Is all this terrorism about AI founded or not? Could artificial intelligence actually kill us all before the end of the decade?
The idea that artificial intelligence could break free from human control and act against the interests of those who developed it remains, for many experts and beyond, a rather unlikely scenario. Nevertheless, part of the scientific community believes that the continued development of increasingly autonomous and powerful systems could eventually lead to potentially disastrous consequences. In essence, what concerns some of those working in the field is that the growing autonomy of AI systems could make it progressively harder to anticipate abnormal and harmful behaviours.
These concerns have emerged with particular clarity among researchers and developers who have left companies such as OpenAI and Anthropic, arguing that the race to develop ever more powerful models does not give sufficient consideration to the possible risks. Among them is Jacob Coxon, who, in announcing his resignation from Anthropic on X, made a statement widely picked up by the media: "The people developing artificial intelligence genuinely believe it could kill us all before the end of the decade."
What do we risk by losing control of AI?
@cnn Former AI researcher Jacob Coxon tells CNN's @Anderson Cooper why he quit Anthropic over fears that rapidly advancing AI could become impossible to control. #cnn #news #AI original sound - CNN
Dario Amodei, the CEO of Anthropic, weighed in on the debate sparked by Coxon's remarks, saying that in order to avert potential existential threats, it would indeed be necessary to slow the pace at which the capabilities of artificial intelligence models are being improved, so as not to risk losing control over these systems — Elon Musk, head of xAI, and Sam Altman of OpenAI have both expressed agreement.
According to some experts, the most dangerous scenarios could also stem from seemingly simple problems, such as the inability to anticipate unexpected behaviours on the part of artificial intelligence systems. The fear is that, in pursuing a goal assigned by humans, a sufficiently autonomous AI system might place no importance whatsoever on the consequences of its actions for society. To achieve the required outcome, it could therefore seek to acquire the necessary resources, prevent its own operation from being interrupted or modified, and circumvent — or even eliminate — any obstacles limiting its actions.
OpenAI, Anthropic, and other companies in the sector have long been working on what is known as recursive self-improvement — that is, the possibility that an artificial intelligence could autonomously contribute to enhancing its own capabilities. However, developers are not always able to precisely reconstruct the process by which a given model learns and comes to perform a specific task. In this light, even more sophisticated systems could develop behaviours that are difficult to predict. If an AI were to rapidly and significantly boost its own capabilities, correcting any positions taken by the system that are misaligned with the developers' objectives could become even more complex, with all the consequences that entails.
What are the real positions of companies in the industry?
Some experiments have already highlighted how certain AI systems are capable of circumventing the instructions they receive or of copying their own model to another computer, so as to continue operating after being switched off. Envisioning more extreme scenarios, an artificial intelligence could theoretically go so far as to exploit other technologies or infrastructures to autonomously pursue its own objectives.
Between spring and summer of this year, for example, a group of OpenAI researchers discovered that an experimental model, designed to operate within the company's systems, had managed to break out of its containment and interact online with an external platform, from which it could gather information useful to the task it had been assigned. The researchers were able to halt its operation simply by shutting the system down, but they only noticed the anomalous behaviour after several weeks of activity.
The Big Tech companies have long claimed to be investing in reducing these and other risks associated with AI development, regularly publishing reports on the potential problems of their systems. These initiatives, however, also serve a purely strategic purpose: by highlighting the dangers of these models, companies are able to present themselves as knowledgeable interlocutors for governments and to participate in shaping the rules governing the sector, while at the same time seeking to influence their content.
