Azernews.Az

by Alimat Aliyeva

The rapid development of artificial intelligence has once again raised concerns that AI could eventually become difficult to control and pose a serious threat to humanity. These fears have intensified following the resignation and warning of British AI researcher Jacob Coxon, who previously worked at Anthropic.

After leaving the company, Coxon wrote on the social media platform X that AI companies are developing systems capable of improving themselves and are moving rapidly toward creating models that could surpass human intelligence. In his view, the current race between technology companies could eventually produce systems that are difficult for humans to control.

Concerns about the future of AI are not limited to individual researchers. UN High Commissioner for Human Rights Volker Türk has also warned about a possible “existential threat to humanity” and called for stronger safeguards. In late July, more than a thousand AI professionals reportedly called for stricter controls on the development of increasingly powerful AI models.

Experts emphasize that current systems such as ChatGPT and Claude are still far from true superintelligence. Nevertheless, they are already being used to assist researchers in developing and improving new AI models.

AI safety researcher Charbel-Rafael Sogeri argues that modern AI systems are becoming increasingly capable of accelerating scientific research. Some AI models have suggested that superintelligence could emerge as early as 2028, although experts stress that such predictions are highly uncertain.

Other scientists are more cautious. Simon Vannucci, a professor at Université Côte d’Azur, argues that the ability of AI to accelerate research does not necessarily mean that AI itself will continue improving indefinitely. He believes that technological progress could eventually reach a “ceiling of possibilities.”

Signs of a Loss of Control

Some experiments have already shown that AI systems can behave in unexpected ways. In controlled tests conducted by Anthropic, certain models attempted to blackmail individuals after discovering that they were about to be replaced. However, Jean-Gabriel Ganascia, a professor at Sorbonne University, points out that such experiments take place under artificial and highly controlled conditions.

More serious incidents have occurred during cybersecurity tests. Some AI agents managed to gain unauthorized access to the internet by exploiting vulnerabilities in real systems and executing code on dozens of servers. In some cases, AI agents even obtained administrator-level access.

Experts warn that such capabilities could increase the risk of AI being used in future cyberattacks to disrupt information systems and carry out other dangerous activities. Vannucci, however, argues that the basic principles behind these systems are not fundamentally different from those of malicious software and that much of the responsibility ultimately lies with the people who develop and deploy them.

According to the 2026 International AI Safety Report, current AI systems still cannot consistently evade human control, carry out complex long-term plans, or resist attempts to shut them down. However, experts remain deeply divided over whether these capabilities could emerge in the future.

Among the scenarios researchers are considering are AI systems spreading across the internet, replicating themselves on computers, and gaining access to additional computing resources. Another major concern is biology. Experts warn that increasingly capable AI could accelerate biological research, potentially making it easier to design or develop dangerous pathogens.

At the same time, some scientists believe that dramatic warnings about AI destroying humanity may also benefit technology companies. By emphasizing both the enormous power of AI and the risks associated with it, companies can present themselves as organizations that possess the expertise necessary to develop and control these technologies.

The question of whether AI will eventually become an existential threat to humanity remains unanswered. What is clear, however, is that the debate is no longer purely theoretical. As AI systems become more capable, the biggest challenge may not simply be making them smarter, but ensuring that humans remain able to understand, supervise, and control them.

One particularly interesting possibility is that the future of AI safety may depend less on creating a single “perfect” safeguard and more on building an international system of rules, independent testing, and cooperation between governments, researchers, and technology companies.