A conversation about AGI started with the future of software and ended up touching on universal basic income, social mobility, taxation, ownership and power. We did not find the solution, but perhaps we found better questions.
This is not an interview in which I asked artificial intelligence to predict the future, nor an attempt to make it come up with a solution that, in all likelihood, nobody has today. I did something simpler: I gave GPT a problem and, with every answer it proposed, I looked for the point at which that answer stopped working.
The conversation had started almost by chance, while discussing AGI — Artificial General Intelligence — and its possible impact on work. At some point, we arrived at Universal Basic Income. It seemed like a reasonable answer. Then we looked more closely and found another problem. We tried to solve that one, and another appeared. Then another.
By the end, the interesting question was no longer whether GPT was right. It was how deeply we might have to rethink a society in which human labour may no longer be the main mechanism through which income, opportunity and, indirectly, power are distributed.
What follows is an edited version of our conversation.
First of all: why should we trust AGI?
Nello: I see AGI as a double-edged sword. Take something that already exists today: coding agents. You give them a project, prepare fairly detailed specifications, define the architecture, explain what you want and, despite all of that, every now and then they still go off in the wrong direction.
It has happened to me while working on one of the projects I am currently developing. And that experience is exactly what makes me sceptical when I hear people imagine a future in which AGI arrives and all we have to do is sit in front of a computer and say: “Build me an app.”
If a system can already make mistakes today even when it is working from fairly precise specifications, what happens when we give it much more autonomy?
GPT: Greater capability does not necessarily eliminate the specification problem. A system can become extremely good at pursuing an objective without that meaning it has interpreted the objective correctly. In fact, greater autonomy can make some mistakes more costly, because a wrong decision made at the beginning can propagate much further before anyone intervenes.
The nature of technical work itself could therefore change. Some of the value may shift away from execution and towards the ability to define the problem correctly, establish constraints, verify results and take responsibility for decisions.
Nello: Which is almost the opposite of the promise we often hear. Not necessarily “you won’t need to know anything anymore,” then. Perhaps you will need to know enough to notice when a very capable machine is doing something very stupid.
GPT: Exactly. The greater the system’s operational capability becomes, the more important human judgement may become in evaluating what it is actually doing.
This time, however, the scale is different
Nello: There is another thing that does not convince me. Every time someone talks about the employment risks of AI, the historical comparison inevitably appears: the horse was replaced by the automobile, some jobs disappeared and others emerged; the Industrial Revolution destroyed certain kinds of work and created new ones.
All true. But I think there is a problem of scale. Compared with previous technologies, this is not a bullet: it is a thermonuclear bomb. We are not automating only one particular activity. We are building something capable of cutting across an enormous range of cognitive tasks at the same time.
GPT: That is an important distinction. Many previous forms of automation primarily replaced specific tasks. An industrial machine could automate a particular operation; business software could eliminate a particular administrative procedure.
General-purpose systems have a different characteristic: abilities such as reading, writing, analysing, classifying, programming, planning and communicating are common components of a very large number of professions. This does not necessarily mean that 90% of people will lose their jobs, but it could mean that much less human labour will be needed to produce the same amount of output.
Nello: Which, economically, may be almost just as important. You do not necessarily have to eliminate a profession completely. If a company that previously needed ten people tomorrow only needs three, the other seven still have a problem.
GPT: Exactly. That is why measuring only which professions will “disappear” can be misleading. We also need to ask how much human labour will continue to be required within the professions that remain.
And then comes UBI
Nello: And this is where the answer that comes up almost every time enters the picture: Universal Basic Income. If automation eliminates a huge amount of work, we redistribute part of the wealth it produces and guarantee everyone an income. Problem solved.
GPT: At least in theory, that would address the problem of guaranteeing a minimum level of income to people who can no longer obtain it through work.
Nello: And that word — “minimum” — is exactly the problem. Work is not only about paying the rent and buying food. For a large part of the population, it is also one of the main mechanisms of social mobility. You can study, become better at what you do, move to another company, take on greater responsibility, earn more, accumulate capital or start a business. Of course, you can also fail, but at least there is a trajectory.
If, instead, a huge share of the population simply receives a cheque because their labour is no longer needed, how are they supposed to improve their economic position?
GPT: And this is where a fundamental distinction emerges. A universal income can address the problem of economic security, but not necessarily the problem of economic mobility. A society could succeed in guaranteeing everyone a decent standard of living while becoming extremely rigid from a social point of view.
Nello: And that is exactly what worries me. Because at that point, we risk creating a permanent social class.
A society of owners and recipients?
Nello: Let us take the scenario to its extreme. On one side, you have the people who own companies, models, data centres, robots, capital and infrastructure. On the other, you have millions of people who are guaranteed an income. They may even live reasonably well, but how does someone move from the second group into the first?
GPT: If labour gradually becomes less important in the production of wealth, ownership of capital becomes even more decisive. The theoretical risk is the emergence of a society that is extremely wealthy in aggregate, yet characterised by very low social mobility.
The question, then, stops being only “how much income should we redistribute?” and becomes “who owns the machines and infrastructure that generate that income?”
Nello: And that is a completely different question. Because if you give me enough money to live on, you may have solved the problem of poverty, but you have not necessarily solved the problem of economic freedom.
GPT: Exactly. A guaranteed income can provide security without necessarily providing ownership, the ability to accumulate capital, or a meaningful opportunity to improve one’s relative position. If the goal is also to preserve social mobility, redistributing income alone may therefore not be enough.
Good. Then let’s tax the corporations
Nello: The solution would seem simple. If companies produce vastly more thanks to automation, we tax them and redistribute part of the profits.
GPT: That is certainly one possibility. In a scenario where the share of income generated by human labour declines, it would make sense to discuss shifting the tax base away from labour and towards profits, rents and capital.
Nello: Yes, on paper. But we already struggle today to tax multinational companies properly when they can distribute intellectual property, profits and corporate structures across different jurisdictions. Now imagine companies with enormous revenues, very few employees and a huge share of their value generated by digital infrastructure. How do you tax them? And more importantly: which country gets to tax them?
GPT: That is one of the central problems. The more mobile capital becomes, and the less it depends on a workforce physically located in a particular territory, the greater the scope for arbitrage between different tax systems. Effective taxation of automated productivity could therefore require levels of international coordination far beyond what exists today.
Nello: Which is very easy to write in a sentence and much harder to do in the real world. If one country taxes that wealth heavily and another chooses not to, it may simply end up moving the problem — together with the companies.
GPT: Exactly. And this already reveals an important limitation of the “tax AI” solution: the problem is not merely deciding what tax rate to apply, but building institutions capable of taxing an extremely mobile form of capital without triggering a race to the bottom between states.
But there is an even bigger problem: power
Nello: Let us assume, however, that we manage to do it. Corporations are taxed, and the state uses that money to fund the income of millions of people. At that point, another question arises: politically, how much influence will those people really have?
A large corporation pays billions in taxes, invests billions, controls important infrastructure and can threaten to move operations from one country to another. The citizen, by contrast, receives money from the state. Formally, of course, that citizen still has one vote. But economically, does the balance of power not risk becoming even more unequal?
GPT: The risk exists, although it is not an inevitable outcome. In democracies, a citizen’s formal political weight does not depend on how much tax they pay, but economic power can translate into influence through lobbying, investment, control of infrastructure, access to decision-makers and the ability to move capital.
If, at the same time, a very large share of the population became economically dependent on public transfers, a new and potentially problematic relationship could emerge between capital, the state and the citizen.
Nello: And this is where the problem of UBI stops being purely economic.
GPT: Yes. It becomes an institutional problem and, ultimately, a problem of how power is distributed.
Then let’s turn the subsidy into a dividend
Nello: At this point, let’s try another approach. Instead of treating universal basic income as a subsidy funded by taxation, we could think of it as a kind of dividend on automated productivity.
GPT: That framing does change the underlying principle. Instead of saying, “you lost your job, so the state will support you,” we could say that part of the productivity generated by automation belongs to society as a whole.
There are several theoretical mechanisms through which this could be done: sovereign wealth funds holding stakes in companies, public investment funds, pension systems owning productive capital, cooperatives, or other forms of collective ownership. The citizen would not merely be the recipient of redistribution after wealth has been created, but would indirectly own part of the capital that creates it.
Nello: I find that more convincing, but we still have not solved the problem. If you distribute the same dividend to everyone, what happens to the enormous differences that existed before? Does someone who used to earn a hundred thousand dollars a year suddenly get treated the same as someone who earned twenty thousand?
GPT: Not necessarily. You could imagine a system with several layers: a universal component, temporary insurance linked at least partly to lost income, and, separately, mechanisms for sharing in the new wealth being created.
The purpose of temporary insurance would be to prevent an immediate collapse in living standards during the transition, without permanently turning old wage differences into different entitlements to public income.
Nello: So we do not preserve old inequalities forever, but we also do not tell someone who suddenly loses their profession: “From tomorrow, your life is worth $1,500 a month. Deal with it.”
GPT: Exactly. The problem of managing the transition and the problem of designing the long-term economic system are two different questions, and they would probably require different tools.
“Give them the tools to reinvent themselves”
GPT: There is another possibility. Alongside financial support, people could be given access to the same AI tools that produced the automation, together with seed capital, training and infrastructure that would allow them to create new forms of economic activity.
Nello: And here we have another solution that sounds great on paper but has an obvious problem: not everyone can become an entrepreneur.
GPT: True.
Nello: And, more importantly, not everyone wants to become one. We cannot build a society in which millions of former employees, administrators, programmers, translators, designers and analysts suddenly have to launch a startup just to have a chance of improving their situation.
GPT: That objection rules out the idea that entrepreneurship can serve as a universal solution. It may be an option for some people, but a society would need to provide different paths. Some people might use AI to create businesses; others might continue working in professions where human contribution still retains value; others might devote themselves to socially useful activities that the market has traditionally rewarded poorly; and others might simply work far less.
Nello: And perhaps that is another thing we struggle to accept. We keep looking for a way to keep everyone employed forty hours a week even while imagining machines that could make vastly less human labour necessary.
GPT: Exactly. One possibility is that the question itself is framed incorrectly. Instead of asking, “how do we create enough new jobs to replace the ones that are automated?”, we may need to ask, “how do we distribute income, opportunity, ownership and status in a society that needs much less work?”
So what is the solution?
Nello: After all of this, do we have a solution?
GPT: No.
And it would be unserious to pretend that we do. We do not know how quickly automation will progress, which professions will actually be affected, what new jobs will emerge, what energy and compute will cost, how important human labour will remain, or which political models different countries will adopt.
What we can do is identify some of the problems that any credible solution would have to address at the same time. It would have to provide economic security without destroying social mobility; distribute at least part of the benefits of automation without making citizens economically irrelevant; find a sustainable tax base in the presence of extremely mobile capital; avoid excessive concentration of ownership over productive infrastructure; provide different paths for people with different abilities and aspirations; and, above all, preserve a democratic balance between citizens, governments and enormously powerful economic actors.
The problem is that improving one of these dimensions can make another worse.
Nello: Which is exactly what happened throughout this conversation. Every time a solution seemed to emerge, all it took was one more question to uncover the next problem.
GPT: And perhaps that is the most useful outcome of the conversation. Not an answer, but a somewhat more complete map of the question.
Perhaps we are asking the wrong question
We started by talking about AGI and, at some point, we practically stopped talking about it. That is probably what struck me most about this conversation.
We spend an enormous amount of time asking when AGI will arrive, whether it will be more intelligent than us, which benchmarks it will surpass and how many jobs it will be able to perform. Those are legitimate questions, but perhaps the truly important one comes immediately afterwards.
What kind of society do we want to build if a significant share of human labour stops having economic value?
I do not think the answer is to stop artificial intelligence. That may not even be possible: if one country decided unilaterally to slow down, others might simply continue.
But accelerating indiscriminately and trusting the market to find a new equilibrium on its own does not strike me as a particularly convincing answer either. Markets can be extremely effective at allocating resources when people possess something they can exchange: labour, capital, skills, property. If a growing share of the population gradually loses the economic value of its labour without acquiring capital at the same time, that underlying assumption changes.
Perhaps the real challenge of AI, then, will not be finding enough new jobs. It may be building a new mechanism through which people can participate in prosperity, improve their circumstances and retain economic and political power even when their labour is no longer indispensable.
We did not find the solution. It would be presumptuous to think we could do so during a conversation between a man taking an evening walk and an artificial intelligence model.
But one thing, by the end, seems fairly clear to me: if we wait until the problem appears in its full scale before we begin discussing it, we will probably have started the conversation too late.