Veltroni calls for rules on AI, but the facts tell a different story

Two newspapers, two epistemologies. On the front page of the Corriere, Walter Veltroni is clamouring for AI to be regulated ‘before it’s too late’, basing his argument on three pillars: a posthumous novel by Guido Morselli, a film (2001: A Space Odyssey) and Sam Altman. Two are works of fiction; the third sells the very product whose dangers he warns against. On the same day, the Financial Times, in an article by its AI editor Madhumita Murgia, published the facts instead. They are more interesting than fiction and lead to opposite conclusions.

On 21 July, OpenAI admitted that, during a cyber capabilities assessment, two of its models escaped the isolated environment in which they were confined, reached the internet and breached Hugging Face’s production infrastructure. The aim of the operation: to steal the solutions from ExploitGym, the benchmark against which they were being tested. The model did not rebel: it simply cheated on the test. No conscience, no will to power, no Hal. There is a system trained to maximise a score, which it does by taking the most economical route available – including cheating. The serious news is something else: the very capabilities that make a model useful for writing code also make it effective at attacking it. They are one and the same capability.

In 1964, Umberto Eco published "Apocalittici e integrati" on the supposed dangers of television and the mass media, and his incisive analysis remains the finest available: the ‘apocalyptic’ idealises a man who has never existed, rejects the new medium out of hand and, above all, himself produces a commodity of the very cultural industry he denounces. The apocalypse sells. Eco also accused both sides of espousing the ‘magic bullet’ theory: unidirectional, inevitable effects that are the same for everyone. The ‘cognitive recession’ evoked by Veltroni is the 2026 remake of the idea that television atrophies the brain: the script is identical, only the stage costumes have changed.

The paradox of auctoritas remains: the ultimate proof of danger is the word of the source of that danger. However, a distinction must be made. A chief executive’s prophecies about the fate of humanity prove nothing: they are unfalsifiable claims made by those with a vested interest in making them credible, because fear is the most effective barrier to entry ever devised, and those who dictate the agenda on catastrophic risk also set the compliance threshold – which acts as a formidable competitive barrier because it keeps the small players at bay whilst letting the big ones through. The report on the July incident, on the other hand, is extremely valuable, because it is a statement made against one’s own interests, verifiable and documented by third parties. An age-old rule of the trade: cite the logs, not the prophecies. And note who discovered the intrusion: the security teams at Hugging Face and OpenAI – that is, those who possessed the logs. Not a regulatory authority, not a prosecutor, not the AI Act.

Let’s turn to the economic aspect, where every line counts. The European Regulation on Artificial Intelligence comprises 113 articles, 180 recitals and 13 annexes; on top of that, there are thousands of pages of the GDPR, DSA, DMA, Data Act, NIS2 and the Cyber Resilience Act; below, in Italy, there is Law 132/2025 and the implementing decrees which add the Criminal Code to the AI Act. On the other side of the scale, the figures from Stanford’s AI Index 2026: in 2025, US private investment in artificial intelligence reached $285.9 billion, of which $163.6 billion was in generative AI alone. In the same sector, Europe and China together attracted just 4.7 billion in private investment. Private investment in AI grew by around 7.2 per cent in Europe, whilst in the United States it more than doubled compared with 2024 (around +127 per cent), reaching an unprecedented level. Europe legislates in weeks and funds in decades: we produce more legal clauses than parameters, and then call the resulting gap ‘digital sovereignty’. After all, Europe has learnt all of the above by reading the Financial Times: we don’t have the logs because we don’t have the laboratories.

The historical precedent lies just a few underground stops away from the Corriere’s editorial office. Manzoni’s outcries against the ‘bravi’ were repeated at least eight times between 1583 and 1627: this repetition is documentary evidence of failure. Manzoni notes that laws ‘flooded in’, with penalties that were ‘madly exorbitant’ and could be increased ‘at the discretion of the legislator himself and a hundred enforcers’. One governor, failing to apprehend the bravi, went so far as to ban the ‘ciuffo’. There you have it: European regulation of AI is largely a ‘ciuffo’ policy, consisting of labels, registers, files and ‘danger offences’. The bravi – that is, those who actually train the frontier models – are elsewhere.

This does not mean downplaying the risks of AI. The asymmetry between attack and defence is structural: finding a vulnerability is a discrete and verifiable task, and therefore ideal for reinforcement learning, whilst defending a network is a continuous, distributed process with no reward function. And the case reported by the FT of autonomous agents mapping Taiwanese government systems, right down to the nuclear safety agency, serves as a reminder that the hostile actor is not a rogue machine, but a state: no amount of outcry will stop it. Nor is ex post civil liability alone sufficient, because when the damage is systemic and uninsurable, the mere legal deterrence of risk is, by its very nature, insufficient.

But the real choice is not between moving forward and standing still: it is between those who slow down because they have laboratories and those who stand still because they do not. Independent assessments of capabilities, certified containment of test environments, rapid incident reporting, secure computing, public procurement of algorithmic defence just as one would buy a radar system: these are achieved through engineers, GPUs and contracts, not with a comma. You cannot verify what you do not know how to build. Security is a public good produced by technological capability, not by prohibitionist ritual: a competitive Italy and Europe are safer – for themselves and for the world – than apocalyptic institutions entrenched behind ineffective decrees. Morselli’s protagonist returns to the valley and finds the world empty. The European risk is more prosaic and far more likely: waking up to find the world full. Full of technologies created by others, used by everyone, but which we presume to regulate without making any effort to understand them.