The Cost of the Ledger, Part II: The Last Job

Opening

In Part I, I argued that we have always counted the wrong things, and that AI has simply made the counting faster. I ended with five principles, a first table of human metrics, and an admission that I didn’t know which of them would survive contact with a real P&L. Two of them have already been withdrawn.

This part is less comfortable, because it assumes the argument gets won and then stops mattering.

Every counter-ledger I proposed in Part I is built on a premise: that human contribution to output is still worth protecting. That there is someone whose skills atrophy, whose judgement is worth paying for, whose cognitive space produces something a machine can’t: a creative director, yes, but equally a junior solicitor, a diagnostic radiographer, an underwriter, a reporter, a software engineer.

What happens to that premise when the machine can?

---

Act I: The timeline of human displacement

There have been three waves, and we have lived through one and a half.

The Mechanical Wave. Industrialisation displaced physical muscle. It reduced humans to extensions of machines, brutally and for a long time, but it left our cognitive dominance entirely intact. The bargain was implicit and it held: the machine does the lifting, we do the thinking. Every labour settlement of the twentieth century rests on that bargain.

The Cognitive Wave. This is where we are: AI is displacing analytical, administrative and intellectual labour, the exact category that the Mechanical Wave promised would be safe. The bargain is being renegotiated without anyone having called a meeting. Note what is happening at the bottom of the ladder in particular: the entry-level analytical work that used to be how people learned judgement is the first thing to be automated. We are removing the training ground while insisting that judgement is what humans are for.

The Creative Wave. This is the one nobody in my industry wants to say out loud, though the comfort blanket isn’t ours alone. Every profession keeps a version of it: clinical intuition, legal judgement, editorial instinct, engineering taste, the relationship a good banker has with a client. The word changes, but the argument doesn’t, and neither does its track record.

We comfort ourselves with a story: creativity is uniquely human, so creative work is safe. But if creativity is the brain’s synthesis of a lifetime of inputs (everything you’ve read, seen, loved, envied, misremembered, recombined), then it is a function of the size and diversity of your input set, plus the quality of the synthesis.

An artificial superintelligence with access to the totality of human output has an input set the size of the species: not one lifetime, but all of them.

I don’t think that makes human creativity worthless. But it does make the “creativity is our moat” argument a comfort blanket rather than a strategy. It is the same reasoning that said chess required intuition, then said Go required intuition, then said poetry required soul.

---

Act II: The last job

Everybody arguing about AI and work is arguing about which jobs survive. I think that’s the wrong question, because it assumes the surviving jobs will be recognisable versions of the jobs we have.

Here is the job I think survives longest, and it is already being written into law.

It is not creative. It is not analytical. It is accountable.

Read the regulation that actually exists rather than the regulation people imagine. The EU AI Act requires meaningful human oversight of high-risk systems: a person who can understand the output, monitor for automation bias, and intervene or override. Colorado gives people who receive an adverse automated decision the right, in some cases, to human review and reconsideration, which means somebody has to do the reviewing.^† Insurers pricing AI liability want a named individual in the chain. Every one of these instruments creates demand for the same thing.

Not a person who does the work, but a person who is answerable for it.

This is the one function whose entire value derives from not being a machine. You cannot sue a model, strike off an algorithm, subpoena a weights file, or put a training run in front of a select committee. Blame requires a legal person, and for now that means a human being with a name, an employer and a signature.

So while every other job is competing with the machine on capability, this one is protected by something no capability improvement can touch: the demand for someone to hold responsible.

What the job actually looks like

Picture the shift.

You arrive at nine and the queue is already built. It holds four thousand items, each one a decision the system has already made and now requires you to own: a claim, a diagnosis flag, a credit refusal, a piece of published content, a moderation call, a shortlist of eleven candidates from four thousand applicants.

Each item has a recommendation, a confidence score, and a green button.

You have, on average, a little over four seconds, not because anyone is cruel but because the throughput was costed on the assumption that oversight is a formality, and the assumption is correct, and everyone involved knows it is correct.

At 14:20 you approve item 2,310. Eight months later it is item 2,310 that matters, and the finding will note, accurately, that a qualified human reviewed and approved it.

You will not remember it. Your name is on it.

The cruellest job ever designed

Strip it back and look at the shape of the thing. Total liability. Zero authorship. Nothing you made, everything you carry.

Every previous grim job in history at least contained the dignity of causation: the miner dug the coal, the clerk wrote the ledger, the operator pulled the lever. This one contains none: it is the first job in which a person’s economic function is to be the place the consequences land.

And here is the part that makes it unstable rather than merely bleak: the oversight is fictional and the liability is real. Forty years of research on automation bias says that a human monitoring a system that is right most of the time stops monitoring. Vigilance decays; it isn’t a character flaw, it’s how attention works.

What’s newer, and worse, is that the standard fix fails too. Every oversight mandate rests on the same intuition: show the human the model’s reasoning and they will engage with it. The evidence points the other way: explanations mostly buy trust, not scrutiny.

• In a randomised study of 457 clinicians, accurate AI predictions raised diagnostic accuracy by 4.4 points. Systematically biased predictions cut it by 11.3 points, and showing the clinicians explanations that explicitly highlighted the model was keying off irrelevant features didn’t rescue them.¹
• 220 radiologists and physicians read chest X-rays with AI advice. When the advice was correct, accuracy hit 92.8%. When it was wrong, accuracy collapsed to 23.6%, and the more local, specific explanations made them trust it faster, cutting the time they spent considering it at all.²
• Radiologists reading mammograms alongside deliberately wrong “AI” suggestions: experienced readers, fifteen-plus years each, fell from 82% accuracy to 45.5%.³
• The foundational HCI finding, from Buçinca, Malaya and Gajos: people don’t analytically evaluate each recommendation, they form a heuristic about whether to trust the system. Adding explanations doesn’t reduce overreliance and may increase it. Only deliberate friction (forcing a pause, withholding the recommendation until the human commits) measurably helps, and participants disliked it.⁴

Ben Green’s review of the policy literature puts it bluntly: human oversight requirements are not supported by empirical evidence that humans can perform the function being asked of them. They serve instead to legitimise the deployment of systems that would not survive scrutiny on their own.⁵ In his and Yiling Chen’s experiment with 2,140 participants, an algorithmic risk score improved the accuracy of people’s predictions while degrading the accuracy of their decisions.⁶

So the reviewer becomes a rubber stamp, which everyone anticipated, and the legal function is undamaged, because the legal function was never to catch errors. It was to have somebody to name.

The role is a fiction the system requires to be legally real.

This is what Part I’s most boring metric was actually for

One number in the Human Ledger suddenly stops being an HR nicety and becomes the most consequential figure in the economy: judgement rate, the percentage of AI output a human materially changed. It now sits inside the standard’s Exposure condition, as the measure for the people who sign off.

If that number is healthy, oversight is real and the human is doing something. If it trends to zero while the liability stays at one hundred percent, then we have built an economy that runs on machine decisions and human blame, and we have no instrument that would tell us, because nobody currently measures it, and nobody is required to publish it.

We mandated the existence of the human in the loop. We never mandated any evidence that the human is doing anything. That gap is the whole argument for the Ledger, and it is a gap in live legislation right now.

The drafters even saw it coming. Article 14 of the EU AI Act explicitly instructs overseers “to remain aware of the possible tendency of automatically relying or over-relying on the output”. That is automation bias, named in the statute.⁷ The remedy for a forty-year cognitive finding is that the person should try to remember it’s happening to them.

And then it ends too

The accountable human is not a stable settlement but the last rung, and rungs get sawn off.

The job disappears the moment liability can be underwritten without a nameable person: model providers self-insuring their outputs, statutory liability caps for certified systems, mandatory insurance attaching to the system rather than the operator. All three are actively discussed, and none of them requires any advance in intelligence.

Which is the detail I find hardest to shake. The final human job does not end because a machine gets better at thinking. It ends because an insurance product gets better at absorbing risk.

That is where Part III of this series begins.

---

Act III: The ultimate barrier (the daily grind trap)

If all of that is true, and there will be no job left, the obvious question is: why not start now? Why not begin reweighting our lives toward relationships, community and craft today, ahead of the curve?

Because we can’t, for three structural reasons.

The cognitive and emotional drain. The modern grind doesn’t just take your hours, it takes the surplus that sits on top of them. People come home with nothing left to give to their families, their neighbours, or their own curiosity. You cannot build a relationship-based society out of people who are running at empty. Anyone who has tried to be a good parent at 8pm after a bad Tuesday knows exactly what I mean.

Structural hostage-taking. Housing, food and healthcare are held hostage by employment. “Reclaim the value of being” is trivially easy to say and impossible to act on when a mortgage payment is contingent on your continued economic productivity. This is a system problem, not a mindset problem.

The AI efficiency trap. This is the direct descendant of Part I’s argument.

We are being sold the idea that AI will give us our time back, and it won’t. Under a system whose non-negotiable KPI is eternal compounding growth, any hour that AI saves is immediately reallocated to more output, because that is what the ledger demands. The treadmill doesn’t stop. It speeds up, and it congratulates you for keeping pace.

We already have the receipts. Every productivity technology of the last forty years (email, mobile, cloud, collaboration software) was sold as time liberation and delivered as availability expansion. AI is the same trade at a steeper gradient.

The steelman: what if abundance actually arrives?

The strongest objection to everything above is that it treats scarcity as permanent, and serious people argue it isn’t.

Elon Musk’s position is that AI and robotics make work optional inside ten to twenty years, that this produces not a universal basic income but a universal high income, and that poverty ends through abundance rather than redistribution: “we’ll basically just issue money to people.” Mustafa Suleyman, running Microsoft AI, forecasts “radical abundance” from collapsing problem-solving costs. Mo Gawdat, who is far gloomier about the path, lands in a similar place at the end of it: roughly fifteen years of dislocation, then something better. These are not fringe voices, and I shouldn’t pretend they are.

And the strong version of their case dissolves an objection I would otherwise lean on. Cost disease, the reason healthcare, education, childcare and elder care get relatively more expensive forever, exists precisely because those sectors resisted automation. That is the premise now under direct attack. If robotics and AI really do deliver diagnosis, tuition, construction and care, then the protected sectors stop being protected and the cost curve breaks. Services become goods, and on a hundred-year view I think that’s more likely than not.

But notice what the abundance case is a claim about. It is a claim about productive capacity, not about distribution, and those have never been the same thing. Housing is already cheap to build relative to what it costs to buy; the price is land, permission and rent extraction, and no robot fixes any of those. Musk knows this, which is why his own conclusion requires a political act: somebody has to decide to issue the money. Abundance does not distribute itself. It never has once, anywhere.

So the two timelines don’t conflict, they nest. The hundred-year destination may well be abundance. The thirty-year path is a transition, and in every previous transition, the suffering happened entirely in the gap between when we could produce the thing and when we agreed to share it. Enclosure, industrialisation, containerisation: capacity arrived first, distribution arrived after a fight, and the people in between absorbed the difference.

The Human Ledger is not an argument against abundance; it is an instrument for the gap.

---

Act IV: The inevitable pivot (crisis and rebirth)

Crisis as the catalyst

Structural change of this magnitude has never been planned; it has been forced. The weekend, the welfare state, the eight-hour day: none of them were adopted because someone made a compelling case. They were adopted after a crisis made the alternative untenable.

The same will be true here. The pivot from doing to being won’t arrive as policy but as a break, an employment crisis severe enough that the link between work and survival becomes politically impossible to maintain.

This is the least satisfying thing I have to say, and I think it’s the truest.

But the ledger is how you shorten the crisis

Here is why Part I still matters, even if the crisis in this act is inevitable.

You cannot govern what you cannot see. Every previous transition was made longer and more brutal by the fact that we had no instruments, no way to observe the damage until it was historical. We measured output beautifully and human cost not at all.

A Human Ledger doesn’t prevent the crisis, but it shortens it, and that is the difference between a controlled descent and a stall.

And this is where I think the trajectory is predictable, because we have watched it happen once already. Sustainability reporting went voluntary → reputational → mandatory → priced into capital, in about twenty years. There is no reason to assume the Human Ledger follows a different path. My prediction, in order:

1. Voluntary. A handful of companies publish human metrics because it’s a differentiator in a talent market.
2. Reputational. Candidates start asking for the numbers in interviews, the way they now ask about D&I and carbon. The first time a scarce senior hire (a creative director, a consultant radiologist, a lead engineer) turns down an offer because a company won’t disclose its judgement rate, the market notices.
3. Mandatory. Governments already require AI job-displacement disclosure in a handful of jurisdictions. Disclosure is the thin end. A Human Ledger threshold, enforced through tax treatment, is the thick end: companies below the line pay more.
4. Priced in. Capital reprices firms that hollow out their human capability, because hollowed-out firms turn out to be fragile ones.

Step two is the interesting one, and it’s closer than it sounds. It doesn’t need legislation. It needs about four hundred people in one industry to start asking the same question in interviews.

What we do in the gap

So the market will keep pushing toward approval rates and liability pricing, because that is what the market is for. Governments will keep regulating the output rather than the conditions. Not because workers don’t vote, but because they don’t vote as workers.

Every voter meets AI’s output: the deepfake, the refused loan, the hallucinated citation, the chatbot that got the diagnosis wrong. A bad output can happen to anyone, so outputs get labelling rules, transparency duties and rights of appeal. That is where every AI law in Act II points.

The conditions behind the output (the hours, the vanished junior roles, the four seconds per decision) are seen only by the people inside them. Those people are scattered across employers, contracts and sectors, with no shared identity and no single lobby. Harm to the public makes headlines. Harm to the workforce makes attrition statistics.

The Ledger arrives when the disruption is large enough to make the current arrangement politically impossible, and not one day earlier.

That leaves exactly one useful thing to do in the meantime, which is to have the instrument built, argued over and legitimate before it’s needed. Every previous transition was made longer and crueller by the fact that we had no way to observe the damage until it was already history. We measured output beautifully and human cost not at all.

The Ledger doesn’t prevent the crisis, but it can make the difference between a controlled descent and a stall.

---

Closing

I don’t have the metrics right, and I’m reasonably sure half of them are wrong.

But the alternative to a bad first draft is another decade of counting only what’s easy, which is how we got here. So I’m putting it somewhere public, versioned, and open to being argued with.

The five principles, the six conditions and the four arguments I can’t finish alone are here at humanledger.report, as a working spec rather than a manifesto. The debate runs in the Human Ledger LinkedIn group. If you want to kill one of them, that’s the place to do it.

One last thing, and it’s the question I can’t fit inside thirty years.

If the accountable human is the last job, and if it ends not because machines got smarter but because insurance got better, then at some point there is no economic answer to the question of what a person is for. Everything in this series so far has been an argument about how to protect human contribution to output. That argument has an expiry date.

What’s left when it expires is a hundred-year question, and it’s the one Kennedy was actually asking in Kansas. I’ll take it up in Part III: The De-materialisation of Meaning.