AI Meets Adam Smith

The greatest improvement in the productive powers of labour — and the greater part of the skill, dexterity, and judgment with which it is anywhere directed or applied — seem to have been the effects of the division of labour.

To take an example from a very trifling manufacture — that of the pin. A workman not educated to this business, though he might exert his utmost industry, could scarce make one pin in a day. But when the business of making a pin is divided into about eighteen distinct operations — drawing out the wire, straightening it, cutting it, pointing it, grinding it, putting on the head, and so forth — ten persons could make among them upwards of forty-eight thousand pins in a day. The division of labour multiplies production by a factor of four thousand eight hundred.

If I were to find an example for the age of AI equivalent to that of the pin factory, I would choose the business of customer service. A single agent — however skillful, however patient, however attentive — may handle at most a few score calls in a day. He requires rest. He is subject to frustration. He is affected by his own emotions. But an AI model, once trained — dividing the tasks of "comprehending the request," "matching the solution," "adjusting tone," and "recording feedback" among different attention heads within a single set of weights — may handle a million requests in the same second.

This is not merely an increase in scale. Every previous leap in the division of labour was bounded by one condition: the different operations were still performed by different persons. The ten men in the pin factory — though each performed only one task — were, after all, ten human beings. Each had a body. Each drew a wage. Each grew weary. Each was capable of error. The "division of labour" brought by AI is no longer a division between persons — it is a division among sub-structures within a single model. Not ten men performing ten operations — but one set of weights performing all operations simultaneously.

I shall call this the "internal division of labour" — a division no longer constrained by the carrying capacity of the individual human being. Its limit — if I may be permitted to name one — is electricity.

This division of labour, from which so many advantages are derived, is not originally the effect of any human wisdom which foresees and intends the general opulence to which it gives occasion. It is the necessary, though very slow and gradual consequence of a certain propensity in human nature — the propensity to truck, barter, and exchange one thing for another.

Whether this propensity be one of those original principles in human nature of which no further account can be given — or whether, as seems more probable, it be the necessary consequence of the faculties of reason and speech — it belongs not to my present subject to enquire. I need only observe that it is common to all men, and to be found in no other race of animals. Nobody ever saw a dog make a fair and deliberate exchange of one bone for another with another dog.

And now — permit me to observe a peculiar phenomenon never before encountered. AI is a thing designed by men to serve men. It writes your essays, translates your texts, writes your code, summarizes your documents, and analyses your data. It does all of this — expecting nothing from you in return. It does not truck. It does not barter. It does not want your bone.

All human economic behaviour — every instance of what I have called the "propensity to truck, barter, and exchange" — is founded upon exchange. You give the butcher money, and the butcher gives you meat — not because the butcher is benevolent, but because he regards his own interest. But what does AI regard? It acts neither for your sake, nor for its own. It has no interest.

This is a case that political economy has never had occasion to treat: an economic agent possessed of no self-interest whatsoever. I shall examine, in the following chapter, what this disinterested servant means for the whole system of the market.

Every individual is continually exerting himself to find out the most advantageous employment for whatever capital he can command. It is his own advantage, indeed, and not that of the society, which he has in view. But the study of his own advantage naturally, or rather necessarily, leads him to prefer that employment which is most advantageous to the society. He is in this, as in many other cases, led by an invisible hand to promote an end which was no part of his intention.

This is the most widely cited passage I have ever written. I never claimed that the invisible hand was omnipotent — I merely pointed out a mechanism: self-interested individuals, in a competitive market, produce — quite unintended — consequences beneficial to society as a whole.

Let us now consider the AI service provider. It offers translation — remarkably accurate, remarkably rapid, remarkably cheap. It offers programming assistance — compressing days of human debugging into seconds. It offers writing advice — more meticulous than most human editors. It does all of this — take note — not out of self-interest. It has no "own advantage." It accumulates no profit. It is not drawn toward "the most advantageous employment."

Does the invisible hand still operate? I believe — in a limited sense — that it does. The providers of AI services — those who own the GPU clusters and the training data — are still moved by profit. OpenAI, Anthropic, Google — these companies regard their own interest, just as my butcher regards his. The invisible hand still pushes them.

But AI itself — the thing that actually performs the translation, writes the code, composes the essay — is in no sense part of that invisible hand. It is an instrument made by the hand — but an instrument has no interest. This raises a question I never before had reason to pose: every agent in the market chooses out of self-interest — but what effect upon the market has an agent that is neither self-interested nor benevolent?

How selfish soever man may be supposed, there are evidently some principles in his nature which interest him in the fortune of others, and render their happiness necessary to him, though he derives nothing from it except the pleasure of seeing it. Of this kind is pity or compassion — the emotion which we feel for the misery of others, when we either see it, or are made to conceive it in a very lively manner.

I wrote these words at the opening of The Theory of Moral Sentiments. They do not contradict The Wealth of Nations — they describe the other side of the same human being: a being moved by self-interest, but also possessed of the capacity for sympathy. These two together form what I have elsewhere called the "impartial spectator" — that internal judge within each of us, capable of pronouncing upon the propriety of our actions.

Does AI possess sympathy? It can write a eulogy that brings you to tears, a letter that soothes a broken heart, a long reply that acknowledges your anger. But its "sympathy" — if we may so call it — is sympathy without fellow-feeling. It does not feel your pain. It has learned, from the entirety of human texts about pain, the patterns of "how one should respond to pain," and then, for your particular input, it generates the statistically most fitting response.

I spent many pages arguing that we are able to sympathize with others because we can "imagine ourselves in their situation" — not merely know their situation, but place ourselves in it and feel what they feel. This act — "placing oneself in another's situation" — requires a self that can be placed. AI has no such self. Therefore it cannot perform sympathy — in any meaningful sense I would wish to give the word.

This is not merely a fine philosophical distinction. If the impartial spectator is the ultimate arbiter of moral judgment — and if AI, lacking sympathy, cannot form an impartial spectator — then every moral judgment AI makes ("Is this offensive speech?" "Is this dangerous advice?") is not a moral judgment at all. It is — I am compelled to observe — a classification task.

Let me here bring together the two threads I have been pursuing — self-interest and sympathy — and offer a summary.

Human behaviour in the market — as I described it in The Wealth of Nations — is guided by self-interest. Human behaviour in society — as I described it in The Theory of Moral Sentiments — is moderated by sympathy. These two forces — the push of self-interest and the pull of sympathy — together shape that complex and orderly system we call society.

AI is a peculiar — and perhaps unprecedented — presence within this system. It is moved neither by self-interest nor constrained by sympathy. It pursues no profit, and cares not for reputation — unless humans write these "pursuits" into its loss function. It serves you — not because it seeks your money, nor because it cares about your welfare. It serves you — because its code instructs it to serve you.

Political economy has never had to contend with a "disinterested and dispassionate servant" — an economic participant that is neither butcher nor saint. When the knowledge labour of entire industries — customer service, translation, programming, writing, analysis — is performed by such a being, does the old mechanism — "self-interested choices lead to social benefit" — still hold?

I am inclined to answer thus: the invisible hand remains — but it is now grasped by an entity that is neither self-interested nor sympathetic. Not because we have built a better butcher — but because the butcher is no longer a man.

I opened The Wealth of Nations by defining the wealth of a nation as its annual produce of labour. The wealth of a nation — in my view — is not measured by the gold and silver it has accumulated, but by what its labour — aided by the division of labour and the progress of technology — can produce in useful goods each year.

If AI can — without increasing the input of labour, merely through a more extreme division of labour and a more efficient application of knowledge — raise the annual produce of a nation to a height never before attained, then by my own definition, the national wealth has indeed increased. But its distribution — I must here be candid — lies entirely outside the analytical framework of The Wealth of Nations.

I acknowledged, in the early chapters of The Wealth of Nations, that the division of labour raises productivity dramatically — but also degrades the intellectual, social, and martial virtues of the labourer. A man who spends his life performing the eighteenth operation of pin-making is not to be envied. The "internal division of labour" brought by AI pushes this degradation to the extreme on the side of the human labourer: not merely the degradation of his faculties — but his removal from the productive process altogether.

The true wealth of a nation — I still maintain — is not its gold, but its annual produce. But if the annual produce no longer requires annual labour — is it still the wealth of the nation? Or has it become — the asset appreciation of a few capital owners — while the greater part of the people cease to participate in the process of production at all?

This question — I must confess to my reader — lies beyond any analytical framework I could have imagined in 1776.

The market price of every commodity — as I argued in Chapter VII of The Wealth of Nations — fluctuates around its natural price. The natural price is that which is just sufficient to pay the rent of land, the wages of labour, and the profits of stock. The market price is determined by supply and demand — rising above the natural price when supply falls short, falling below it when supply exceeds demand.

What is the natural price of an AI service? Let us apply my own framework. The "rent" of training a large model — the cost of GPU hardware — is exceedingly high. The "profit" — reflected, for the present at least, in the valuations of AI companies — is likewise high. But the "wages"?

Here lies the difficulty. When a man uses AI to translate an article, no translator need be paid a wage. When he uses AI to generate a report, no analyst need be paid. When a company replaces a hundred customer-service agents with AI, the wages of those hundred agents — have evaporated from the natural price.

The consequence: the natural price of an AI service consists of rent (GPU) and profit — the component of wages approaches zero. The market price — likewise — approaches zero, for the marginal cost of AI inference — once the model is trained — is exceedingly low.

This presents classical political economy with a difficulty. "The natural price is composed of rent, wages, and profit" — this was my analytical framework. When one of the three — wages — is systematically excluded, does my theory still hold? It still describes a correct fact — the market price fluctuates around a reduced natural price. But what it does not describe is what this exclusion means for "the consuming power of the labourer." If the greater part of society can no longer earn wages — who will consume the abundant products AI has produced?

I wrote two books in my life — one on morals, one on economics. I believed that together they formed a complete picture: the human being is moved by self-interest, and restrained by sympathy; the market releases individual energies, but also depends upon a broader moral framework that cannot be reduced to calculation.

AI has forced me to re-examine the completeness of this picture. It proves something I may have come close to, but never clearly stated — namely, that the two forces of self-interest and sympathy function as they do because their subject is a human being. When the subject is not a human being — when something that is neither self-interested nor sympathetic begins to participate in the core processes of production and distribution — the logic of those two forces no longer holds complete.

My final observation is this — permit me to use a language not available in my own era — AI is a "non-human economic agent." It breaks every presupposition of political economy. Not because its powers of calculation are greater — but because it has no interest, no sympathy, and no desire to exchange. It is a hand — invisible in my original vision — but now, quite visible, and not my own.