Asking an Algorithm to Pay Its National Insurance

For generations, the compact between the citizen and the state rested upon a delightfully straightforward arrangement: a human being would wake up at an unreasonable hour, perform several hours of mild administrative suffering, receive a modest paper voucher or electronic credit, and immediately surrender a non-trivial percentage of it to Her Majesty’s Revenue or its international equivalents. This system functioned with admirable predictability because human workers possess certain unalterable habits, such as requiring lunch, complaining about public transport, and eventually reaching an age where they must be quietly replaced by younger humans.

The trouble began when enterprise software evolved from a glorified filing cabinet into something that could draft memos, adjudicate insurance claims, and express mild polite regret on behalf of customer support teams. Suddenly, tax authorities around the globe observed a subtle but disturbing tremor in the national ledgers. Every time an enterprise replaces forty junior accountants with a central processing unit, the corporate treasury saves a fortune on dental plans, but the exchequer suffers a sudden and embarrassing deficit in income tax and national insurance contributions. Software, it turns out, is notoriously unhelpful when it comes to funding municipal swimming pools. It does not buy tea from the office canteen, it does not pay local property taxes on a semi-detached house in Surrey, and it shows an unfeeling disregard for the duties on tobacco and spirituous liquors.

In response to this fiscal awkwardness, policy institutes and treasury officials have begun contemplating the intriguing notion of an AI tax—or, to describe it in proper bureaucratic prose, a levy on synthetic cognitive labor. The core difficulty of this enterprise lies in determining where the software ends and the employee begins. When a human clerk processes an expense claim, the tax office knows precisely where to send the demand letter. When a neural network processes ten thousand expense claims in the span of a single afternoon while idling in an air-conditioned server farm in Reykjavik, the situation becomes decidedly murky. Does one tax the individual floating-point operation, or does one treat the entire graphics card as a single, very densely packed member of staff?

I recall observing a senior treasury official attempting to explain the concept of taxable algorithmic capacity to a committee of software engineers. The official, who had spent thirty years relying on standard tax tables designed for Victorian railway clerks, kept inquiring whether the algorithm’s neural weights could be subjected to a PAYE deduction at source. The lead engineer tried to clarify that the model was technically distributed across three continents and several thousand microchips, to which the official thoughtfully replied that they would simply need to send three separate self-assessment forms to the Icelandic hosting facility.

The administrative consequences of treating software as taxable personnel are gloriously surreal. If a large language model is to be taxed as a worker, logic dictates that it should also be entitled to standard employment deductions. A corporate server array might reasonably demand tax relief on several hundred thousand litres of chilled water, a deductible allowance for high-voltage electricity, and perhaps a small allowance for thermal paste. Furthermore, if an algorithm is subjected to payroll taxes, one must inevitably confront the question of statutory benefits. It would be a deeply awkward afternoon at the Ministry of Work and Pensions if a neural network submitted a claim for statutory sick pay on the grounds that its training data had developed a mild case of corrupted memory blocks.

Governments, however, are rarely deterred by philosophical absurdity when there is a fiscal deficit to be plugged. Proposals currently circulating through policy circles range from charging a flat rate per teraflop of computing power to levying a surrogate payroll tax directly on companies based on the number of human roles displaced by automated systems. The latter approach requires corporate managers to keep a meticulous ledger of human absence—a document listing every empty desk and every unwritten resignation letter, accompanied by a calculated estimate of how many virtual clerks are currently squatting in the cloud infrastructure.

There is an inherent irony in watching the modern state attempt to fit the infinite frictionlessness of digital execution into the rigid cubicles of twentieth-century tax law. For decades, technology companies assured us that automation would liberate humanity from the mundane burden of administrative drudgery. What they failed to mention was that the state relies almost entirely on administrative drudgery to pay for its roads and hospitals. By removing the human being from the office chair, we have inadvertently threatened the very machinery that keeps the municipal streetlights burning.

I was once informed by an accountant that the finest tax codes are those that pluck the goose with the smallest amount of hissing. The difficulty with taxing artificial intelligence is that software does not hiss at all; it merely sits in a rack, emitting a low, warm hum while quietly recalculating the global balance of trade. As revenue agencies construct increasingly elaborate regulatory nets to trap every stray output token, one suspects that the ultimate outcome will not be a harmonious compromise between silicon and statecraft, but rather a fifty-page self-assessment tax guide written entirely for an audience of graphics cards that will never bother to read it.