AI Is Quietly Creating a New Class System
Two groups are forming: the people who direct AI, and the people it directs.
Sahir Maharaj12 min read
Every major technological shift in the history of industrial society has reorganised the class structure. The Industrial Revolution created the factory owner class and the factory worker class, and the tensions between them defined much of the social and political history of the nineteenth and twentieth centuries. The information technology revolution created a professional class of knowledge workers and an underclass of gig and service workers whose labour was devalued relative to credential-heavy cognitive work. We are in the early stages of a third reorganisation, and the class structure it is producing is coming into focus quickly enough that we can describe its outlines. The organising principle is not ownership of capital or possession of credentials but something more specific: the relationship of the person to the AI systems that are increasingly doing the substantive work of the economy.
The clearest version of the emerging divide can be stated simply. There are people who direct AI systems, who define the problems those systems address, who evaluate and interpret their outputs, who make the consequential decisions those outputs inform. And there are people who are managed by AI systems, whose work is monitored, evaluated, and assigned by algorithms, whose pace is set by a queue rather than a judgment, and whose value to the employer is increasingly determined by how efficiently they execute tasks the AI has structured. The first group is growing more powerful and more economically valuable. The second is growing in size and experiencing pressure on wages and conditions. The gap between them is widening.
The people who are best positioned in the emerging AI class structure share a specific profile. They have domain expertise deep enough to evaluate AI outputs in a specific field with enough understanding to know when the output is useful and when it is wrong. They have the ability to formulate the right problems, which is a different skill from solving them: AI is increasingly capable of solving well-specified problems, which means the scarce and valuable skill is specifying the problem correctly. And they have the human and institutional skills to translate AI-generated analysis into action in organisational contexts, which requires the kinds of relational intelligence, communication capacity, and political judgment algorithmic systems do not possess.
The formation of the directing class is not simply a matter of education, although education matters. It is a matter of experience accumulation in roles where the skills of problem specification and output evaluation are actually practised. This creates a specific trajectory problem: the entry-level roles that have historically provided the apprenticeship experience through which junior professionals developed the judgment to eventually become senior ones are precisely the roles that are being most rapidly automated. The junior lawyer who reviewed contracts learned what mattered in a contract by doing the review, accumulating judgment that could eventually become the basis for higher-level advisory work. If the review is done by AI, the learning it produced does not happen, and the pipeline of senior professionals who developed through that apprenticeship is not refilled from below.
The people who are most clearly in the directed category are those whose work is most completely specified and monitored by AI systems: gig workers whose task assignment, routing, and performance evaluation are entirely algorithmic; content moderators whose queues are AI-generated and whose throughput is AI-tracked; warehouse workers whose movement, pace, and error rate are monitored in real time by systems that set the parameters of acceptable performance; customer service workers whose scripts are AI-generated, whose calls are AI-transcribed, and whose performance is AI-scored. In all of these cases, the human worker is executing within a structure that the AI has defined, at a pace the AI has set, to a standard the AI is evaluating. The direction flows from the algorithm to the person, not the other way around.
What is less visible but equally significant is the expansion of the directed category into roles that were previously much more autonomous. The middle manager whose decisions are increasingly framed by AI-generated recommendations. The financial analyst whose work is primarily the interpretation of AI-generated analysis. The doctor whose clinical decisions are structured by AI decision-support tools whose recommendations carry implicit institutional endorsement. Professionals who previously exercised substantial independent judgment are finding that judgment increasingly structured and constrained by algorithmic systems, moving them along the spectrum without any explicit decision to reorganise the role.
The historical pattern of technological class formation is not encouraging. The benefits of major technological transitions have consistently accrued disproportionately to capital owners and to the workers whose skills are most complementary to the new technology, while the costs have been concentrated among workers whose skills are displaced or devalued. The periods of transition have been extended enough that the workers who bore the costs of the displacement often did not live to see the new equilibrium in which, on aggregate, living standards eventually improved. The policy responses that ameliorated the worst outcomes, strong labour protections, redistributive taxation, investment in public education, were politically contested and achieved only through sustained collective action over decades.
The AI transition is happening faster than previous technological transitions, which compresses the timeline in which policy responses need to be developed and implemented. The risk is not that AI creates a permanent new underclass from which there is no exit, though that risk exists. The risk is that the transition period, during which the distributional consequences are severe and the long-term equilibrium has not been established, is long and painful for a large fraction of the people living through it.
What would change the trajectory requires being honest about the fact that it requires political choices. Investing in education that develops problem-specification skills rather than just execution skills requires reforming educational systems oriented toward producing competent executors. Expanding access to capable AI tools rather than allowing them to concentrate in the most well-resourced organisations requires either regulatory mandates or public provision of AI infrastructure. Addressing wage pressure on directed workers requires labour protections that extend into contexts where algorithmic management has been used to sidestep traditional employment relationships. None of these is technically radical. All of them are politically difficult. And the AI class system will continue to consolidate in their absence.
You might also like
View all
Are We Giving AI Too Much Control?
Control does not transfer in one moment. It seeps out one small decision at a time.
The Internet Is Being Flooded With AI Slop
Near-zero production costs, and what happens to the signal beneath the flood.