The conversation about artificial intelligence has been stuck in a familiar loop: first comes the prediction that machines will replace workers, then the debate over whether the replacement will be fast or slow, and finally the question of whether new jobs will appear quickly enough to offset the losses. But the discussion is shifting again—quietly at first, and now with more urgency—toward something less cinematic and more consequential: what happens after the productivity shock, when governments discover that their existing policy toolkits were designed for an economy where value was created primarily by people and taxed through familiar channels.
In that sense, the real question isn’t only whether mass underemployment arrives. It’s whether the policy framework arrives with it.
Underemployment is not the same as unemployment. It can look like people who still have jobs but fewer hours, lower bargaining power, reduced wages, and shrinking career ladders. It can also show up as “employment without leverage”: workers performing tasks that are increasingly standardized, monitored, and optimized by AI systems, while the gains from efficiency accrue elsewhere. When automation reduces the demand for certain roles, the labor market doesn’t always snap cleanly into job loss and re-hiring. More often, it compresses work—turning full-time roles into part-time arrangements, pushing workers into lower-paid categories, and making transitions harder because the new jobs require different skills and arrive unevenly across regions and sectors.
That is why the tax question is no longer a niche technocratic concern. If AI changes how wealth is generated—especially if it enables firms to capture productivity gains without hiring proportionally more labor—then the fiscal system that funds public services and social insurance may start to drift out of alignment with economic reality. Governments could find themselves paying for the consequences of labor-market strain while collecting less revenue than expected, or collecting revenue in ways that distort incentives at precisely the wrong moment.
This is where the idea of “a new tax code for the wealth AI creates” enters the debate. The phrase can sound provocative, even simplistic, but the underlying issue is concrete: modern tax systems were built around assumptions about how income is earned, how profits are measured, and how value is attributed to production. AI complicates all three.
Consider what AI does in practice. It can reduce costs, speed up decision-making, improve forecasting, automate parts of service delivery, and generate outputs that substitute for human effort. In many cases, the firm that deploys AI captures the benefit as higher margins, lower unit costs, or new revenue streams. Yet the tax base—corporate profits, payroll taxes, capital gains, consumption taxes—doesn’t automatically adjust to reflect the new mechanics of value creation. If AI-driven productivity increases profits faster than it increases wages, the distribution of taxable income shifts. If AI reduces headcount or hours, payroll tax receipts may fall relative to the scale of economic gains. If AI enables intangible assets—models, datasets, proprietary workflows—to become central to competitiveness, then the question becomes how to tax value that is hard to measure and often mobile across jurisdictions.
The challenge is not simply to raise more revenue. It is to design rules that are robust enough to handle a world where the “labor input” to production is less visible, while the “capital input” includes software, compute, data pipelines, and intellectual property that may be owned by entities spread across countries. A tax code that works well in a world of factories and payroll-heavy industries can become less effective when the center of gravity moves toward algorithmic capability and platform-like business models.
But there is another layer: the political economy of taxation. If governments respond to AI disruption with ad hoc measures—temporary levies, emergency subsidies, or one-off adjustments—they risk creating uncertainty that discourages investment and undermines trust. Workers, meanwhile, may experience the transition as unfair if they see productivity gains rising while their own prospects shrink. That is why the call for a coherent framework matters. It is not just about technical design; it is about legitimacy.
A credible framework would likely include three interlocking components: tax rules that can account for AI-created value, labor-market and social policy that stabilizes incomes during transitions, and administrative capacity that can implement these policies without turning them into bureaucratic traps.
Start with the tax side. One approach often discussed in policy circles is to broaden the tax base beyond payroll and traditional corporate profit measures. If AI reduces the labor share of income, then relying heavily on payroll-linked revenue becomes increasingly fragile. Some proposals aim to shift toward taxes that better track economic rents—returns above normal rates of profit—especially those associated with market power and technological advantage. In theory, this could capture some of the gains from AI deployment without requiring policymakers to precisely attribute every output to a specific worker or model.
Another approach is to refine corporate taxation so that profits linked to AI capabilities are taxed where value is created, rather than where ownership is located. This is where the debate about “where value is created” becomes unavoidable. AI systems can be trained in one place, hosted in another, integrated into products sold globally, and supported by data sourced from users worldwide. The old logic of permanent establishment and physical presence struggles to map onto digital and AI-driven business models. Policymakers have tried to address this with minimum tax regimes, destination-based concepts, and anti-avoidance rules. Yet the pace of AI adoption may outstrip the ability of tax administrations to keep up.
A third approach is to consider taxes tied to automation itself—sometimes framed as “robot taxes” or levies on the use of labor-replacing technology. These proposals are controversial, partly because they can be blunt instruments. Automation can complement labor rather than replace it; it can increase demand for certain high-skill roles while reducing demand for others. A tax that penalizes automation could discourage beneficial productivity improvements or push firms to restructure in ways that avoid the levy without improving worker outcomes. Still, the underlying intuition—that the fiscal system should not ignore the labor-displacing effects of automation—remains influential. The key is to design such measures carefully, perhaps by focusing on outcomes (like reductions in employment hours or wage bills) rather than on the mere presence of technology.
There is also a more subtle possibility: instead of trying to tax “AI wealth” directly, governments could tax the distributional consequences of AI productivity. If AI increases profits while reducing labor income, then a tax system that adjusts the balance between capital and labor taxation could help fund transition support. That could mean changes to capital gains taxation, corporate integration with personal taxes, or reforms to how stock-based compensation is treated. These are politically difficult choices, but they may be more administratively feasible than attempting to define and measure “AI-generated value” in a way that satisfies both economists and auditors.
Yet even the best tax design cannot solve the labor-market problem alone. Underemployment requires income stability, retraining pathways, and employer incentives that align with workforce transitions. If AI reduces demand for certain tasks, the policy response must treat the transition as a process, not an event.
A framework ready for mass underemployment would likely include automatic stabilizers that trigger when labor markets weaken in specific ways. For example, governments could expand wage insurance schemes that top up earnings for workers who accept lower-paying roles due to displacement. They could strengthen unemployment benefits and shorten the time it takes for workers to qualify, especially in economies where gig work and contract employment blur the lines between “employed” and “unemployed.” They could also invest in rapid reskilling programs that are connected to actual hiring demand rather than generic training catalogs.
But reskilling is often discussed as if it were a single lever. In reality, it is a system: training providers, credential recognition, employer partnerships, and the availability of jobs that match new skills. Underemployment adds urgency because workers may not have the time or financial cushion to retrain. If their hours are cut, they may struggle to attend courses, relocate, or take unpaid internships. That means income support and training must be coordinated. A tax framework that raises revenue from AI-driven productivity could be paired with a dedicated “transition dividend”—a ring-fenced pool used for wage insurance, training, and mobility support. Ring-fencing is not a magic solution, but it can improve political durability by making the link between AI gains and worker protection explicit.
Employer incentives matter too. If firms can capture productivity gains while externalizing adjustment costs onto workers and taxpayers, the transition will be slower and more painful. Governments could encourage firms to share benefits through mechanisms such as profit-sharing requirements tied to layoffs, or tax credits for investments in workforce development that are measurable and audited. Another possibility is to create standards for “responsible deployment” of AI in workplaces—standards that include workforce impact assessments and commitments to redeployment before layoffs. These are not purely regulatory ideas; they can be integrated into procurement rules, public-private partnerships, and eligibility for certain tax advantages.
All of this raises a question that often gets overlooked in the AI debate: what kind of underemployment are we talking about? The term can cover multiple realities. There is underemployment caused by demand shocks—when the economy slows and firms hire less. There is underemployment caused by skill mismatch—when workers remain employed but cannot command wages commensurate with their capabilities. And there is underemployment caused by task substitution—when AI automates parts of jobs, reducing the number of hours needed or the number of workers required to deliver the same output.
Task substitution is the most distinctive AI-related mechanism. It can reduce the need for coordination and supervision, compress workflows, and make certain roles redundant. But it can also create new roles that are hybrids: people who oversee AI systems, validate outputs, manage exceptions, and handle customer relationships that require judgment. The problem is that these new roles may not appear in the same places where displaced workers live, and they may require different credentials. Without policy intervention, underemployment can become a long-term condition rather than a temporary phase.
This is where the “unique take” on the tax question becomes important. The goal should not be to craft a tax code that perfectly identifies AI’s contribution