Radio ExpressTV
Live
How to Tax Artificial Intelligence
Sami Mehrroum: Founder of Spark X, previously held positions at INSEAD, the OECD, and Nesta.
The commitment by newly elected New York City Mayor Zahran Mamdamani to raise the minimum wage to $30 an hour illustrates a dilemma that plagues most advanced economies. Even with increases in minimum wage, workers feel a decline in job security. However, by responding with blunt instruments like minimum wage, policymakers overlook a deeper issue: the disconnection between human work time and economic value.
For two centuries, labor markets have relied on this relationship. Wages, contracts, and social protections all assumed that time is a reliable indicator of productivity. But artificial intelligence has severed this link. Now, disease diagnostic experts who have spent years mastering pattern recognition compete with systems that process cases in seconds. Lawyers using AI tools can complete tasks in minutes that once took hours. The question is not whether minimum wage risks missing the mark; rather, it is whether time-based compensation still reflects reality at all.
So what is to be done? First, financial frameworks must recognize machine time as a distinct productive input and price it accordingly. This does not mean proposing a “robot tax,” but something closer to a measurable input tax, using data that is already being tracked in real time.
Direct attempts to tax companies based on their level of automation have never succeeded, as such processes are ambiguous from a regulatory standpoint. The algorithmic decision-making embedded in programs is often nearly invisible, and attempts to define “robot” have proven legally nonsensical (Is Excel a robot?). Furthermore, automation has long been seen as essential for maintaining global competitiveness, posing a “problem” primarily for those who provide manual labor.
Policies that penalize productivity tend to be hopeless. Notably, in 2017, the European Parliament considered a proposal — embedded in a report by Maddy Delvaux on robots and artificial intelligence — to levy a “robot tax” on companies that replaced workers with automated systems. The idea was framed as a way to fund social protection and retraining, but it was explicitly rejected in the final council vote amid fears it could stifle innovation and harm Europe’s competitive edge. Since then, proposals for a robot tax have remained largely theoretical across EU member states, reinforcing the political reality that productivity-penalizing policies face significant obstacles.
The solution may lie in taxing “AI hours”: the computational time consumed by AI systems performing productive economic tasks. As computational time is already precisely measured for billing purposes in the cloud industry, which supports most AI services for businesses, AI hours represent one of the rare tax bases with built-in and automatically recorded audit trails.
Consider a medium-sized law firm in the United States using AI to automate contract review, replacing 25 full-time legal assistants averaging $65,000 annually. The hourly wage for this displaced workforce is $32.50 (calculated as $65,000 divided by 2,000 standard working hours per year). If the firm consumes 50,000 AI hours annually in this automation, the taxable value of the displacement is $1,625,000 ($50,000 hours × $32.50). Imposing a 15% tax on this displacement value generates $243,750 in annual tax revenue. This framework taxes the economic value of human labor that AI systems replace, measured automatically based on wage levels across various professions, linking the tax burden to the extent of labor displacement.
While this might seem modest, it becomes significant when aggregated across all legal services, financial analysis, and medical diagnostics. By 2028, AI infrastructure spending in enterprises will reach $200 billion globally.
It is true that edge computing, which processes data close to its source, complicates matters, as usage records are localized. In these cases, the tax base would shift from data flows to capacity agents, such as chip specifications and energy consumption. Since regulatory bodies already use these metrics for carbon footprint reporting and energy audits, firms would merely need to report devices supporting AI over designated thresholds and provide quarterly usage estimates, cross-checked against capacity and energy consumption remotely.
Of course, there is a risk that companies may manipulate the system to qualify for lower tax rates. Clearly defined definitions will be crucial. The two key terms that require utmost clarity are “augmentation” and “replacement.” Augmentation refers to cases where humans use AI but remain the primary decision-makers. Replacement occurs when AI entirely executes a workflow, with human intervention only in exceptional circumstances. Therefore, a radiologist reviewing the anomalies identified by AI, yet making the final diagnostic decision, is augmented. Conversely, an AI system that automatically approves 95% of loan applications is replacing humans.
With these terms clearly defined, it becomes feasible to levy a tax on AI hours at scaled rates, say 5% for augmentation and 15% for replacement. Workflow audits will determine classifications. If a company claims to use augmentation while humans actually handle less than 40% of cases, the system will be reclassified. While the precise rates can be debated, the fundamental principle mirrors the current tax differentiation between operating expenses and capital expenditures. Even with a 15% tax, AI hours will remain significantly cheaper than human labor.
The purpose of this tax is to nudge companies toward hybrid systems where human judgment enhances outcomes, rather than toward fully autonomous setups. This would render labor displacement slightly more costly than augmentation without stifling innovation. The industry itself should welcome such a policy. The real threat to AI progress is not taxes, but a violent political backlash. A predictable, rule-based tax could ensure social legitimacy for AI by ensuring shared gains.
But what about the global dimension? As AI hours become a core element of production, goods and services generated using unpriced automated time may undermine those not produced with such aid. The encouraging aspect is that we already have a method to address such arbitrage. A digital boundary adjustment mechanism (DBAM) could require businesses to disclose the content of AI hours in the creation of traded goods and digital services across borders, similar to existing carbon border adjustment mechanisms. Implementation would not be difficult, as internal pricing rules already compel multinational companies to provide comparable documentation.
An AI hours tax would have broad appeal. Under it, companies would receive predictable rules instead of politicizing wages for purpose; workers would gain a mechanism to benefit from the technological surplus (without hindering its adoption); and financial conservatives would find stability as payroll taxes erode, while progressives would obtain a tool to narrow inequality gaps without resorting to potentially unconstitutional wealth taxes.
Coordination remains the chief challenge. Beginners in adopting this mechanism may face pressures affecting their competitiveness in the short term until digital border adjustment systems mature. However, multilateral agreements — as with carbon pricing — could create a critical mass.
The debate over whether the minimum wage should be $16 or $30 misses the fundamental point. AI systems operate continuously at minimal costs. Without a new financial framework, we will continue to drift toward an economy with a dwindling tax base. An AI hours tax could convert machine productivity into public revenue without stifling creativity and innovation. It is the best available response.
