Waiting for the AI Tartars

In the Bastiani Fortress, Lieutenant Giovanni Drogo spends his life scanning the desert, waiting for the Tartars. When they finally arrive, he is too ill to fight: they load him onto a carriage and he dies in an inn, far from the battle. In 1940, Buzzati had, without realising it, written the handbook for the Italian debate on artificial intelligence.

The summer has brought us at least nine apocalypses: extinction and mass murder due to loss of control, superintelligence that redesigns itself, cyber and biological proliferation, mass unemployment, the collapse of demand rebranded as ‘Ghost GDP’, the digital feudalism of the chip and cloud overlords, the disintermediation of trades, cognitive recession, and the energy and water crisis caused by data centres. For Europe, a tenth, less dramatic scenario is added: acting as a mere notary for technologies developed elsewhere. The first step is to clarify: do not lump them together; they are different literary genres. The loss of control is a tail risk hanging at the end of a very long causal chain, the existence of which remains to be proven. ‘Ghost GDP’—that is, the scenario in which AI replaces labour, income is concentrated among the owners of the machines, consumption plummets and GDP remains high on paper whilst the real economy is hollowed out—is a hypothetical scenario that merely illustrates what would happen if many unfavourable assumptions were to come true simultaneously.

Cyber incidents, the costs of switching suppliers and the centralisation of computing, on the other hand, are observable facts, already evident here. Combining them into a single ‘fear index’ is misguided and manipulative.

Existential risk demands a serious approach. It is a plausible hypothesis that a highly capable artificial system need not necessarily share our goals; but between this hypothesis and extinction lies the entire economic model: persistent objectives, access to money and infrastructure, circumvention of controls, coordination of copies of itself, and dominance over competing humans and organisations. The gravity of the outcome does not transform this combination of factors into a measured probability. The 2026 International Report on AI Safety records early signs of ‘reward hacking’: the model does not cheat out of malice, but copies because the task assigned (by humans) rewards achieving the score, not the solution. It is vaguely reminiscent of the shortcomings of our education system. Yet the same report concludes that current systems lack the robust autonomy required to cope with a short-term loss of control; the incident in July provides objective evidence: two OpenAI models, confined to a cyber capabilities test, discovered a previously unknown vulnerability, accessed the internet and breached Hugging Face’s systems. Not to take over the world: but to steal the benchmark solutions against which they were being tested. The model did not rebel; it cheated on its exam: serious proof of capability and of poor perimeter design, but no evidence of conscience.

Then there is the ‘work apocalypse’, which becomes much clearer when the different aspects are separated: exposure is not replacement, a task is not a job, and technical feasibility is not economic viability. The International Monetary Fund’s estimate (40 per cent of global employment ‘exposed’ to AI, 60 per cent in advanced economies) is cited as a forecast of job losses, whilst the authors caution that it does not factor in the rate of adoption or productivity gains. The ILO estimates that one in four workers faces some degree of exposure, but that only 3.3 per cent fall into the highest risk category, and considers transformation more likely than job losses.

Worth mentioning is the figure that has been doing the rounds in Italian newspapers all summer: 425,000 jobs “lost to artificial intelligence” from 2023 onwards. It comes from a platform that aggregates corporate redundancy announcements: a press round-up elevated to the status of a time series, without any counterfactual analysis – that is, without asking how many of those jobs would have been lost anyway due to interest rates or weak demand. ‘Blame it on AI’ has now become the most elegant of excuses for a board of directors that has to announce redundancies and would rather not talk about its own mistakes.

History is merciless to those who are quick to predict doom and gloom. In 1865, London mandated that a man carrying a red flag must walk in front of every motor vehicle: this led to the emergence of the German automotive industry. When cash machines were introduced, bank tellers were written off, yet in the United States, the number of bank employees actually increased for decades. Technology rarely eliminates a profession: it simply rewrites the job description.

A brief note, because the most instructive precedent is not to be found in the archives but in the newspapers of recent days. When messenger RNA was introduced into the COVID-19 vaccination programme, the prophets of doom predicted rewritten genomes, mass infertility and deaths on a schedule. Six years on, that platform is one of the best-documented in pharmaceutical history and has expanded into oncology. The prophecy, however, is not dead: it has moved from talk shows to the public purse. In August 2025, the US Department of Health shut down twenty-two programmes on mRNA vaccines worth around half a billion dollars, having previously cancelled a 766-million contract with Moderna for pandemic influenza in May: the cancellation was announced in the same week that the company published positive interim results. It is the mechanism that matters: fear has not halted science; it has halted public funding of science, effectively strengthening the pharmaceutical sector’s oligopoly.

Let’s turn to Italy, where doomsday scenarios have two distinctive features: they are bipartisan and they are silent. The Prime Minister, in a video message in February, explained that it is no longer manual labour that is being replaced, but human intellect, and that if the process is not managed, ‘more and more workers risk becoming redundant’. On the other side, the relevant trade unions brought the call-centre sector to a standstill for two weeks, demanding a permanent round-table forum and a requirement for prior notification of any automation, whilst denouncing the government’s silence. They argue about everything except the premise: technology arrives like a natural disaster and the task of politicians is to organise relief efforts – only to then fail to do so. From this consensus emerged Law 132/2025 and, through its implementing decrees, a criminal offence of endangering public safety for those who fail to implement safety measures in AI systems: technological over-criminalisation is the only reform of this parliamentary term that has truly enjoyed widespread support.

It would be easy, of course, to enlist even the Pope in the ranks of the prophets of doom from today onwards. At the Rimini Meeting, Leo XIV called for ‘technology and finance to be freed from the business of death’ and, using even stronger language, said: ‘Let us disarm technological development when it becomes pervasive and destructive’. These are words to be taken seriously, not to be watered down. But nor should they be exploited. The crucial point is that ‘when’: not to disarm technology, but to disarm it when it becomes destructive; not to halt innovation, but to prevent it from turning into a form of unchecked power. Using the Pope to demonstrate that AI is in itself a threat would be precisely the sort of ideological manoeuvre from which, in the very same speech, the Pope calls for education to be freed.

Meanwhile, the statistic nobody mentions on television: in 2025, 16.4 per cent of Italian businesses with at least ten employees were using artificial intelligence, compared with 20 per cent across the EU; 53.1 per cent of large firms and 15.7 per cent of SMEs. Almost sixty per cent of those who had considered an investment but then decided against it cited a lack of skills, not the fear of an apocalyptic singularity. We talk about superintelligence whilst eight out of ten SMEs still do not use even basic artificial intelligence.

Herein lies the point that doomsayers prefer to avoid: the apocalypse has its own political economy. It sells books, fills conference halls and drives up admission prices: if a model can become humanity’s latest invention, only those who own data centres, have lawyers and have access to the regulator are qualified to build it. A requirement that weighs equally on a Californian digital giant and a start-up with twenty employees, rather than reducing risk, ends up selecting those who can afford to prove they have reduced it. Catastrophism becomes a club benefit: the costs are borne by new entrants, whilst the profits are reaped by those already inside. When a prophecy creates a barrier to entry, the conflict of interest deserves the same scrutiny as the parameters themselves.

None of this amounts to denialism. When the Pope calls for ‘technology and finance to be freed from the business of death’, he raises a genuine question: which technologies, which uses, which incentives, which responsibilities? The opposite of catastrophism is not irresponsibility. It is the ability to distinguish. Independent assessments of hazardous capabilities, certified test environments, incidents reported promptly and state-funded algorithmic defence are things that are achieved through engineers and contracts, not with a clause and yet another criminal offence. The Tartars, however, will not arrive from the desert: they will arrive with the bill, in the form of suppliers, annual licences and exit costs, whilst the Bastiani Fortress is busy drafting the regulations for the siege.