As AI systems continue to be adopted at scale, they are increasingly performing tasks carried out by human employees. From drafting documents, to writing code, to handling customer queries and concerns have grown about the economic risks this shift may pose, including job displacement and the erosion of the income taxes generated by human labour.
This rapid automation of roles raises difficult questions around labour displacement, reduction and redeployment. A direct economic consequence of this is the growing fiscal pressure that governments may face due to loss of income tax. This has renewed interest in proposals for a robot tax, alongside the emergence of AI policy research on alternatives such as token tax and FLOP tax, each targeting a different point in the AI value chain: labour displacement, AI-driven usage and compute power.
In short, as companies increase the pace and scale of AI adoption, there is an increasing likelihood that today’s efficiency gains will become tomorrow’s tax liabilities. Understanding if and how governments are proposing to tax AI deployment is therefore a new, but critical component of effective business strategy and tax planning.
Robot tax
One potential economic consequence of businesses automating roles previously performed by human employees is a reduction in income tax revenues generated from labour.
A key policy question is whether AI systems that perform tasks traditionally carried out by workers should be taxed as capital assets, rather than as a substitute for labour. Treating it as capital can create a tax-driven incentive for firms to replace human workers with automated systems, and a robot tax would eliminate this fiscal advantage and restore tax neutrality between human and automated labour.
This would also prompt businesses to consider the real cost of labour replacement by weighing the benefits of human labour against automation. A robot tax can preserve government revenue that would otherwise be lost to automation and can be used to fund retraining programmes or unemployment support for displaced workers. However, taxing the use of machines can disincentivise investment in new technology, and the method of taxation would require a careful exercise of balancing labour interests with advancing innovation.
Token tax
Tokens are small units of data that come from breaking down whole words into parts of words, or punctuation, that AI models process language in. This is a computational unit, and not a linguistic one. AI providers typically use token-based pricing for AI models, with both inputs and outputs being chargeable.
Unlike a robot tax, which targets labour substitution, a token tax would seek to capture value from AI usage directly, regardless of whether specific jobs are displaced. A token tax would be a levy applied to the provider’s billed token cost and may be simpler to implement as tokens are already tracked and billed by AI providers, making it a measurable proxy for AI-driven business activity.
One of the challenges is that taxation may vary by model type. Models with less efficient architecture generate far more tokens for the same task than advanced models. The tax burden would depend heavily on which models a business happens to use rather than the value or complexity of the work performed.
FLOP tax
Significant computational resources are required to train and operate increasingly powerful models. The most capable AI systems depend on vast amounts of compute, measured in FLOPs. FLOP regulation is already within regulatory scope of the EU under the EU AI Act, whereby compute usage beyond a certain threshold would serve as a proxy for model-capability and trigger enhanced safety controls for general-purpose AI models that pose systemic risks.
Beyond systemic risks, however, compute can also raise economic challenges as resource concentration can create barriers to entry. A FLOP tax (also called compute tax) would be borne by the AI model provider, and impose a levy based on the volume of compute used to train or run AI systems.
Critics argue that a FLOP tax could discourage investment in AI innovation because taxing the very infrastructure needed for AI development is likely to be a self-defeating policy. And, if a FLOP tax were introduced, it would be surrounded by challenges around accurately measuring and monitoring compute usage given the global nature of the AI infrastructure. Compute resources, data centres, cloud networks, semiconductor manufacturers and model developers are geographically dispersed, with a handful of jurisdictions occupying critical chokepoints in the AI supply chain. This creates significant international interdependence, while also raising questions about where a FLOP tax should be imposed and if and how it could be enforced. In the absence of international coordination, AI providers may relocate compute-intensive activities to lower-tax jurisdictions, potentially undermining the effectiveness of the tax, increasing geopolitical competition for AI investment and complicating national efforts to achieve AI sovereignty.
Conclusion
Regardless of whether AI use is taxed, it seems clear that widespread AI adoption will not only reshape the future of work, but will also shift the cost base of most, if not all industries, leading to AI-driven fiscal challenges. Enterprises that are currently adopting AI should therefore stay abreast with developments as AI efficiencies could potentially introduce costs in the future that are not currently anticipated.
