On the Conservation of Judgement, the Two Ledgers of AI, and Why the Only Body That Can Fix This Is the One We Trust Least
By Dr Luke Soon
Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol
July 2026

Here is the sentence I have been circling for months, and I have not found a way to soften it.
The only institution on earth with the mandate, the reach and the time horizon to manage what is coming is government. And it is the institution we currently trust the least.
That is the whole essay. Everything below is the working out.
I. THE ARITHMETIC NOBODY CONVERTS TO A DAY
Begin with a number, because philosophy without arithmetic is just mood.
The better forecasting houses now converge on a figure for the back half of this decade. Beginning around 2028, somewhere north of thirty two million jobs a year will require reconfiguration, redesign, splintering or fusion.
Nobody quotes it, because thirty two million a year sounds like a statistic. So divide it by a working day.
Roughly one hundred and fifty thousand jobs evolving daily through upskilling. Roughly seventy thousand more requiring outright rewriting.
Every day. Not once. As a standing condition.
Read that again.
Not thirty two million jobs destroyed. Thirty two million jobs changing shape, annually, forever, or at least for as long as anyone reading this will be working.
Now consider the instruments we intend to meet this with. Redundancy law designed for events. Training systems designed for cohorts. Human resources operating models designed for annual cycles. Ministries organised around five year plans and four year electoral terms.
We are bringing a calendar to a stopwatch fight.
That mismatch, not automation, is the actual emergency. The technology is not moving faster than we can learn. It is moving faster than we can legislate, and those are very different problems with very different owners.
II. THE MAP IS NOT A MENU
There is a four scenario map doing the rounds in boardrooms. It plots two variables: how much autonomy an organisation surrenders to AI, and how far it is willing to transform the work itself. It yields four futures.
Fewer workers doing what the machines cannot. Many busy workers doing more with machines. Many innovative workers pushing past the frontier with machines. And, in the far corner, almost nobody at all, running an enterprise that has been rebuilt around autonomy.
Every executive I show it to does the same thing. They point at a quadrant and say: that one.
They have misread it entirely.
The genuinely useful idea buried in that map is not the quadrants. It is the ripple: the unbudgeted, unintended, entirely predictable consequences that fire the moment AI changes how work is done.
Take the most ordinary example available, the contact centre.
You choose the efficiency corner. Deflect the volume, reduce the headcount. Clean. Defensible. Board approves.
Then the ripples arrive.
The number of people serving customers falls, and the number needed to train, tune and extend the machines rises. Your experienced staff, the ones you were quietly hoping to release, turn out to be the only people who can teach the system the long tail, so they are redeployed into becoming exception handlers of extraordinary sophistication. The surviving human role drifts upward into designing the experience itself. And within eighteen months the service representative, silicon or carbon, is carrying a revenue target.
One decision. Four staffing realities. None of them in the business case.
The map is not a menu. It is a metabolism. You do not choose a quadrant. You choose an entry point, and the system digests you through the rest.
III. THE FIRST LAW
Which brings me to the proposition I want to put on the table properly, because I think it is the load bearing idea for the next five years.
Automation does not reduce the quantity of human judgement an enterprise requires. It relocates it.
Call it the conservation of judgement.
When you automate a task, the judgement inside that task does not evaporate. Judgement is not a cost line. It is closer to energy. It moves into the specification of what the machine should do. Into the supervision of whether it did it. Into the handling of what it could not. And into the governance of what happens when it is confidently, fluently, scalably wrong.
The task leaves the org chart. The judgement does not. It reappears somewhere more senior, usually unbudgeted, and generally unnoticed until quality quietly collapses and everyone blames the vendor.
This is why the ripples are not an oddity. They are the necessary consequence of a conservation law. You cannot delete judgement by automating its container.
And it explains the most quietly alarming statistic in this year’s labour data.
Analysis of more than a billion job advertisements across twenty seven countries finds that entry level roles in AI exposed occupations are now seven times more likely to demand traditionally senior capabilities. Judgement. Leadership. Strategic thinking. Stakeholder communication.
Those roles grew thirty five per cent since 2019. Other entry level roles fell ten per cent.
Look at what has happened. The judgement did not disappear when we automated the junior tasks. It was pushed downwards, into the job description of the very person who no longer has access to the tasks through which judgement was historically acquired.
I called this mechanism seniorisation in an earlier instalment, when it was still a hypothesis. It is now a hiring standard.
We are advertising for the output of an apprenticeship whose apprenticeship we have automated.
There is a word for asking someone to arrive already possessing what they were meant to learn from you. Several words, in fact. None of them are flattering.
IV. TWO LEDGERS
So. Is AI disrupting human work, or augmenting it?
I now think this is a badly posed question, and I can show you why using two datasets that appear to contradict each other and do not.
The first. A large linked study matching survey responses to actual usage for thousands of frequent AI users found overwhelming positives. Eighty six per cent report gains in speed, eighty two per cent in scope, sixty nine per cent in quality. Sixty eight per cent say they are learning more. Fifty seven per cent say AI has made their skills more valuable.
Then the finding that turns the received wisdom inside out. The people who delegate most aggressively are the most optimistic, across every dimension measured: pay, job security, employability, meaning, autonomy, even human interaction. The predicted hollowing out simply does not appear. The heaviest delegators report learning at the same rate as everyone else.
The second. Brynjolfsson, Chandar and Chen at Stanford, working from payroll records across more than seven hundred occupations, find that workers aged twenty two to twenty five in the most AI exposed occupations have suffered a relative employment decline in the region of thirteen to sixteen per cent. For young software developers, headcount is down roughly a fifth since late 2022. Their older colleagues in the same occupations are fine.
The effect has not mean reverted. Through April 2026, nearly four years of data, it has grown at roughly half a percentage point per month. Strip out the entire technology sector. Isolate remote work. Control for interest rates. It holds every time.
And critically, it is not a redundancy story. It is a hiring story. The vacancies simply stopped appearing.
So which is it?
Both. In the same motion. Because they are the same transaction, booked twice.
Every act of delegation is entered as augmentation on the incumbent’s ledger and as disruption on the entrant’s. The senior analyst who hands the first draft to a machine experiences leverage, and is telling you the truth. The graduate who would have written that draft experiences a vacancy that was never advertised, and is also telling you the truth.
Augmentation is disruption with a beneficiary.
The graduate is not being replaced. She is being un-hired, which is considerably worse, because there is no severance for a job that never existed, no notice period for an absence, and no statistic that counts her.
This is the direct continuation of the argument I have been making for a year now: the AI productivity dividend is not a paradox, it is a transfer. The dividend is entirely real. It is simply financed, and the financing is invisible because it takes the form of work never commissioned rather than work terminated. Shadow jobs. The roles that quietly failed to be born.
Here is the whole thing in one picture.
HUMANS WANT AI TO DO THE WORK
| FEWER WORKERS | ALMOST NO WORKERS |
| doing what AI cannot | an enterprise rebuilt |
| | |
| Incumbent ledger: + | Incumbent ledger: + / minus |
| Entrant ledger: 3 minus | Entrant ledger: 4 minus |
| Judgement relocates UP | Judgement relocates OUT |
| MANY BUSY WORKERS | MANY INNOVATIVE WORKERS |
| using AI to do more | pushing past the frontier |
| | |
| Incumbent ledger: + + | Incumbent ledger: + + + |
| Entrant ledger: 1 minus | Entrant ledger: + |
| Judgement relocates ACROSS | Judgement COMPOUNDS |
HUMANS WANT TO DO THE WORK WITH AI
Work unchanged <

Answer it soon. The clock is not a metaphor. It is a hundred and fifty thousand jobs a day.

Dr Luke Soon is an AI ethicist, futurist and computer scientist based in Singapore, and the author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: The Superintelligence Protocol. The views here are his own.