the same shape the U.S. payroll and postings data describe, arriving from a second, entirely independent direction: the entry rung and the rung above it appear to be what is removed first.

The aggregate numbers establish that the cliff is real. They do not answer the more useful question for anyone managing talent today: within this generation of new entrants, who is actually absorbing the impact — and is it evenly spread, or concentrated in a specific, identifiable band of the workforce?

Who Gets Hurt Most: The Squeeze on the Middle

The last section asked how fast entry-level work is disappearing. The more useful question is which work — and who it used to belong to.

The work being removed is the most ordinary work in the building.

The honest answer isn't dramatic, and that's exactly the point: it's ordinary office and administrative work — processing forms, handling paperwork, reconciling records, first-pass coordination — that has quietly made up one of the largest categories of entry- and mid-level jobs in many organisations.

Revelio’s posting data point in the same direction: highly AI-exposed entry-level roles are down more than 40% relative to early 2023, in the same August 2025 analysis. Routine paperwork is the kind of work generative AI is best placed to substitute for, so this is plausibly where much of that decline sits — though we have not seen it broken out by task category.

Not the Extremes

When researchers break the decline down by the tier of university a junior worker attended, it isn't the extremes that absorb it. Hosseini and Lichtinger, using résumé and job-posting data covering more than 60 million workers, compare firms that adopted generative AI with firms that did not. At the adopting firms, they find the steepest relative declines in junior employment among graduates of solid (Tier 3) and less selective (Tier 4) institutions, and smaller declines for graduates of elite, strong and lowest-tier schools. The data come from Revelio Labs, the same provider as the postings figures above.

Why the extremes fare better is interpretation, not a result of the study. One reading is that the most elite graduates compete on a different axis AI hasn’t yet touched, and that work at the lower tier may more often need a physical presence AI can’t yet replace. The study does not establish either.

In this data, it's the broad, ordinary middle that takes the hit: the group that makes up the bulk of most graduating classes.

(Economists have a name for a milder version of this shape — job polarization, the decades-old finding that middle-skill work shrinks while work at the top and the bottom grows relative to it, Autor & Dorn, 2013 — but the version showing up here is sharper. It isn't wages compressing in the middle. It's the entry point itself disappearing.)

Not everyone who becomes exceptional looks exceptional at twenty-two.

The people losing this work are the same people the résumé was already failing.

This is also where the report's two threads meet. The people in this ordinary middle never had an elite pedigree to open doors, and they weren't competing on hands-on, physical work either. What they had was the paper trail — a decent degree, a clean résumé, steady experience that added up to “capable and reliable.” That paper trail is exactly the mechanism the next section shows is now failing them a second time.

One well-documented career shows what this filter would have discarded.

Kazuo Inamori is a useful check on this argument. He founded Kyocera, later founded DDI (now part of KDDI), and in his late seventies was asked to come out of retirement to rescue Japan Airlines from bankruptcy — restoring it to profitability within two years.

But his own biographers and his company's official history describe a young man whose early years gave no clear signal of what was coming: modest regional roots, early setbacks, a first job as a young researcher at a small, unglamorous ceramics maker in Kyoto that nobody would have flagged as the launchpad for one of Japan's most consequential business careers.

What followed took years, not one interview or one line on a résumé — persistence, a management philosophy built through trial and error, the willingness to keep showing up in an ordinary job long enough to become extraordinary at it. None of that is visible on day one. It only becomes visible in hindsight, built from years of exactly the kind of ordinary work this report has been describing.

If the ordinary middle is being squeezed out of the jobs that used to prove people like this, the next question is whether the one tool left to identify them — the résumé itself — can still do the job.

What's Left When Both Sides Use AI

The same technology now used to decide who gets cut is also being used to decide who gets hired — and, increasingly, to write the application in between.

Two shifts, each real on its own, compound into something more serious together.

Employers Changed What They Screen For. The Ground Is Still Moving.

On the hiring side: most large employers have already moved past résumé and GPA screening toward skills-based hiring. 70% now report using it, up from 65% a year earlier — applied most heavily at exactly the two stages that matter, screening and interviewing.

The World Economic Forum puts a number on how fast the ground is shifting under that judgment anyway: employers expect 39% of workers’ core skills to change by 2030.

On the firing side, the same technology has gone from a minor factor to the leading one within a single year. In January 2026, AI was cited in 7% of announced U.S. layoffs. By March it had become the single most-cited reason for job cuts in America, ahead of market and economic conditions for the first time on record. By May it accounted for roughly 40% of that month's cuts — 38,579 of 97,006 — the highest monthly total recorded since tracking began in 2023, and the third consecutive month in which it led all stated reasons.

(A caveat worth keeping honestly in view: some of this is what commentators call “AI washing” — companies using AI as convenient cover for cuts they would have made anyway. Both things can be true at once. Either way, “AI” is now the label companies themselves reach for first when explaining who they no longer need.)

The sharper problem is not the volume of jobs. It is the instrument used to allocate them.

Put the hiring side and the firing side together and a narrower, sharper problem emerges. It isn't about jobs disappearing. It's about the tool used to sort who gets the ones that remain.

Both sides now use generative AI in the process — employers to screen résumés, candidates to write them. Each side adopted the technology for the same reason — speed. The result is a closed loop: AI writing, being read by AI, at scale, on both ends.

A 2025 study out of the University of Maryland, National University of Singapore, and Ohio State University tested exactly what that loop does to a candidate's odds. Researchers ran matched pairs of human-written and AI-generated résumés through a set of major commercial and open-source models acting as evaluators — GPT-4o, DeepSeek-V3 and LLaMA among them — across 24 occupations.

Among the larger models tested, the researchers measured a self-preference bias of 67% to 82% in favour of résumés the model had written itself over human-written ones. Smaller models showed weaker effects.

When the researchers then simulated realistic hiring pipelines end to end, candidates whose résumés had passed through the same AI model as the evaluator were 23% to 60% more likely to be shortlisted than equally qualified applicants who submitted human-written résumés — the gap widest in business-related roles. In the researchers’ account, the reason was not that the AI-shaped résumé was more qualified. It was that it sounded more like the system reading it.

This does not look like a formatting problem, and a better prompt is unlikely to solve it.

And the same broken signal is now being read by the same machine everywhere at once.

A biased screen at one employer is bad luck. A biased screen at every employer a candidate applies to is something else, and until recently the data to tell the two apart was scarce.

In May 2026, researchers at Stanford, Chapman and Northeastern published the first large-scale study of algorithmic hiring conducted on live data rather than simulations. They followed 3.4 million people submitting 4 million applications to 1,700 job postings across 150 employers in 11 industry sectors. Every one of those applications was screened by algorithms built by a single third-party vendor.

Algorithmic screening is now widespread among U.S. employers — an estimate attributed to the World Economic Forum puts it above 90% — and many rely on the same few vendors, which is precisely the condition the study was designed to test.

Two findings matter for this report, and the second is the easier one to overlook.

The first is that rejection stopped being independent. Applicants who submitted multiple applications to positions screened by that one vendor were rejected from every position at a rate higher than would occur if those employers were deciding independently of one another. Ten percent of applicants who submitted four applications were rejected from all four.

The researchers then ran the same test against the largest previous study of hiring outcomes — 83,000 applications to 108 Fortune 500 firms over a similar period, with no particular focus on AI — and found no such effect there: rejection-from-everywhere ran in line with what independent decisions would predict.

The difference does not appear to be the candidates. It is that one judgment is now being applied in many places simultaneously.

The second finding is about measurement itself, and it generalises well beyond hiring. The researchers tested the vendor's recommendations for adverse impact under the EEOC's four-fifths rule. Pooled together — treating the vendor as one enormous hiring process — no adverse impact appeared. Examined position by position, as an adverse impact analysis is normally conducted, it appeared in many positions.

The same data, the same tool, and the disparity was visible or invisible depending on the level at which the results were aggregated. Averaging across a warehouse role and a finance role lets one pattern cancel the other, and the aggregate looks clean.

A proxy only works while producing it requires the thing it stands for.

A résumé was always a proxy — a record of what someone has already done, on the assumption that past output predicts future capability. That assumption held reasonably well when producing the record required the underlying skill. It stops holding the moment the record itself can be generated independently of the skill it's supposed to prove.

Skills-based hiring was meant to be the fix for exactly this — proficiency instead of pedigree. But if the “skill” itself can now be demonstrated by a machine on someone's behalf, the fix inherits the same flaw it was built to solve.

If the résumé — humanity's oldest hiring signal — is being read by machines and increasingly written by them too, the real question is no longer how to screen better. It's what AI may structurally be unable to do, however well it's prompted, that still has to be found some other way.

The Responsibility Gap

AI can generate a thousand ideas before lunch. It has never once had to live with what happens after someone acts on one.

The résumé analysis ended on a question: what is it, exactly, that AI may structurally be unable to do — however well it is prompted? The answer starts with something anyone who has used AI for more than a few months has already noticed.

It gets something wrong. It apologises. It tries again — and sometimes gets the correction wrong too, in a slightly different way. The loop is so familiar by now that most people barely register it.

But notice what the apology actually is. It costs the system nothing to say it. Nothing is at stake, and nothing changes in how the system operates because it said the word. It is language that resembles accountability, without any of the weight behind it.

Once the stakes rise, the absence of a bearer of consequence stops being a curiosity.

The stakes are not always this small. A public database tracking AI-fabricated content submitted to courts worldwide had logged roughly 1,490 cases by May 2026 — more than 1,000 in the United States alone — and was still climbing.

In one of the most-cited recent rulings, a U.S. federal appeals court fined two lawyers $15,000 each after their filings contained more than two dozen fabricated case citations. The court did not find that AI had been used, but it cited the principle that it makes no difference whether a citation came from generative AI or any other source: the lawyer who signs the filing is responsible for verifying it.

In a separate case, the Nebraska Supreme Court suspended an attorney from practice until further notice after a brief in a divorce appeal was found to contain fictitious cases or misquotes in 57 of its 63 case references.

Asked directly by a justice whether he had used AI, the attorney said no. Two days before the suspension order, he admitted in an affidavit that he had, calling it a “grave error of judgment.”

In both cases the consequence landed on the person who signed the filing, not on the tool.

This is unlikely to be a bug that the next model update removes. It follows from how these systems work.

AI can generate an almost unlimited number of ideas, drafts, and options. What it cannot do is answer for the result — take one of those options, act on it in the real world, and be the one left holding the outcome when it goes wrong. Execution is where ideas turn into results. It is also where responsibility starts, because someone has to answer for what the action actually did.

This is not a new observation. It has had a name in the literature for twenty years.

Two academics, two decades apart, named different pieces of this same problem.

In 2004, computer scientist Andreas Matthias pointed out that once a learning machine's behaviour can no longer be predicted even by the people who built it, the traditional way of assigning legal and moral responsibility has nowhere left to attach. He called this a responsibility gap.

In 2016, philosopher John Danaher pushed the idea further. Even when we can trace a harmful outcome back to a specific AI system, he argued, punishing or blaming that system satisfies no one — it has no remorse to feel and nothing at stake to lose. He called this the retribution gap: correctly identifying what went wrong, with nowhere real to direct the consequence. It is the academic name for the apology loop above.

Put together, these point at something more concrete than a philosophical puzzle. “Responsibility” is not one undifferentiated thing. At least two distinct jobs have to be done every time AI is involved in a decision.

Someone has to catch and correct the error — because AI will make one, regularly, and it may not know it has. And someone has to own the consequence once an action is taken, whether it succeeds or fails. AI is genuinely useful for the first half of almost any task: generating options, drafting, calculating, predicting. It has no mechanism for the second half of either job.

That gap has a shape. The next section names four specific qualities inside it — the parts of judgement and accountability that, on the evidence so far, AI does not do and a person still has to.

Judgment & Accountability: The Four Things a Résumé Can't Show You

Many people are learning how to talk to AI. Far fewer are discussing what it takes to be the one AI answers to.

The responsibility gap named in the last section has a shape. And the management world, without quite naming that shape, is already circling it.

A June 2026 McKinsey piece on HR describes an emerging role it calls the “orchestrator”: someone who designs and oversees human–agent workflows, deciding where to automate, where to augment, and where human judgement is irreplaceable. The role is being named.

Naming the role is the easy half. Finding the person is the hard half.

But naming the role is the easy half. The harder problem is the one this report has been building toward ever since the résumé stopped working: how do you actually spot who has what the role needs — not through a résumé, and not through a skills test that only measures the layer AI has already learned to fake?

Much of the current investment in AI skills goes into the same layer: prompting. How to phrase a request, how to chain instructions, how to get a cleaner output. That is a real, useful skill.

But it sits on top of something résumés, skills tests, and training courses all rarely touch — capability: the deeper, more transferable capacity that shows up precisely when there is no prompt template for the situation in front of you, no SOP, nothing to look up.

One common way to think about the layering underneath this is as follows. Knowledge is what you know. Skill is what you can reliably do — and, importantly, skill can be taught, standardised, and copied. Give someone a script and enough practice, and their execution will rarely be excellent, but it also won't be disastrous.

Capability sits below both. It is harder to see, harder to train quickly, and it is what decides what happens the moment the script runs out.

Neither Layer Works Without the Other, and Organisations Often Fund Only One

None of this makes skill unimportant. A capable person with no skill can be dangerous — good judgement, delivered through clumsy execution, can still produce a bad outcome. A skilled person with no capability is fragile — reliable until the situation stops matching the script.

The organisations that get this right build both, deliberately, at the same time. Two real examples show what that looks like when it works, and what happens when a company tries to skip the harder half.

Put these two side by side and a pattern comes into focus. What separates them isn't technology, or even skill. It's whether a human being was still positioned to notice, judge, and answer for what happened — and whether the organisation had actually selected for that person in the first place.

So if that's the real difference, what specifically should an organisation be looking for? Watching for it in practice, four qualities keep recurring — not a finished theory, but a pattern we think deserves a wider conversation. Here they are in full, before we unpack each one:

Four qualities worth looking for

• 1

  Correction

  corrected by reality, and changes because of it

  A bullet point can claim “learned from setbacks.” It cannot show the moment the setback actually landed, or what changed afterwards.

• 2

  Connection

  connects what doesn't obviously connect

  A career that zigzags across industries can read as inconsistency to a screen, and be filtered out before anyone asks what it taught.

• 3

  Commitment

  commits before the evidence is complete

  A résumé can list the decisions someone was involved in. It cannot show what they were willing to lose if they were wrong.

• 4

  Ownership

  stays to answer for it

  A résumé lists outcomes. It essentially never shows whether the person stayed to own the one that went badly.

Each one shares the same problem, and it's the problem this whole report has been circling: none of them reliably shows up on a résumé. Not because résumés are badly designed, but because a résumé was only ever built to show what someone has already done — and every one of these four qualities is about what someone does in a situation their record has never covered before.

① Correction — corrected by reality, and actually changes because of it.

The apology loop from the last section never resolves anything, because the system doesn't update the way a person does. Someone who watches a pitch fail in the room — the specific silence, the question that landed badly — carries that forward into the next one, often without being able to fully explain how.

A model that got something wrong yesterday has no equivalent loop; ask it again tomorrow and, unless someone has retrained it, it may fail the same way, the same day, for the tenth time.

This is hard to see on a résumé. A bullet point can claim “learned from setbacks.” It cannot show you the moment the setback actually landed, or what changed afterward — and an AI-polished bullet point makes that claim exactly as confidently whether or not it's true.

② Connection — connects what doesn't obviously connect.

The most useful ideas in most organisations rarely come from someone who stayed in one lane. They come from someone who spent time somewhere else — a different industry, a different function, a different kind of pressure — and noticed that a pattern from over there applies over here.

This may be the quality most easily penalised by conventional screening, not just overlooked by it. A career that zigzags across industries can read, to a keyword-matching ATS and to a time-pressed human recruiter alike, as inconsistency — a candidate who “can't decide what they want to do,” filtered out before anyone asks what that zigzag actually taught them.

The same nonlinear background that a screening system reads as noise can be a source of the insight a linear career does not produce.

③ Commitment — commits before the evidence is complete.

Nassim Taleb's Skin in the Game (2018) makes a related point about risk: a decision made by someone with nothing to lose from being wrong is worth less than the same decision made by someone who bears the consequence.

AI can generate a thousand plausible options and a probability for each. It cannot pick one and commit — because committing means closing off the other 999, and there is no one inside the system for whom that closure costs anything. A résumé can list the decisions someone was involved in. It cannot show what they were willing to lose if they were wrong, because it was written after the fact, by someone who already knows how the story ended.

④ Ownership — stays to answer for it.

This is where the responsibility gap stops being structural and becomes personal. Being willing to be the name attached to a decision that turned out wrong — not just technically liable, but actually answerable to the people affected by it — is, on everything gathered in this report so far, still something only a person can do. This, too, is hard to see on paper. A résumé lists outcomes. It essentially never shows whether the person stayed to own the outcome that went badly, or quietly moved on before anyone had to ask.

• 1Correction
• 2Connection
• 3Commitment
• 4Ownership

Where each of the four is actually read

The obvious objection to all four is that they are soft — the kind of thing everybody claims to select for and nobody can measure. That objection is half right.

They are hard to measure by questionnaire, because a questionnaire asks the person to report on themselves, and self-report is exactly the layer AI has already learned to produce. But they are observable in behaviour, provided someone is watching real work over time and recording it against a written standard rather than an impression. That is a different problem from an unmeasurable one, and possibly a solvable one.

Our own instrument, VITALIS, is designed to read them in people already doing the work, not in candidates — and to read them in more than one place.

VITALIS reads a person on three axes rather than one score. X — the Capability Skeleton is what the person can currently deliver. Y — the Growth Engine is the rate at which that capability is changing. Z — Core Vitality is how much of X and Y actually reaches the work. The four qualities of this report do not sit in one place; they distribute across the three, which is itself worth noting.

The specific dimensions inside each axis, their scoring ladders and their evidence requirements are taught in VITALIS Foundation certification and are not published here. What matters for this report is the structural claim: these four are read from observed behaviour against a written standard, by a trained and calibrated assessor — not inferred from a document the person wrote.

If these four qualities matter for whether someone can grow into a role AI is unlikely to fill, it should show up somewhere concrete — not just in what companies say they value, but in where the jobs that are actually still growing sit.

What Is Still Growing: Not Every Job Is Disappearing

The jobs still growing keep needing the same thing: someone who can be trusted to be there, judge the situation, and answer for what they decide.

The last section ended with a specific, testable claim: if Correction, Connection, Commitment, and Ownership are what AI does not appear able to replicate, that should show up somewhere concrete — not in what companies say they value, but in where the jobs that are actually still growing sit. This section looks at whether the data is consistent with that claim. It cannot confirm it.

Everything up to this point has been about what's disappearing — entry-level and routine work contracting, a résumé that can no longer reliably show who has what's left. All of it real. But it is only half a picture: it says nothing yet about what's growing, or why. That's the other half of the test.

Where Hiring Continues, a Person Often Has to Be in the Room

Healthcare added 22,000 jobs in June 2026, below its average monthly gain of roughly 38,000 over the prior twelve months. That is a slower month than the recent average, not a reversal.

Nurse practitioners specifically are now projected among the fastest-growing occupations in the entire U.S. economy over the next decade — the latest BLS projection (2025–35) puts employment growth at 41%.

Consider what a nurse practitioner, a hotel concierge, and a hospice social worker have in common. Not the industry. Not the pay grade. What they share is that the job is difficult to do without someone physically present, making a judgement call, and being the person who answers for it if the call is wrong. (Only the first example is backed by projection data cited here; the other two are illustrations.)

Inside the Growth Data: For Capable Workers, Augmenting AI May Enlarge the Job

There's a second thing happening inside this data that's easy to miss if you're only counting headcount.

One working paper (Marguerit, 2025) that separates “automation AI” (which substitutes for a worker) from “augmentation AI” (which extends what a worker can do) finds a pattern: automation AI reduces employment and wages in lower-skilled work — the pattern already documented in the first half of this report. For higher-skilled workers the pattern differs: augmentation AI is associated with the emergence of new work and higher wages. The study covers the United States, 2015–2022, and does not show that these gains offset the losses elsewhere.

This matches what capable people report about using AI well, long before a study examined it. Used well, AI may not make a good employee's job smaller. It can remove the parts of the job that never needed a person in the first place, and hands back the time for the parts that did — the conversation that needed judgement, the edge case nobody had written a procedure for, the higher-difficulty work there was never previously time to do properly.

Someone with the four qualities from the last section, working alongside AI rather than being replaced by it, may become more valuable to an organisation than before. Not less busy. Busier, with work that actually needs them.

If all of this holds — the cliff is real, the middle is squeezed, the résumé can no longer show who has what's left, and what's left may be worth more than it has been — the last question is the one this report has been circling ever since the résumé stopped working: once you can no longer trust a résumé to answer it, what should an organisation actually measure instead?

A Different Starting Point: What We Measure Now

The question is not how to screen faster. It is what a résumé was never built to show.

The previous section ended with the question this report has been building toward since its opening pages: once a résumé can no longer be trusted to answer it, what should an organisation actually measure?

What Has Been Established

Here is the shape of what's been established along the way. Entry-level and routine work is contracting, and the contraction appears concentrated in the ordinary middle — not the most elite candidates, not the lowest-tier graduates, but the broad group in between who used to trade steady, procedural competence for a stable start.

The résumé itself, now read by AI on one side and increasingly written by it on the other, has stopped functioning as a reliable signal of much beyond how well someone can operate the tools that produce résumés.

AI has a structural gap sitting exactly where responsibility should be: it can generate options endlessly, but it is not corrected by living through being wrong, does not connect what it hasn't personally lived, does not commit before the evidence is complete, and does not stay to answer for what happens next.

And many of the jobs still growing are ones that require those four things — Correction, Connection, Commitment, Ownership — from an actual person.

Put together, this is a specific, practical problem, not a philosophical one.

A faster screen cannot fix a broken signal, because the signal is the part that broke.

If judgement and accountability are what's genuinely scarce, and the tool historically used to spot who has them has stopped working, the fix is not a faster or smarter résumé screen. A faster screen still finds the same signal, only quicker — and the signal is the part that broke.

What's actually needed is a way to observe the four qualities directly: whether someone can be corrected by reality and actually changes because of it; whether they connect what doesn't obviously connect; whether they'll commit before the evidence is complete; and whether they'll stay to answer for it. Little of that shows up in a work history. All of it can, in principle, be observed — and it is worth the effort of building a way to observe it properly, rather than continuing to ask AI to read a document that was never built to contain it.

The group with the most to gain from a better instrument is the group this report opened with.

This matters most for exactly the people this report has spent the most time on: the ordinary middle, filtered out early by a résumé that was never built to see what they might become. A way of seeing past the credential and the linear career — of observing these four qualities in the work itself rather than inferring them from a document — would hand back something the résumé quietly took away: the chance for someone whose capability does not show up on paper to have it noticed once they are doing the job.

Kazuo Inamori's résumé, if he'd had reason to write one at twenty-two, would have shown a failed entrance exam, a regional university, and a first job at a small, unglamorous ceramics maker nobody had heard of. Nothing on it would have predicted Kyocera, DDI, or a man in his late seventies coming out of retirement to pull Japan Airlines back from bankruptcy.

What separated him from other unremarkable graduates that year was not visible on paper. It may have been something closer to Correction, Connection, Commitment, and Ownership — and it took him decades of ordinary work to show it. An organisation able to see that in the ordinary course of his work — years earlier than the record eventually showed it — might have known something about him that no document was going to reveal.

That is not a finished answer. The four qualities named here are a working position, not a settled theory, and the coverage gaps are marked where we found them.

But they are not unmeasurable by every method. They are hard to measure by questionnaire — which is a different problem, and a narrower one.

VITALIS is not a hiring instrument and is not used to make employment decisions. It addresses a different point in the sequence: once someone is already in the work, whether an organisation can see what is actually happening, early enough and specifically enough to act.

It does that by recording observed behaviour against written standards, repeatedly, with the named work instances kept alongside the reading. Founding deployments will test whether that record holds up between assessors and over time.