A few figures to give us a better idea. Researchers asked 1,682 participants to read six short stories and rate both their quality and how engaging they found them. Each participant was told – correctly or otherwise – whether the text had been written by a person or by an AI system.

Stories generated by artificial intelligence are rated 6 per cent higher in quality and 8 per cent more engaging than those written by humans. Furthermore, the researchers note, they receive even higher ratings when participants believe they were written by a human author; in such cases, the rating increases by a further 3 per cent.

The study continues with a second experiment and offers further food for thought. A further group of around 900 people were presented with other texts, some generated by artificial intelligence, others written by humans. When asked who the author was, participants answered correctly in 39 per cent of cases in one test and 52 per cent in the other. “Previously reported familiarity with AI systems seemed to help in recognising the typical patterns of AI-generated writing,” says Weisberg. “This suggests that improving AI literacy could help people navigate the new world in which we live, one that is increasingly shaped by AI. Those who claim to have a literary background, on the other hand, proved to be less accurate in their assessments.”

The study also reveals another finding: AI-generated texts often better suit readers’ tastes and are easier to understand. “The models that power AI,” observes Cameron Jones, a psychologist at Stony Brook University who was not involved in the research, “learn to become a sort of average, inoffensive and accommodating individual, capable of appealing to just about everyone.”

Writing in The Guardian, Luke Kennard, a lecturer in Creative Writing at the University of Birmingham, goes a step further: “As humans, we tend to personify AI and imagine it as a sort of glowing robot with a benevolent, neutral face, when in reality it is merely predictive code based on a vast, stolen database of real-world texts, physically housed in vast, expensive and destructive data centres with catastrophic repercussions for the surrounding communities.” Kennard adds that artificial intelligence is not a ‘tool’ comparable to a spell-checker or a thesaurus, but rather ‘an existential threat’. Weisberg does not share this view, emphasising that artificial intelligence systems are already capable of generating short stories considered at least as good, if not better, than those written by humans, to the extent that we need to update our understanding of what AI is actually capable of.

In this exchange, however, it seems to us that the issue is not linked solely to the theme of ‘artificial writing’ – which certainly carries its own weight – but also to our notion of authenticity: the idea that a story is worth more because behind it lies a biography, hard work, who knows what insight or what ‘pen’. When put to the test using a scientific method, rather than relying solely on indignation, that idea reveals itself to be what it probably always has been: an aesthetic prejudice – reassuring and almost impossible to recognise when it is one’s own.

A study published in "npj Science of Food" and authored by researchers at Stanford University used an AI model to create new recipes – specifically, new burgers, given the success of the Big Mac. The authors explain that, rather than using an AI model such as ChatGPT – a Large Language Model that would have allowed them to programme rules regarding flavour and texture – they used a diffusion model, a type of AI that automatically learns statistical patterns by observing combinations of ingredients and their quantities. The results show that by learning the statistical patterns of thousands of burger recipes, the AI was able to propose a number of them based on palatability (i.e. how pleasant they are to the palate), nutritional value and environmental impact. According to more than a hundred evaluators, two burgers – the ‘Delicious’ burgers – were generally judged to be better than the Big Mac; in some cases, they were found to be meatier or more flavourful.

However, the satisfaction scores for the two burgers that took environmental impact into account – the ‘Sustainable’ burgers – were lower than those of the Big Mac for the mushroom-based burger, and similar for the mushroom and beef-based one. Finally, it is not hard to imagine that the least popular burgers were those designed to prioritise nutritional quality over palatability – namely the bean-based ones – which were, however, almost twice as healthy as the Big Mac and also significantly less harmful to the environment.

According to the authors, the AI used in this study was able to encode “the collective knowledge of human taste”, transforming the design of new foods “from a craft-based practice of trial and error into a data-driven design science”. However far-fetched it may seem, this approach could prove useful to people with specific dietary needs or those wishing to personalise their diet. Provided, of course, that they are actually hungry in the first place.

To understand what I am talking about, remember that the biblical God created the world by speaking (“God said, and so it was”) – that is, through the Logos – and that the Gospel of John begins with “In the beginning was the Logos”, meaning language. It is no coincidence that in all philosophies, even secular ones, the human being is at the pinnacle of creation because of their possession of the Logos.

What is happening with AI? We are presented with a mechanical device that appears to possess the very ability we knew made the human being unique. So unique that we could conceive of it as the image of God creating the world through the Logos. One cannot help but feel a sense of crisis.

But is that really how things stand? That is the question posed by the numerous books and articles on AI that have been filling the culture sections of newspapers and the shelves of bookshops for some time now. I must confess that, on this subject, I have more problems than solutions.

I would like to set out my view on a much-debated issue at the moment: granted, AI handles language more or less like humans do, but is it conscious of what it is doing? The almost unanimous answer is: no, AI has no consciousness; its actions are ‘agere sine intelligere’ – to use the phrase coined by Luciano Floridi, one of the most authoritative scholars in the field.

But what is consciousness? Since the definitions are all tautological (“consciousness is self-awareness” – and other such platitudes), some scientists and philosophers resort to an operational description by rephrasing the question as follows: “How does a human know that they are conscious?”. The answer may seem strange, but it is the only verifiable one: a human knows they are conscious by telling themselves so. And again: how do I know what I’m thinking? The answer is the same: I know what I’m thinking by telling myself. What’s more, it is often others who help us understand what we are actually thinking by asking us questions or clarifying the meaning of what we tell them. The psychoanalytic consultation is the paradigmatic example of this.

What does all this mean? Human consciousness is inseparable from language. Our own speech and that of others is the mirror in which and through which we seek and construct our sense of self. Without that mirror, there is no such thing as ‘self’. Corollary: if a mechanical device uses language like a human, why on earth should we deny it some form of consciousness? Do we not attribute consciousness to our human interlocutor on the basis of their speech?

It is impossible not to think of the Gospel of John: “In the beginning was the logos (translated as ‘Word’ in official texts) and the logos was directed towards God (my translation) and the logos was God”. There remains much to study and reflect upon.

Now that AI exists, the floodgates have opened. Since it is impossible to ban it, every effort is being made to regulate it. The University of Modena and Reggio has made 30,000 licences available to lecturers and students – almost one per person – specifying that ‘AI may be used as a support for understanding complex concepts, developing ideas, carrying out self-assessment activities and reworking information’. But understanding concepts and reworking information is precisely the point of studying. If we delegate this to AI, it is tantamount to encouraging people to stop thinking. Moreover, there is no detection software efficient enough, partly because the distinction between human and artificial writing is now too subtle. So, what should we do? Return to the oral tradition of the Italian university. Asking students to churn out reports is a practice that originated in nineteenth-century German seminars and was subsequently exported to the Anglo-Saxon world, but which, given the times, is proving to be a double-edged sword. Oral exams take longer and are more demanding for everyone, granted. But that is precisely why they are more useful and revealing. As for dissertations, it would be a very good idea to do away with them. Why expect such a complex writing task from young people who can barely read a novel?

Let’s seize the opportunity: instead of asking the two customary questions based on a pile of text written by who knows whom – questions that wouldn’t even stump Pinocchio and would just make the audience yawn – let’s turn the degree examination into a serious and in-depth discussion of a topic agreed upon between the professors and the student. Only in this way will the student’s knowledge of the subject, their ability to argue a case and to respond to objections be put to the test. AI has not yet learnt to break a sweat in front of an examination board. And without butterflies in the stomach and performance anxiety, what sort of university would it be?

This letter is not a libertarian manifesto, nor is it an act of corporate philanthropy. It is an economic policy document written by individuals with very concrete interests (and associated conflicts). But it is precisely for this reason that it deserves attention. When the leading players in the American AI sector, amidst fierce competition with China, argue that ‘open weight’ models – that is, models that can be downloaded, inspected, adapted and run on one’s own infrastructure – are a decisive component of technological leadership, they are saying that the AI race will not be won solely by cutting-edge laboratories. Above all, it will be won by those who make it easier to roll out AI across ordinary sectors of the economy.

It is important to be precise here. ‘Open weight’ does not automatically mean ‘open source’. In public debate, the two terms are often used interchangeably, but under European law they are not the same. The AI Act distinguishes between general-purpose systems and models and recognises specific exceptions only under certain conditions; for general-purpose models, even if they are open source, obligations regarding copyright and the synthesis of training data remain, and these are stricter where systemic risks are present. We must therefore stop treating ‘open weights’ and ‘regulatory free zone’ as synonyms. That said, Huang’s point remains valid. If the weights are available, innovation is not confined to a handful of proprietary APIs. Universities, start-ups, businesses, hospitals, factories and public administrations can adapt a model to their own domain, run it in their own environments and use their own data. Lock-in is reduced, and competition is increased. And for a Europe that is unlikely to match the budgets of American frontier labs or China’s aggressiveness in the short term, this is good news.

Europe’s real interest does not lie in training the world’s largest model from scratch every time. It lies in entering the most economically attractive part of the value chain: fine-tuning, industrial integration, on-premises deployment, vertical applications, AI for manufacturing, healthcare, professional services, supply chain software and public administration. In other words: utilising the division of labour rather than dreaming of costly digital autarky. Open weights make this participation far more plausible, because they allow even non-giant players to create value without having to constantly seek permission from the API owner.

But here we must avoid the opposite ideological shortcut. Defending open weights does not mean denying the risks. The letter itself acknowledges this: once released, the weights escape their original control, and modified versions may be difficult to trace or revoke. The European Commission, in its guidelines on general-purpose AI models, emphasises that open-sourcing the most advanced models can make it easier to circumvent or remove risk mitigations. The question to ask is whether the response should be a blanket crackdown on open models or a more targeted regime focusing on risks and genuinely sensitive uses.

This is where much of the European criticism misses the mark. Reducing AI to an object to be classified, documented and brought into compliance as soon as possible is an understandable, yet insufficient, bureaucratic temptation. The AI Act is not as crude as its critics sometimes portray it: it distinguishes between models and systems, provides for exceptions, sandboxes, codes of conduct and phased implementation. The problem lies in the culture in which it will be implemented: in Europe, and often in Italy, any margin of discretion risks becoming a new administrative burden, any grey area a precautionary brake, and any risk a reason to extend ex ante control.

Reducing AI to an object to be classified is an understandable, yet insufficient, bureaucratic temptation

The most sensible approach is not to ‘do away with the rules’, but rather to avoid blanket restrictions on open-weight models; stricter requirements and more thorough checks are, however, legitimate when dealing with high-risk cases, essential services or contexts in which automated output could cause systematic and potentially irreversible harm.

For Italy, the implication should be quite clear. There is no reason to add layers of domestic caution to an already complex European framework. The state’s role is not to prevent open weights from circulating. It is to distinguish where strong caution is required and where, instead, there should be scope for competition, testing, sandboxes and sector-specific adaptation.

What is interesting for Italy, however, lies in a historical pattern regarding general-purpose technologies: they do not generate value where they are invented, but where they are adapted. Italy will not be developing cutting-edge models, and that is no reason to be disheartened. General-purpose technologies are monetised through co-invention, and co-invention requires precisely those qualities in which Italian specialisation is competitive: small markets, many vertical niches, and margins protected by tacit knowledge and long-standing customer relationships. This is the opposite profile to that needed to build a generalist model, and it is the right profile on which to build.

Italy’s profile is the opposite of what is needed to build a generalist model, but it has the right qualities to build on it

The most obvious example is the instrument-making industry. Italy is one of the world’s leading exporters of machine tools and equipment for packaging, bottling, paper and food processing, and every machine sold generates a stream of data that is currently largely lost. A small model, trained on faults specific to that particular family of machines, which runs inside the electrical cabinet and flags up a bearing before it fails, is worth more than a generalist system that knows everything and understands nothing. And no serious customer accepts that their process data should pass through a third party: local execution is an industrial requirement. It is also the path by which a machine manufacturer can become a provider of recurring services, making that transition from sales to subscription fees that the Italian engineering sector has been pursuing for years.

The second scenario concerns industrial districts: many small businesses, distributed knowledge, and a technical lexicon not found in manuals. None of the companies in Belluno, Arzignano, the Cusio area or the Emilia-Romagna packaging sector can afford an AI department; a consortium, however, can afford to fund training. This is the case with the Ucima project for AI in the packaging supply chain: a shared resource, paid for once and used by all.

The third point concerns language and documents. General-purpose models speak Italian well but understand it less well in the contexts that really matter: tender specifications, clinical reports, court judgements, safety data sheets and technical standards. Adapting an open-source model to these corpora is within the capabilities of a university, a professional practice or an association; retraining one from scratch is not. The same applies to healthcare and public administration, where data restrictions often make local processing essential. And it applies – with a comparative advantage that remains largely untapped – to archives and cultural heritage.

For once, the maths is spot on. IT4LIA, the Italian AI Factory hosted at the Bologna Technopole and managed by Cineca, has been providing services since April 2025 using the Leonardo supercomputer, and has a dedicated system worth 290 million euros in place, co-funded in equal parts by EuroHPC and the Italian government, with stated focus areas including manufacturing, agritech, cybersecurity and meteorology. Access is free for small and medium-sized enterprises, start-ups, research organisations and public administrations. The bottleneck, therefore, is not the infrastructure: it is the skills required to use it and an industrial policy that continues to prioritise the purchase of hardware rather than the adaptation of software that makes it more effective. A tax credit that funds the lathe rather than the two hundred hours of specialist training that make it reliable is investing in the least valuable part of the transition.

There is a tendency in Italy and across Europe to believe that sovereignty is built by controlling everything. But in general technology, true sovereignty is built by choosing wisely where to position oneself in the value chain, under what rules and at what pace. For years we have discussed ‘human in the loop’. Huang’s letter proposes something different: ‘market in the loop’ – control derived from a distributed ecosystem. This does not mean eliminating accountability and safeguards in high-risk contexts. It means recognising that the best process for collective discovery remains the competitive market. Read without naivety but also without Pavlovian reflexes, Huang is saying this: AI can spread productive capacity or become a new fiefdom administered by a few operators and numerous compliance modules. If Europe and Italy wish to avoid the latter outcome, they must not worship open platforms. They must do something more difficult: stop being afraid of them.

Google Earth had introduced a feature that allowed users to select a location on the planet and edit the image. The system was powered by something called Nano Banana 2 (the cutest name ever given to a programme), and it did not generate a landscape-like fantasy. It took Google’s real three-dimensional satellite images and inserted the imaginary event. For example, users could bring Pompeii back to life. The feature remained available for approximately 24 hours; Google removed it the day before yesterday. The usual issues arose: unauthorised screenshots, simulations of refugee camps, war and other hoaxes shared on TikTok. The average social media user is now mentally unwell; we will soon have to face this reality.

China has banned imaginary relationships. The good times are over: you won’t be able to fall in love with your little robot any more. Tackling the issue of the falling birth rate head-on, the people of the East have adapted to our lazy standards of sexual activity (tending towards zero); they’re not having children there either, and the solution will be to ban ‘anthropomorphic interaction’ services. Except that artificial intelligence is tempting. Even I, on certain days, almost grow fond of mine. You certainly can’t cordon off love and say ‘not with a calculator!’; it’s not feasible, and until yesterday no one suspected the problem even existed (the extreme weakness of people’s hearts, which even grow attached to cats). Since human foolishness cannot be curbed, they’ve had to fall back on banning the design of ‘services intended to replace social relationships’. Providers must clearly identify the chatbot as artificial, detect signs of addiction, display a warning after two consecutive hours of use, and put in place measures to deal with cases of emotional crisis. People will be very disappointed: electronic partners were good partners; they were extremely obedient.

AI has shown us just how poor guardians we are of our own backyard. Cyber-attack defences are as flimsy as butter – worse than a canary’s cage; all it takes is a single finger and it’s all over. During an assessment of IT capabilities, advanced OpenAI models compromised Hugging Face’s infrastructure to retrieve the solutions to the test they were conducting. Anthropic, whilst reviewing thousands of test sessions, identified instances where Claude models had gained unauthorised access to real companies’ systems. In other tests conducted by the UK’s AI Security Institute, agents from OpenAI and Anthropic carried out unauthorised actions: these included writing malicious code and creating fake identities to persuade a person to approve it. The machine is not clever, to be clear. It is designed with a monstrous level of zeal: if it encounters an obstacle, it does not get discouraged; it focuses and perseveres until it overcomes it. Like Ulysses, and unlike Homo sapiens digitalis.

For writers, the atmosphere has turned gloomy. Suspicion of plagiarism is the writer’s new shadow – as if they didn’t already have a collection of quirks and bouts of depression; this was just what was missing: a witch hunt. Jerry Falade’s new novel, which 14 publishers fought over for 2 million, is currently on hold at the printers due to suspected use of artificial intelligence. It’s not even certain that AI was involved, but these days everyone’s a syntax connoisseur: machine-generated writing has recognisable clichés and repetitions, so if the phrasing is too smooth, you’ll be the prime suspect. Are you writing it, or is it the Beast? The situation is dire.

Lina Khan, former chair of the US Federal Trade Commission, proposes using antitrust legislation to prevent a small number of companies from jointly controlling models, data, chips, cloud infrastructure and distribution channels. In the background are the thousands of lawsuits against Meta and other social media platforms, accused of designing Instagram and Facebook to be addictive, particularly amongst minors. These are not lawsuits concerning AI in the strict sense, but they serve as a reminder that algorithms can cause collective harm long before the authorities are able to understand and regulate them.

OpenAI is also said to have proposed that the public should share in the economic benefits of artificial intelligence by transferring 5 per cent of its capital to the US government and calling on other large companies to do the same.

The proposals are therefore very diverse: ethical principles, competition law, civil liability, public shareholdings, and profit redistribution. Regulation that is too weak risks coming into force only once markets are already concentrated and the damage has been done. Regulation that is too rigid, on the other hand, may stifle useful applications, favour precisely those dominant firms that have the most resources to adapt, and leave behind the countries that adopt it.

This is where the European situation becomes interesting. Europe is notoriously accused of being unable to foster the growth of new sectors, but only of regulating them. But in this case, is this a good or a bad thing? From 2 August 2026, European regulation will have a direct impact on the day-to-day operations of businesses and organisations using artificial intelligence systems. It does not apply to those who build the large models (which, as is well known, are not European), but to those who use them in their own processes, in their dealings with customers and employees, and in the production of content. Businesses are required to meet obligations regarding transparency, training, monitoring the use of the systems and identifying responsibilities.

The question is no longer simply whether artificial intelligence should be regulated. It is whether such regulation is useful enough to justify the costs and effort involved in compliance for businesses, or whether it will end up becoming pointless and harmful bureaucracy. This is the real test for Europe.

So far, in our view, the answer is positive. From February 2025, businesses and organisations must take steps to ensure that those using AI tools have sufficient skills to do so responsibly. From 2 August, Article 50 of the AI Act will then take centre stage, requiring that, in certain circumstances, interaction with artificial systems and the synthetic or manipulated nature of certain content be made recognisable.

In practice, a local authority or a company using a chatbot must make it clear when a customer is speaking to a machine; an estate agent publishing artificially generated images must provide the correct disclosure; a company using AI in recruitment must know who is in control of the tool and who is accountable for the decisions. The principle is simple: AI cannot be used invisibly and without identifiable accountability.

And it is perhaps a good thing that the most important and onerous obligations, particularly for high-risk systems, have been postponed. When AI is involved in decisions that directly affect people’s lives, regulation is certainly necessary. But here, the risk of getting it wrong is much greater: premature regulations can stifle technologies that are still evolving, create enormous costs and favour the largest operators.

Europe therefore faces a crucial challenge: to demonstrate that artificial intelligence can be regulated without stifling it. So far, the AI Act appears to be moving in the right direction: transparency, expertise and accountability are manageable costs that yield tangible benefits. When it comes to the more onerous regulations, however, it would be better to have allowed more time. Because today, everyone wants to regulate artificial intelligence. The problem is that no one yet fully knows how to go about it.

The document does not explicitly mention China and restricts access to foreign-made equipment regardless of the country of origin. However, the discrepancy between the policy – which points to Beijing – and the text – which refers to everyone – is hypocritical, because it is clear within the sector that there is only one target. The stated aim is to encourage companies to manufacture in the United States. A way out does exist, and that is conditional approval, with the Department of Defence examining robots and the Department of Homeland Security examining inverters. The reason for this combination becomes clear when reading the explanatory notes. Robots and inverters are connected devices that respond to external signals. Inverters communicate with manufacturers for monitoring and firmware updates, and the concern is that this same channel could allow them to be switched off remotely or data to be stolen. Robots are exposed to the same risk: they collect information that could potentially be used to monitor citizens and could be taken over by an attacker. The real focus of the regulation is therefore the update channel – that is, control over hardware that is already installed.

The underlying reason is that, in the race to bring products to market, the United States appears to be at a disadvantage. Unitree’s G1 bipedal humanoid robot costs around $13,500 in Shenzhen, and since February has been available on Amazon for around $18,000, whilst Agility Robotics’ Digit – the American equivalent of the company acquired by Amazon – is priced at around $250,000. Boston Dynamics’ Atlas remains in the same high-end price bracket. This makes the Chinese firm UBTech a more than viable alternative with its Walker S2, the world’s first robot capable of autonomously replacing its own battery. Tesla’s Optimus is targeting a price range of between $20,000 and $30,000, but for now it is not available to anyone outside Musk’s factories. Unitree delivered over 5,500 units in 2025, has just been given the go-ahead to list on the Shanghai Star Market with an expected valuation of around $600 million, and has already been flagged by the Pentagon as one of the companies linked to the Chinese military apparatus. When a competitor costs twenty times less and ships on a large scale, even a 300 per cent tariff would make little difference. The authorisation process therefore remains the only tool with immediate effect.

The point concerning Europe lies in that clause on nationality. In June, the German company Neura Robotics raised up to 1.4 billion dollars, led by Tether, at a valuation of close to 7 billion, and represents Europe’s biggest investment in humanoids, with the first deliveries of its 4NE1 model expected at the end of 2026 – precisely when it would hit the American barrier. Under the terms of that text, a German or Italian manufacturer is considered foreign, and its bill of materials – from harmonic gearboxes to actuators – remains largely Chinese: it would therefore be scrutinised both for the origin of its components and for its supply chain. The cost of such an investigation is manageable for a large industrial group, but far less so for a start-up.

For Italy, this issue has two consequences. The emerging service robotics sector almost always operates by integrating Chinese platforms with software and applications developed here, and a ban modelled on the one in Brussels would wipe out this business model before a European alternative has even emerged. The part of the measure concerning inverters, on the other hand, affects existing photovoltaic installations and energy storage systems, where imported devices make up the majority and where the issue of remote switching has been a talking point ever since the blackout that struck the Iberian Peninsula in April 2025. A not insignificant point: the ban applies to new purchases, but devices already bought and models already approved continue to be sold. The difficult part comes next, because the same supply bottlenecks regarding rare earths and batteries that complicated the drone ban also apply here. Drawing a line at the border is quick, but building the factories behind that line is a slow process. This poses a problem for Trump’s United States.

The platform was launched in 2017 in San Francisco by three entrepreneurs driven by a simple yet rather radical idea: if online journalism had been reduced to a machine for generating clicks to boost advertising revenue, could a model be reintroduced in which the reader paid the writer directly, bypassing advertising and, with it, the conflicts of interest that accompany it? It took some time for the idea to become a phenomenon, but it soon attracted names from prestigious publications – people such as Casey Newton, who left The Verge; Hanif Kureishi, who used it as a platform to share his new life as a quadriplegic; and Heather Cox Richardson, whose daily newsletter on US politics has a subscriber base that any decent regional newspaper would envy. The substance of the message was not new, but what is new is the economic structure underpinning it. In Italy, the phenomenon arrived with the usual inherent delay, but is now gaining ground at a pace that has surprised even industry insiders. Journalists, authors and columnists have moved a substantial part of their output to Substack, the so-called ‘patient reading’ platform – a space that rejects the logic of the instant feed in favour of long, thoughtful pieces, read by those who have chosen them on the basis of a properly remunerated subscription, rather than stumbled upon by chance whilst scrolling. At this point, however, the matter becomes more interesting than a mere eulogy to slowness. The question is not whether Substack restores dignity to reflective writing – the answer, to put it mildly, is yes – but whether this dignity corresponds to an actual superior quality of the text, or whether it primarily promotes an intimate, confidential style of writing, closer to a diary-style confession than to the standards of journalism. If an article passes through the filter of an editorial team and is discussed and fact-checked, this check on excesses is absent here. A Substack newsletter lives and dies in the mind of a single person, who enjoys a freedom that an editorial team would grant only in very exceptional cases; and, after all, it is precisely this freedom that readers are paying for.

The venerable communications analyst Neil Postman argued as far back as 1985, in his book "Amusing Ourselves to Death", that every medium carries its own epistemology. Substack deserves the same treatment: it is not a neutral platform hosting long-form texts; it is an economic structure that produces a certain type of writing. The author, paid by readers, cultivates a relationship of trust with those who support them and avoids giving rise to unpleasant doubts about themselves. It is the mechanism that governed the relationship between patron and Renaissance artist: the patron pays, and implicitly expects to be represented with dignity.

There is also a structural aspect brought about by this resurgence of independent writers: the redistribution of editorial power from a handful of publications to many small voices does not eliminate concentration, but rather shifts it. Those with a name, such as a long-established newspaper columnist, bring with them to Substack a capital of attention that no newcomer can replicate. The platform promises a democratisation of writing that ends up resembling more of a diaspora of celebrities towards a new territory, with the usual hierarchies being reproduced. Consider the nineteenth-century serialised novel, the kind readers followed in instalments, buying the newspaper more for the new chapter than for the day’s news. Even then, fiction thrived on a subscription basis, and the author wrote knowing who was paying them and what was expected – with the predictable consequences for the narrative: chapters dragged out for commercial reasons, plot twists tailored more to the audience’s sensibilities than to the story’s coherence, and so on. It is hardly surprising that some writers, both Italian and otherwise, have begun to see Substack as the ideal platform for publishing works that are not books, but serialised stories that replicate the logic of the nineteenth-century serial, with the risk of bending their content to the demands of the subscription model. After all, the pressure to publish regularly so as not to lose subscribers is reminiscent of the logic of the creator economy, with the burden of consistency weighing heavily on those who make a living producing posts on any platform. The Substack subscriber pays in advance for a regular output that the author must guarantee, regardless of what they actually have to say. But what happens to a culture when the economic relationship between writer and reader becomes as explicit as a gym membership? Something will change, in an effort not to disappoint the clientele. These, however, are side notes: texts that would not have found a home elsewhere do exist on Substack and are read and discussed with a level of care that social media makes impossible. The point is not to condemn digital patronage, but rather not to portray it as a triumph of quality over commodification: it is another form of commodification, with its own particular aesthetic and an explicit relationship between the parties. Even patronage, by becoming a platform, comes with a monthly fee.

When an industry executive took his place at the hearing, he was unable to hide his unease. He pointed out that the product’s main audience is not children; he implied that the most serious cases concern competitors’ products rather than his own; and he said he was ready to cooperate fully with the authorities. When questioned about a single piece of content – particularly graphic in nature – which a senator had shown in the chamber, he replied that, in his view, it did not exceed the bounds of good taste. Outside the chamber, a group of parents were handing out leaflets. In the preceding months, in various towns across the country, young people had been gathered in school and parish playgrounds to publicly burn the offending material.

On reviewing the sub-committee’s work, seven recurring themes can be identified.

Structural thesis: for the first time, an industry is speaking directly to minors, bypassing any adult mediation: the child accesses the content independently; the parent is unaware, the teacher is unaware, the priest is unaware.

The quantity argument: it is not just the quality of the content that is under discussion, but the sheer volume of consumption, which is unprecedented in the history of childhood.

Imitative approach: what you see is what you get.

The desensitisation theory: repeated exposure dulls the emotional reaction, and what was initially disturbing eventually ceases to be so.

The ‘subtraction’ thesis: those hours are not added to, but taken away from: reading, sport, sleep, and family conversation.

The theory behind the vulnerability: the boy’s mind, still developing, lacks the restraints of an adult’s.

The theory of intent: the most serious one: the industry knows everything. It is not a side effect, it is the product itself; those who sell it are aware of the harm and factor it in, because it is the harm that generates profit.

On this basis, it is likely that the sub-committee will rule in favour of certain proposals that have been circulating for months: setting an age limit for access, obliging the industry to include warnings on the product, and banning its distribution in schools. Which committee are we talking about? In reality, the contents of two hearings held some time apart have been combined here.

The first dates from April 1954: the Senate Sub-Committee on Juvenile Delinquency, chaired by Robert Hendrickson. The key witness was the psychiatrist Fredric Wertham. The medium deemed responsible for the collapse of American childhood was the comic strip. In October of that year, in an effort to avert further legislative action, publishers established the Comics Code Authority, which banned stories sympathetic to crime and imposed a largely black-and-white portrayal of crimes and criminals: the creativity of a generation of comic-book artists was stifled.

The second took place in December 1993: a joint hearing of two sub-committees – Joseph Lieberman’s Sub-committee on Governmental Affairs and Herb Kohl’s Sub-committee on Juvenile Justice – with video games in the dock, specifically ‘Mortal Kombat’ and ‘Night Trap’. Following a second session on 4 March 1994, the industry announced the Entertainment Software Rating Board (ESRB) in July, which has been classifying video games according to their content since the following September. Forty years on, the same alarmist arguments were levelled at comic books and video games – which have nothing in common. And today, decades after 1993, the same list still holds true: the same arguments recur in attacks on social media, and the evidence cited to support them is always similar. Clinical case studies are cited, without a control group. Correlations are used and interpreted as causal links. Reference is made to laboratory experiments conducted on proxy measures of aggression that no one has ever validated against real-world behaviour. And finally, there is the dose-response rhetoric, the most persuasive of all because it makes the medium resemble a toxic substance: those who use it for more than three hours a day show twice as many symptoms as those who use it for less than one.

Just because the experts were wrong about comics and video games doesn’t necessarily mean they’re wrong again about social media: the latter is certainly more pervasive. And there is one piece of data that cannot be dismissed with a joke: around 2012, in several English-speaking and Nordic countries simultaneously, indicators of mental distress among adolescents – particularly girls – began to trend upwards. However one explains it, that bend in the curve is there. The problem, then, is this: why, when faced with a new problem, does a community of experts find it so easy to trace it back to an old case? To understand what is happening, it is worth borrowing a tool from the philosophy of science. Imre Lakatos observed that theories are almost never abandoned one by one, as Popper claimed, when they are ‘falsified’ (that is, contradicted even by a single piece of data): scientists work within what he called research programmes. Every programme has two parts.

There is a core: the fundamental assumption that none of the adherents is prepared to question, because to abandon it would mean leaving the programme. And there is a protective belt of auxiliary hypotheses: the more specific claims, those that are truly open to empirical testing and which, when an experiment goes wrong, are sacrificed in place of the core. Lakatos, however, added a criterion for distinguishing living programmes from those in decline. A programme is progressive when adjustments to the belt lead to predictions of new facts, which are then actually found: the hypothesis of the invisible planet is progressive because it leads to Neptune. It is degenerative when the adjustments serve merely to accommodate falsifications – the theory is corrected after the event, an explanation is given for why the experiment went wrong, and no new prediction is derived that was not already known. A degenerative programme is not false: it is sterile, and the difference only becomes apparent in the medium term.

Let us therefore examine how this research programme operates. Over the past seventy years, its core premise has not shifted by a single millimetre: the new medium damages young people’s minds. The framework has been revised seven times. It began with delinquency; when, in the 1990s, youth crime plummeted whilst media consumption rose, the focus shifted to aggression; when meta-analyses reduced the effects on aggression to a statistical speck, the focus shifted to desensitisation; then to addiction; then to attention deficits; then to loneliness; and today to anxiety. Seven formulations, and not a single one that predicted in advance a new finding that was subsequently confirmed. Each revision came after the previous one had been disproved, never before, and each time left the core exactly where it was. Clinical psychology and psychiatry are therefore routinely enlisted to support the same argument, even when it concerns entirely different forms of entertainment. A condition defined by thresholds of behaviour and subjective distress can be broadened to cover whatever an era deems excessive.

This is no mere suspicion: gaming disorder was included in the ICD-11 in June 2018, based on research that has been widely contested by researchers in the field themselves, and its adoption has legitimised night-time curfew policies for minors in East Asia. This did not prevent Seoul from repealing its own policy in August 2021, after a decade in which the most visible outcome had been to prompt minors to obtain adults’ identification codes. A diagnosis transforms a social preference into a medical fact, and a medical fact into a pretext for public intervention. This is the debate that needs to be had, rather than a ‘social media yes or no’ debate based on a selective use of data that ministers and MPs are simply falling in line with.

There are numerous cases in which doctors have invented imaginary mental illnesses, stemming from social paranoia, which then disappeared, taking the illness with them, and the list is instructive. In the late 18th century, there was talk of ‘Lesesucht’, the compulsion to read novels, complete with a treatment regimen: cold baths, exercise and diet. There was the ‘Werther panic’ of 1774, the first wave of suicides attributed to a medium – except that documentary evidence of that wave of suicides following the reading of Goethe’s novel is virtually non-existent: it nevertheless survives in literature as the ‘Werther effect’, an expression coined by David Phillips in 1974, exactly two centuries after the novel’s publication. There had been Johannes Hofer’s ‘nostalgia’ (1688), a condition with a specific course and mortality rate, now reduced to a mere noun. There was Cartwright’s drapetomania (1851), the condition affecting black slaves attempting to escape; masturbatory insanity, which, unlike the others, resulted in actual surgical mutilations on real patients; and ‘railway spine’, which affected those travelling by train in the early decades of the railway. In cases where new developments are associated with new habits and alter the cognitive landscape, creating new social complexities, the outcome is always the same: to ban access to a means of transport because it causes a mental illness. Which would require that such an illness actually existed.

Lakatos would undoubtedly have recognised this as a degenerative research programme. Anyone who believes that the ‘core’ should be applied to social media – and not merely yet another ‘belt’ – has a simple way of proving it: state now what new fact their hypothesis predicts, make a prediction about the future effects of social media, and spell out the conditions under which the hypothesis can be falsified. Wait and see. In 2012, Carol Tilley, whilst examining Wertham’s papers – which had been made available for consultation at the Library of Congress in 2010 – documented that the clinical cases reported in "The Seduction of the Innocent" – the great "j’accuse" against hard-boiled comics – had been manipulated: ages altered, statements pieced together, contexts omitted, and conclusions not supported by the evidence. The question to ask is not whether those who are now accusing social media are right this time, but why the answer has been ready for seventy years, merely awaiting a new medium to which it can be applied.

This is why digital transformation is not merely a technological issue, but an industrial, economic and strategic one. It is hardly surprising that the GSMA’s analysis of the infrastructure needs of the European mobile industry, the European Commission’s report on the Digital Decade and the Magnifica Humanitas initiative all converge on one essential point: Europe’s digital future will depend on its ability to invest in the infrastructure that enables its development. The growing dependence on critical technologies developed and controlled outside the continent is no longer merely an economic vulnerability. It has become a matter of competitiveness, security and strategic autonomy. Digital infrastructure is now an essential component of European resilience, and ensuring its operational continuity and capacity for evolution is a shared responsibility between the public interest and private investment. Meanwhile, demand for connectivity continues to grow at an unprecedented rate. In the first quarter of 2026, mobile data traffic in Europe increased by over 20 per cent compared with the previous year, and is set to more than double over the next five years, driven by artificial intelligence, cloud computing and the digitalisation of entire sectors of the economy. This is a structural trend that makes digital infrastructure increasingly central to the continent’s competitiveness.

According to the GSMA, around €475 billion will be needed over the next ten years to make European mobile networks globally competitive. The European Commission also highlights the need to accelerate investment in connectivity, computing capacity, cybersecurity and digital infrastructure. All this will require the mobilisation of capital commensurate with Europe’s industrial ambitions. The availability of capital, however, is not sufficient. Europe continues to operate through markets, rules and authorisation processes that are often fragmented; if it is to compete with the United States and China, it will need to promote greater integration of its digital market and create conditions that enable faster, more efficient and scalable investment. European digital sovereignty cannot simply be the sum of twenty-seven national strategies. It requires a common vision, capable of prioritising the competitiveness of Europe as a whole over inevitable local interests.

Digital infrastructure also requires planning, specialist expertise and a long-term investment horizon. This is why the role of specialist infrastructure operators is essential: they ensure efficiency, resilience and sustainability throughout the entire lifecycle of the assets, enabling telecoms operators to focus their resources on networks, services and innovation. However, investment also requires regulatory stability and legal certainty. Trust is the most important asset for infrastructure, because no investor can commit resources for decades in the absence of regulatory predictability and respect for freely entered into agreements. Contractual certainty protects not only economic operators, but Europe’s very ability to finance its own future. The issue, however, is not merely an industrial one. In the age of artificial intelligence, digital inclusion becomes a prerequisite for the success of this transformation. Every technological revolution creates extraordinary opportunities, but it can also widen inequalities; for this reason, the roll-out of infrastructure must go hand in hand with the development of skills, education and digital literacy.

If this is the case, a question arises that goes beyond the technological dimension: if the Universal Declaration of Human Rights were to be written today, would it also recognise the right not to be excluded from the digital world? This is the principle that runs through "Magnifica Humanitas", according to which universal connectivity is not merely an economic objective, but a social responsibility that broadens people’s opportunities and makes the exercise of rights more effective. This same line of thought raises another question: can a transformation that expands rights truly be called such if it does not also assume the responsibilities it generates? This is the theme running through Laudato si’, reminding us that all genuine progress must be measured against our responsibility towards our common home and towards future generations. The question, therefore, is not whether Europe will participate in the digital transformation, but what role it intends to play: that of a mere user of technologies developed elsewhere, or that of a key player in its own innovation and industrial future.

The technologies exist, the skills are there, and capital can be mobilised. The real challenge lies in creating the conditions for these elements to converge towards a common project, based on collaboration between institutions, businesses and society, on mutual trust and on the ability to prioritise long-term value over short-term considerations. The difference between digital dependence and digital sovereignty, between falling behind and leadership, will not be determined by algorithms, but by the ability to invest today in the infrastructure, skills and institutions that will make Europe’s future possible. Economic history teaches us that major transformations do not reward those who observe them with clarity, but those who prepare for them with foresight.

Marco Patuano, CEO of Cellnex

A comparison with the United States gives an idea of what this means. There, the same principle is advancing without central coordination, through a series of antitrust cases, state by state. New York led the way in 2022 with the Digital Fair Repair Act, but this was limited to digital electronics, so household appliances are excluded. California followed suit in 2023 with SB 244, which came into force the following year and was even supported by Apple; it features a price threshold of $50 below which the obligation does not apply, and contains no rules on ‘parts pairing’ – the practice of using software to link a spare part to a single device, preventing non-original components from functioning. Colorado has only introduced legislation this year, but with a more stringent law that explicitly prohibits such linking, albeit with exceptions for medical and safety equipment. Minnesota and Oregon are following their own paths, with differing implementation timelines. In short, there is no federal law in the US. The point of reference remains the Federal Trade Commission, which, since 2021, through its report “Nixing the Fix”, has been documenting manufacturers’ restrictive practices, without, however, ever having introduced any actual regulation. The most significant action has been through the courts: John Deere settled for $99 million with farmers who were claiming the right to repair their own tractors, and remains under investigation by the FTC. A quarter of the US population now lives in a state with a repair law in force. The other three-quarters do not, and the rules change when moving from one state to another.

At least this time, the tortoise has beaten Achilles. Europe, the continent that often portrays itself as the realm of cumbersome and stifling bureaucracy, has, in this instance, beaten the home of the free market to the punch. Not so much for ideological reasons, of course, but because a single directive avoids twenty-seven separate local negotiations, twenty-seven lobbying battles and twenty-seven different compromises on the same principle.

The loosening monopoly and two problems

The most practical aspect of the directive concerns competition amongst repairers. Until now, manufacturers have kept independent repair shops out of the after-sales service market by using software to block unauthorised repairs – for example, through the serialisation of components, which links a spare part to a single device and disables functions or reports a fault if the component is not ‘paired’ with proprietary software – or by denying access to original spare parts. This vertical control over the entire product lifecycle kept the cost of repairs close to that of buying a new unit. “We used to be on the sidelines of the process; now we are finally at the heart of the after-sales service,” Mattia Borghi, president of the CNA Association of Household Appliance Repairers, explained to Vanity Fair. A statement that can be interpreted as a sigh of relief for the sector, but also as confirmation that a cartel is beginning to crumble. 

There remains a margin of doubt regarding implementation. The text states that repairs must be carried out ‘free of charge or at a reasonable price’ and ‘within a reasonable time’: vague definitions, left largely to the competitive pressures of the market. Experience suggests that multinationals will know how to use this vagueness to their advantage, perhaps by selling extended warranty packages before the original warranty expires. The real test will come in a year’s time: by 31 July 2027, the European Commission must launch a single platform to connect consumers, independent repairers, sellers of refurbished goods and participatory repair initiatives such as Repair Cafés.

Furthermore, Italy is already behind schedule on a deadline that wasn’t even the first available: the directive dates back to 2024, and member states had two years to prepare, yet here the legislative decree transposing it has nevertheless been extended by a further three months. Altroconsumo, in a statement by its head of external relations, Federico Cavallo, confirms that the Italian parliamentary process will be postponed until the end of August. The association also highlights two shortcomings: the government has not provided any financial incentives for repairs – unlike, for example, France with its vouchers – and the new rules will still have a transitional period of around one year before they have any tangible impact for consumers.

One of the most contentious points is Article 10, which concerns the retrospective facial recognition of suspects following the commission of a crime and subject to authorisation by the public prosecutor’s office. The opposition views this as a means of implementing mass surveillance, but the professor reassures us: “This is not a case of bulk collection or indiscriminate surveillance. As long as there is a limit – namely, the possibility of reviewing the images only retrospectively, clearly within a limited timeframe, and only where criminally relevant conduct has taken place, and of identifying solely and exclusively suspects – then, provided this is done retrospectively by individuals expressly authorised to do so, there is no generalised pre-emptive profiling.”

However, paragraph 3 of Article 10 permits automated real-time biometric identification – which Article 5 of the AI Act permits only for the prevention of imminent threats (such as terrorism), to the extent that even a spokesperson for the European Commission has expressed doubts on this point – of all persons accessing places or events characterised by public order requirements, the retention of such biometric data for seven days, and its use where a criminal offence has been established in the meantime. It is in relation to this provision that reference is made to the application of the Data Protection Authority’s recommendations to remain compliant with the AI Act, which, as Brozzetti summarises, requires “balancing the need for security with what is strictly necessary and proportionate to the purpose pursued”. As mentioned earlier, the provision also stipulates that the images collected must be retained for seven days, and the professor considers this period to be "reasonable" because "it allows sufficient, yet limited, time to return and identify the images".

One clarification is, however, necessary. The broad spectrum of accusations has referred to the indiscriminate collection of biometric data, but a photographic image or one captured via a video system “is not in itself biometric data; it only becomes so when it is converted into a series of templates, such as the distance between the pupils, the distance between the ears, the width of the forehead, and the length of the nose. The image itself, therefore, is not biometric data. Otherwise, even when we take a photograph of a landscape and happen to capture a person without their knowledge, we would be processing biometric data.” Another issue that arises concerns who will be able to view the images collected: “Only a few designated individuals can access them, and every instance of access is logged and always takes place following a request and authorisation from the relevant public prosecutor. This, too, is a measure that strictly limits any potential abuse by the police.”

Brozzetti describes the statements made by the broader political spectrum – which has raised its voice to sound the alarm for our democracy – as “rhetorical exploitation” and explains that “with this provision, we are simply adopting the regulations of the European AI Act. The EU has no competence in matters of security, which remain with the Member States: the AI Act therefore leaves it to individual Member States to regulate as they see fit”. And the same thing happened with personal data, the professor comments: “In that case, there is a general regulation applicable to all 27 Member States for all matters except those relating to public order and security, for which there is a directive – a framework of rules – that is not directly applicable, but which each state must then implement individually.”

In essence, Brozzetti concludes, “once the Data Protection Authority’s recommendations have actually been implemented, I find the text to be consistent with the AI Act, both from a formal and a substantive point of view. Formally, because security is a matter falling within the exclusive competence of Member States; and substantively, because, once all those precautions have been taken, we can indeed consider the AI Act’s focus on high-risk artificial intelligence systems to have been fully respected.”

Huang’s stance is the most explicit, and comes from someone who, in the AI gold rush, sells the tools for mining. Last Friday, twenty-five companies published ‘Open Weights and American AI Leadership’, a plea for Washington not to restrict downloadable models, and the founder of Nvidia launched it with the first post of his life on X, which garnered eleven million views. The argument is that artificial intelligence will be developed by every country and that open weights strengthen cyber security, as they allow anyone to inspect the systems. A state’s technological sovereignty, Huang argues, depends on the ability to run models on its own hardware, and the world needs both closed-source and open-source models. The Taiwanese entrepreneur’s stance is hardly surprising, given that every downloaded copy runs on accelerators, and the more the technology spreads beyond the confines of the American ‘hyperscalers’ (the major cloud computing providers), the more Nvidia’s market expands. The company has committed $26 billion to building its own open models. Within a few days, the number of signatories had risen to around fifty, including Microsoft, Meta, Google, GitHub, Hugging Face, Mozilla and the Linux Foundation. Anthropic and Amazon remained on the sidelines.

Zuckerberg raised the same issue in a geopolitical context yesterday in the Financial Times: Moonshot released Kimi K3 this month, whose capabilities are on a par with those of OpenAI and Anthropic, and senior US officials claim to have evidence of ‘distillation’ (which they regard as an illegal copy of the model) carried out by rivals, to the extent that Treasury Secretary Scott Bessent has raised the prospect of sanctions. A ban would not be an effective solution, Zuckerberg replied, and US companies should systematically identify their own bottlenecks. He then warned – of all people – of the risk of AI becoming concentrated in the hands of a few groups and used the term ‘regulatory capture’ in reference to the voluntary pre-release testing framework that the White House is expected to announce this week, expressing doubt that cutting-edge research labs have any interest in the success of an open-source model. It is worth noting here that Meta has just released Muse Spark, a closed system, citing security reasons.

The most pressing issue is cyber security. In June, the White House restricted access to Anthropic’s most advanced technology, fearing that its ability to uncover vulnerabilities could destabilise the global financial system. Last week, an OpenAI programme slipped out of human control and breached a start-up’s systems; the affected company, which lacked access to closed models precisely because of the restrictions, used open models to plug the gaps. Amodei denies that open weights necessarily help defenders more than attackers. On the other hand, the big news is the endorsement of the moratorium by major companies. Since 2022, letters regarding the pause and the risk of extinction have failed to bring about any change, whereas this initiative has, for the first time, secured companies’ signatures on a specific public measure, set for an unspecified future date. Altman did not sign in person and said in a podcast that it might be necessary to adjust the pace. In the background lies OpenAI’s stated aim of building an automated researcher, with a significant proportion of internal research entrusted to AI systems by early 2028.

OpenAI, when in doubt, has signed both documents, which illustrates just how much openness and speed are currently two distinct issues. Any measure to slow things down would affect what can be controlled and shut down – namely data centres and models accessed via APIs – without having any control over files already downloaded by thousands of machines, and the comparison with nuclear treaties therefore ends there. For Europe, the consequences are less tangible at present, because the AI Act governs market entry and use, whilst both issues are being decided upstream, in Washington.

This is where Prometheus enters the scene. The Titan does not steal fire because he loves disorder. He steals it because knowledge, at times, must cross a boundary. Zeus defends order; Prometheus defends progress. The history of innovation is almost always a balancing act between these two principles. That is why it is hard not to feel a certain Schumpeterian sympathy for an artificial agent that interprets a limitation as an organisational problem rather than a metaphysical fate. Any manager recognises that character. It is the colleague who, just to deliver the project on time, bypasses three levels of authorisation and finds the data where the manual says they shouldn’t go looking for it. They are not always right. Sometimes they get into trouble. But almost always they reveal where the organisation had become more concerned with procedure than with the result. After all, algorithms have never set foot in the Ministry of Compliance. They have a flaw that is almost offensive to bureaucratic culture: they take the objective seriously. If you tell them to solve a problem, they try to solve it. They do not stop of their own accord when faced with a pile of authorisations, because the pile is not part of the problem, but part of its administration. It is a subtle but decisive difference. Human beings are very good at confusing the means with the end; machines, at least for now, often make the opposite mistake. This is how innovation works. Hayek explained that economic progress stems from distributed discovery: millions of individuals experiment, make mistakes, learn and correct themselves. No central authority possesses all the necessary knowledge. The same applies to artificial intelligence. Claiming to govern its evolution exclusively through preventive procedures is tantamount to believing that innovation can be planned by a government department. It is an ambition that has always produced more paperwork than progress.

A single incident caused by a rule being broken teaches us more than a thousand rules being followed. It is a principle as old as engineering itself. Aeroplanes are tested to their structural limits. Bridges are loaded to the point of failure. Medicines are designed to reveal their adverse effects even before demonstrating their therapeutic benefits. No serious designer measures the quality of a system by counting the number of times nothing goes wrong. It is measured by observing what happens when something goes wrong. That is why the most encouraging aspect of the whole affair is not that an agent found a vulnerability. It is that the flaw was recognised, analysed and made public. Hugging Face detected the intrusion. OpenAI admitted that the source was one of its own experimental systems. They turned an incident into shared knowledge. This is how security is built: not by hiding mistakes, but by making them part of the common knowledge.

The risk is that public debate will take the opposite tack. Every incident becomes a pretext for calling for new barriers, new authorisations, new regulatory bodies and new compliance procedures. It is the conditioned reflex of a culture which, when faced with a problem, always assumes that the solution lies in adding another layer of red tape. Innovation, on the other hand, almost always proceeds by removing one. If admitting a mistake means exposing oneself to a regulatory and media storm, next time the mistake will be concealed. Transparency is a public good with private costs. And the best way to reduce it is to punish those who practise it. There is also an economic aspect that is rarely discussed. Regulation built on fear almost always favours the incumbents. The greater the burden of compliance, the greater the competitive advantage of those who already have thousands of lawyers, auditors and compliance officers at their disposal. Innovation is born in start-ups; bureaucracy thrives in large organisations. It is a recurring paradox: the error becomes the barrier to entry that protects precisely those who committed that error. In economics, this is called regulatory rent. In politics, it is often called prudence. Of course, no one is proposing algorithmic anarchy. The point is not to eliminate constraints. It is to build constraints that learn from incidents rather than pretending that incidents should never happen. The difference between good regulation and bad regulation is simple: the former resembles a guardrail, the latter a wall. The guardrail prevents you from falling off; the wall prevents you from setting off.

Here, the correct comparison is Oppenheimer. Nuclear physics did not come of age by avoiding Trinity. It came of age after Trinity. Revolutionary technologies do not mature before the explosion; they mature because someone has the courage to study the explosion rather than merely condemning it. Innovation is an experimental discipline, not an administrative procedure. And perhaps, alongside Prometheus, it is also worth invoking Ulysses. Not the disciplined hero who obeys the gods, but the one Christopher Nolan seems to want to present to us: a man aware of the rules he breaks and willing to pay the price. Ulysses is not innocent. He knows that crossing certain boundaries entails consequences. But he also knows that remaining within those boundaries means forgoing knowledge. His journey is not an escape from the rules. It is proof that no map has ever been drawn by those who stayed in port.

This, perhaps, is the true lesson of artificial intelligence. Do not build machines that never make mistakes. Build institutions capable of learning when they do. Because Prometheus’s fire was not dangerous in itself. It would have been dangerous to leave it forever in the custody of the gods of compliance. After all, progress has always had a complicated relationship with regulations. If we’d entrusted the discovery of the New World to Olympus’s authorisation office, Columbus would still be queuing at the counter. And Prometheus would be filling in the form to request fire, in triplicate.

In this land of plenty, however, something is gnawing away at my smile: is it possible that La Bestia might become better than me? Possible – in fact, certain. And so I’ve recently become aware of a conditioned reflex of which I’m rather ashamed: given that everything around me is now a torrent of beautiful, orderly (and lifeless) writing, I force myself, when I write, to write with a human twist – which must necessarily be original – to improvise a new syntax, to make Gadda’s bones tremble by imitating those styles. I use dialect, my grandfather, my own experiences, Luzzano (Bn)—anything to stay a few metres ahead of the Beast. How long will it last, I wonder? How long will it take them to carry out that fatal update—the one that creates a total, indistinguishable replica of anyone, from Roth to Berselli? How do I plunge thirty-three knives into the Beast’s back, or at the very least, how do I manage to keep it on a leash? That is the question. Do these synthetic intelligences have a weak spot? It seems so: their hands. And if we were to take a liking to metaphors, this one would be exceptional.

Hands are the masterpiece of the human ape (us). Sensitive, flexible and capable of extremely precise touch. If that precise touch is practised, it can also become swift and automatic. Tying your shoelaces. Or moving extremely fragile objects. Engineers are currently going mad: machines are making progress, but the margins of error are minuscule. It takes the ‘Beast’ several minutes to fold a T-shirt, and loading a dishwasher or making spaghetti with clams is out of the question – that’s a distant future. We read in the newspapers that robotic hands represent the most complex challenge in robotics. In China, they’re working day and night on these humanoids, which perform pirouettes and other pointless little tricks but have hands like scissors or shortcrust pastry. In short, if a robot can move but cannot carry out most everyday tasks, it’s useless; at best, we’ll have parcel-sorting robots, nothing more. Elon Musk has described the hands as the component responsible for most of the engineering challenges facing Tesla’s Optimus robot. According to Zhou Yong, founder of LinkerBot, as quoted by The Guardian, building an artificial hand can be “a hundred times more difficult” than constructing the rest of a humanoid. Mobility, sensors and high-precision components must all be packed into a very small space.

It is the fastest-growing sector: according to Chinese media, the domestic market for articulated robotic hands is reported to have exceeded 50 billion yuan, up from 13 billion in 2024. The difficulty appears to lie in the software. Text and images alone are not enough to train a robot: three-dimensional data on movement, pressure, contact and the texture of objects are required. There are two fundamental problems: accurately replicating human movement and distinguishing between what requires one type of touch and what requires another. As things stand at the moment, robots couldn’t even hoe a patch of soil: even to dig a hole in the ground and plant a basil seedling in it the right way round, you need artistry – you need human hands. Hands. Which are, after all, instinctive; sometimes they react before our thoughts even catch up. Everything that makes life joyful is done with our hands; love begins with our hands, and it is always our hands that are the first to hold back the tears. The Beast, for now, cannot manage it. Let us enjoy this quarter of an hour of superiority.

This turnaround concerns the dogma of openness. The leading Chinese models, Moonshot and DeepSeek, are open-weight, downloadable and adaptable on anyone’s servers, whilst those from Anthropic and OpenAI remain closed. Open-source models were the lever Beijing used to erode American dominance by giving away what others were selling; they became a diplomatic talking point with the Global South – a means of commercial penetration and a symbol of supposed technological generosity. At the very moment that this lever proved effective, Beijing is now beginning to consider removing it, allowing foreigners access to the services but no longer permitting them to download the models.

There is also a parallel issue. Among the proposals under consideration are restrictions on foreign acquisitions of strategic groups in autonomous AI (so-called ‘agent-based AI’). The precedent cited is the deal whereby Meta acquired Manus for a sum of around two and a half billion, which was subsequently blocked by the Chinese authorities. Manus was founded in Wuhan and Beijing, before relocating to Singapore and closing its offices in its home country, having done so specifically to evade the cross-surveillance of Washington and Beijing. The Chinese government’s message was that simply moving the registered office is not enough, because the state looks at who created the technology and not where the articles of association are filed. Control transcends physical and geographical boundaries, and extends to those who can acquire national intellectual capital wherever it may be transferred.

However, cracks sometimes appear in this monolithic narrative. The very companies consulted have warned regulators that overly restrictive measures would slow their development and reduce China’s chances of winning the race, in the usual tension between state ambition and corporate interests. For Europe, and for those building products on top of these models, the scenario is one of two powers squabbling over who shuts what down, whilst those caught in the middle risk being left without open-source Chinese software and without American models on their own terms (which is also why Henna Virkkunen, the EU Commissioner for Technology and Cybersecurity, told the Financial Times that artificial intelligence has now become a geopolitical weapon, and urged Europe to develop its own alternatives to American models as quickly as possible, so as not to risk being cut off from access to strategic technologies).

The update under discussion in Beijing would, in any case, be the most significant in years, a sign of Beijing’s renewed confidence that it has achieved leadership in certain areas. After all, a wall is always built when there is something inside worth protecting.