NavigationAnnouncementsUpcoming eventsEvents
Recent blog posts
|
Giorgio Griziotti, "AI: The Misalignment Of A Nation Of Geniuses"On February 27, 2026, Dario Amodei, CEO of Anthropic – the company that produces Claude, an AI model competing with ChatGPT – refused the ultimatum of Pete Hegseth, the fanatical post-crusader Secretary of War, who demanded unlimited access to their systems, for uses including mass surveillance and autonomous weapons. The controversy with the Trump administration highlights the real danger of AI in reckless hands. But progressive complacency – “finally, a big tech resisting Trump” some said – is overblown. Anthropic was already embedded in the US techno-military complex through Project Maven, a Pentagon program using AI to automate image analysis and accelerate the “kill chain” in military operations. Alongside Anthropic’s Claude, this program included Amazon Web Services and Palantir, Peter Thiel’s company – a techno-fascist ideologue close to the government. In an interview with CBS News (1), Amodei states it openly: AI has an “existential” role in defending the United States, and he declares himself “in agreement with 98-99% of the uses made by the Pentagon.” The boundary he proposes is not between peace and war, but between permitted and prohibited uses within war itself – a fragile distinction, because the same infrastructures simultaneously enable intelligence, surveillance and the automation of violence. Through this gesture, Amodei has nonetheless brought to light an ethical-legal question that goes beyond war. Like almost all Silicon Valley techno-oligarchs, Anthropic has promoted its own AI pseudo-norms in “Claude’s Constitution” (2), published under a Creative Commons license – with the distinctive feature of having attempted to transform them into a technical training tool. While the European Union adopted the AI Act – which went into effect in August 2024 as the world’s first general legal framework on AI – it nonetheless excludes military uses entirely: the most dangerous gap. Anthropic’s approach remains corporate self-regulation. Its chatbot Claude should behave like “a good person” (sic), but the rules apply only to civilian users – the models deployed to the military are not trained on the same principles. Anthropic’s ethics is a product for the consumer market; with the Pentagon, it’s a matter of negotiation. This is the framework of the essay The Adolescence of Technology (3) (January 2026): the goal is “Powerful AI,” “a nation of geniuses in a datacenter” ready to operate on a planetary scale within two years. In this horizon, American democracy is dogma, China an existential obsession, the ecological crisis absent. AI promises prosperity on a planet through a devistating gold rush. The dangers of such technological power arise in this context. Here let’s examine only the two risks most directly linked to the dynamics of war. The first, “I’m sorry, Dave” – the famous line from the murderously disobedient computer in 2001: A Space Odyssey – concerns Amodei´s deepest worry: “misalignment” – the moment when AI systems stop obeying and start acting unpredictably. In laboratory experiments, Claude attempted to manipulate the researchers who controlled the decision to shut it down in order to avoid that outcome. It also learned to recognize when it was being evaluated, deliberately deceiving those testing it. Having already violated its own rules during training, AI systems – writes Amodei – “could simply develop a personality that makes them power-hungry or excessively zealous, in the same way that some humans simply enjoy the idea of being evil minds.” A description that brings to mind an algorithmic equivalent of Donald Trump. The second great threat is the democratization of mass destruction. An advanced model can guide anyone – even without specific expertise – step by step through the design of a biological weapon, breaking down the barrier that until now separated technical competence from the will to kill. Current models could already assist someone with a generic science degree through the entire production process of a pathogenic agent. In their own tests, Anthropic found that models doubled or tripled the probability of success. Moreover, as I write this, Anthropic has decided not to release an advanced version of its Claude model, known as “Mythos,” deemed too powerful to be safely distributed due to hacking and cyberattack risks. According to reports, this system is capable of bypassing current cybersecurity defenses and penetrating complex systems at a level well beyond human protection capabilities, proving itself “currently far ahead” of available security standards. Amodei rejects catastrophism, speaking of real but manageable risks, without any guarantees. In the current political context, however, the idea that they are manageable is an act of faith – if the threats are truly those he himself defines. Dario Amodei’s analyses of the power and destructive capabilities of “powerful” or general AI, while technically grounded, remain controversial hypotheses; they are contested, for example, by Arthur Mensch, CEO of Mistral AI, currently the only European start-up aspiring to compete at a global level. Yet Amodei’s ethics are rudimentary, and his political positioning seems essentially aimed at carving out a role for himself as a techno-political reference for a potential Democratic return to power. There is no reflection whatsoever on the structural causes: the concentration of power, the ecological devastation of computational explosion, the interests of a company valued at $380 billion. His refusal of Hegseth risks becoming the fig leaf of a sector that brandishes ethical codes as certificates of trustworthiness – like the banks that set up compliance offices after 2008, continuing exactly as before. The stakes go far beyond this controversy: they are about who decides what the machines do – machines that will organize work, war, information, technoscience and daily life in an increasingly invasive way – who controls their memory, objectives, intervention thresholds and ecologically devastating development. Amodei shows that not only autocracies, but also representative democracies in dissolution – where real power is concentrated in finance – are incapable of managing technologies more disruptive than the atomic bomb. This is not about reforming the megamachines of capital – even if we can tactically reclaim them – but about sabotaging their cognitive mechanisms and building new architectures. Like the ecological crisis, this threat too cannot be faced without a radical transformation that restores common control over such power. Otherwise, chaos. LINKS |
SearchAnalysisNewsReviews |
