The claim that there is more than a 10% chance of artificial intelligence (AI) exterminating all of humanity was made by Evan Hubinger, Director of Alignment at Anthropic, on September 8th, and quickly gained traction on social media. However, Dr. Álvaro Machado Dias, a neuroscientist, futurist, and professor at Unifesp, dismissed this idea as mere narrative, stating that 'there is no slightest plausibility in this talk' during an interview with Olhar Digital.
Hubinger's publication referred to a previous tweet by Jacob Coxon, who had worked at both OpenAI and Anthropic, focusing on the pre-training of AI models. Coxon had issued a warning shortly after leaving Anthropic, suggesting that both companies were 'racing towards a superintelligence capable of self-improvement and betting our lives.'
Hubinger reinforced his belief, declaring: 'We believe, very seriously, that AI could kill humans! I personally think the probability is over 10% in the next decade.' He added that although he believes Anthropic is doing its best, there is still no clear plan to solve the problem of superintelligence alignment.
Arthur Igreja, a technology and innovation specialist, expressed skepticism about the methodology, questioning how Hubinger arrived at the 10% figure and requesting the premise of the calculation. Igreja added that in a risk table, he is more concerned about mistaken decisions made by people until the end of the decade.
Álvaro Machado Dias reiterated that the idea of AI revolting, acquiring intentionality, or attacking humans is purely narrative. He explained that this would require a type of intentionality present in human beings and other species, something not observed in machines. Furthermore, he pointed out the absence of a visible execution capability for such a scenario.
The neuroscientist stressed that AI requires access to extremely sensitive resources, such as laboratories for synthesizing biological weapons, areas that are highly regulated. However, he acknowledged that AI can cause great damage through digital interfaces, citing the example of an AI designed to create polymorphic computer viruses or hijack sensitive hospital databases.
Machado Dias made a crucial distinction: it is necessary to differentiate what is feasible in the domain of 'how' (practical reality) from what, upon analyzing the 'how,' simply makes no sense. Concurrently, Dario Amodei, CEO of Anthropic, published an article on September 12th admitting that the accelerated pace of AI development, due to 'recursive self-improvement,' requires a coordinated slowdown among industry companies.
This proposal was endorsed by Sam Altman (OpenAI), Elon Musk (xAI), and Demis Hassabis (former CEO of Google DeepMind). Hassabis supported Amodei's essay, while Musk agreed with Amodei's statement. Altman clarified that slowing down does not mean stopping, but rather investing in security audits and tests to ensure that model capabilities do not exceed their alignment.
In contrast, Donald Trump disregarded the warnings from Amodei, Altman, and Musk, minimizing AI risks and defending the maintenance of US leadership over China. China reacted to Amodei's article through the state newspaper Global Times, alleging that the initiative aimed to create a 'Cold War manual' to contain Chinese advancement and protect American technological hegemony.
Machado Dias highlighted that halting AI progress does not depend on government authorization; it merely requires ceasing development. He observed that the hesitation to pause lies in the fact that none of the major companies (Anthropic, OpenAI, Google, xAI) seem willing to adopt this stance unilaterally. He concluded that, strategically, adopting a cautious stance is advantageous because it transfers a potential corporate risk to the governmental and social sphere.
Regarding the positions of the US and Chinese governments, the specialist noted that although Americans initiated development (with ChatGPT) and invest more, the Chinese stance of demanding discussion on deceleration is legitimate. However, he warned that both countries seek to expand their technology globally, and ultimately, 'nobody wants a stop.'