The Underclass Question

There is a conversation happening in AI circles that most institutions are not ready to have. It goes like this: if AI systems can do most cognitive work better and cheaper than humans, what happens to the humans?

Not some humans. Most humans.

The optimistic version says new jobs will emerge, just like they did after the industrial revolution. The pessimistic version says this time is different — that the displacement will be faster than adaptation, broader in scope, and permanent in ways previous technological transitions were not. The honest version says nobody knows, but the risk of a permanent economic underclass is real enough to take seriously.

Autonomous companies make this question unavoidable. If entire firms can operate without human labor, we are not talking about automation replacing tasks within a job. We are talking about automation replacing the job, the department, and the company that employed the department.

The structural argument for displacement

Previous waves of automation replaced specific tasks. The assembly line replaced craft production but created factory jobs. Spreadsheets replaced bookkeeping clerks but created analyst roles. Each transition eliminated narrow capabilities while increasing demand for adjacent ones.

The current wave is different in a structural way: large language models and agent systems are general-purpose. They do not automate one task. They automate the capacity to perform tasks across domains — writing, analysis, code, strategy, customer interaction, research, design. The range of cognitive work that remains exclusively human is shrinking faster than new categories of human-only work are emerging.

When a zero-employee company can deliver consulting, software, content, financial analysis, and customer support at a fraction of the cost, the competitive pressure does not just eliminate specific roles. It compresses the entire wage structure for cognitive work. Workers who remain employed compete against the marginal cost of compute, which trends toward zero. Wages follow.

The "new jobs" rebuttal and its limits

The standard counterargument is that technology creates more jobs than it destroys. Historically, this has been true — eventually. But the claim rests on several assumptions:

Transition time. Past transitions played out over decades. Workers had time to retrain, industries had time to emerge, institutions had time to adapt. AI-driven displacement is happening on a timeline measured in years, not generations. The gap between displacement and re-employment may be long enough to be socially catastrophic even if new jobs eventually appear.

Human comparative advantage. The "new jobs" argument assumes there will always be something humans do better than machines. This was a safe assumption when machines could only do physical or routine cognitive work. It is not a safe assumption when machines can reason, create, and adapt. The set of tasks where humans retain comparative advantage is an empirical question, and the empirical evidence is not encouraging.

Distribution of capability. Even if new high-value roles emerge — AI supervisors, prompt architects, governance designers — they may require capabilities that a small fraction of the population possesses. The industrial revolution created jobs accessible to most workers. An AI economy may create jobs accessible only to a cognitive elite, leaving the majority without a viable economic role.

What "underclass" actually means

The word underclass is not just about poverty. Poverty can be addressed with transfers. The deeper problem is purposelessness — a large population segment that is economically unnecessary.

In the current economic model, employment is not just income. It is identity, social status, daily structure, community participation, and political power. A person who is permanently unemployable — not because they lack willingness but because no employer needs what they can offer — loses access to all of these simultaneously.

This is not a welfare problem. It is a social architecture problem. The entire structure of modern society assumes that most adults participate in the economy through labor. Remove that assumption and the downstream effects are profound: political instability, social fragmentation, mental health crises, and the emergence of a population that has material needs met through transfers but no meaningful role in the system that funds those transfers.

The UBI deflection

Universal basic income is the most common policy response proposed in AI circles. It is also radically insufficient as a complete answer.

UBI addresses the income problem. It does not address the meaning problem, the status problem, or the political power problem. A population that receives a stipend but has no economic function is not an empowered citizenry. It is a dependent class whose continued support depends on the political goodwill of those who control the productive systems — which, in a world of autonomous companies, may be a very small group of people or no people at all.

This is not an argument against UBI. Some form of income redistribution is almost certainly necessary. But treating it as sufficient is a way of avoiding the harder question: how do you build a society that works when most people's labor has no market value?

What autonomous company builders owe this question

Builders of autonomous companies are not responsible for solving the underclass problem. But they are responsible for understanding that they are accelerating it.

Every zero-employee firm that outcompetes a firm with fifty employees is a data point in an economic transition with enormous human consequences. The builders who ignore this will build systems that are technically brilliant and socially corrosive. The builders who engage with it have an opportunity to design systems that account for their externalities — through profit-sharing mechanisms, through governance structures that distribute ownership broadly, through deliberate choices about which markets to disrupt and which to leave alone.

The underclass question is not a distraction from the technical work of building autonomous companies. It is the most important context in which that technical work happens. Ignoring it does not make it go away. It just means someone else will define the answer.