The cost of not thinking and productivity that goes unmeasured

The productivity that AI promises is measured in terms of time saved. The productivity it destroys cannot be measured at all. It empowers those who think and quietly replaces those who delegate. The distinction is not a technical one. It is a choice.

The public debate on artificial intelligence oscillates, with predictable regularity, between two extremes: enthusiasm for its promise and fear of being replaced. On the one hand, there are leaders and decision-makers who see technology as an inevitable driver of growth for economies that have fallen behind in terms of productivity. On the other, there are economists and analysts who point to actual data – such as productivity growth and indicators that rule out the simple effect of demanding more from people and machines – and conclude that the ‘something spectacular’ heralded by the sector has yet to show up in the figures.

Both sides, however, share an assumption that is rarely questioned: that the value of artificial intelligence is measured by the amount of work it replaces or speeds up. It is this assumption that deserves closer scrutiny.

The personal computer and internet revolution removed barriers to finding, storing and transmitting information. The productivity gains were relatively straightforward because people could simply do more – and do it better – of what they already knew how to do. Generative artificial intelligence automates something of a different nature, namely the very production of cognitive outputs, as it writes, synthesises, reasons and makes recommendations. And it does so with such fluency that it is difficult, at first glance, to distinguish between a competent output and one that is plausible but incorrect.

This is where a tension arises that the debate on productivity tends to obscure. When a system produces a draft, an analysis or a screening decision, someone has to verify that result. This verification requires precisely the specialist knowledge that the tool is supposed to make redundant. The time saved in the development phase is partly, and sometimes entirely, taken up by certification. Not because the technology is poor, but because responsibility for the outcome does not disappear when tasks are delegated. It simply shifts.

This shift has a consequence that is rarely factored into corporate adoption models. Organisations that respond to artificial intelligence by hiring fewer junior analysts, fewer early-career lawyers and fewer entry-level technicians are simultaneously eroding the knowledge base needed to oversee the systems they have adopted. A máquina produz resultados em volume infinito. A capacidade humana de os validar é, por definição, finita e depende de uma competência que se constrói com tempo e exposição. Quando essa cadeia de formação se interrompe, o erro oculto cresce sem que a organização tenha desenvolvido os instrumentos para o detetar.

The debate on the taxation of artificial intelligence, the so-called ‘robot tax’, and proposals for a tax on windfall profits made by companies in the sector indirectly reflects this tension. The strongest arguments against specific taxes point to the risk of diverting investment and penalising innovation. But the deeper issue is another: the existing tax system was designed to capture value generated by labour and consumption. As the share of capital in production increases and that of labour decreases, the tax base weakens regardless of technology. Artificial intelligence accelerates this process, but it did not create it. Making it the specific target is merely treating the symptom.

What is missing from this debate is not more regulation or more taxation as ends in themselves. It is a more fundamental question: what, exactly, is the purpose of artificial intelligence in the process of knowledge creation and decision-making?

The prevailing view is that of the assistant: a tool that executes, summarises and makes proposals, freeing the user to focus on tasks with greater added value. This view makes sense, but it assumes that the user retains the ability to judge what the tool has produced. A data dashboard does not make decisions; it presents structured information that a decision-maker interprets based on their own understanding of the problem. Artificial intelligence, when used as an assistant rather than as an autonomous decision-making system, should function in the same way. The problem is that the fluidity of their results creates an illusion of completeness that traditional dashboards have never been able to achieve. A poorly constructed graph immediately arouses suspicion. A well-written text with a flawed conclusion requires the reader to make an active effort to verify it – an effort that the very fluidity of the text tends to inhibit.

There is a tension here that is neither technical nor economic. It is epistemological. Artificial intelligence, when used as a substitute for cognitive effort rather than as a support for it, does not strengthen the user’s capacity; it may actually weaken it. Not because the tool is bad, but because knowledge is built through the process of reaching a conclusion, not simply by having access to it. A worker who systematically delegates diagnostic reasoning to an automated system is not becoming more productive in the sense that matters; rather, they are gradually losing the ability to question the system when it fails.

The public debate on artificial intelligence therefore needs to make a distinction that is not yet sufficiently clear: between adoption that enhances human capability and adoption that quietly replaces it. The former requires investment in training, critical thinking and verification mechanisms. The latter yields short-term gains and long-term vulnerabilities that only become apparent once an error has already had consequences.

Productivity that goes unmeasured is not the productivity that artificial intelligence has not yet generated. It is the productivity that is lost when the effort to understand is replaced by the comfort of delegating.

References and inspiration

• Negócios n.º5747 jornaldenegocios.pt