Scientists warn again about artificial intelligence

No, such a scenario is not entirely new: back in 1991, a Hollywood film hit the cinemas, which already paints a very dark picture in the opening sequence: “Terminator 2”. The film begins in a distant future ravaged by war. Killer robots and spaceships hunt down the last humans still fighting for survival. These machines are controlled by a military artificial intelligence that has gotten out of control, called “Skynet”. The “distant future” in which this dystopian vision takes place is the year 2029.

If warnings from a few AI developers are to be believed, a similar scenario could come true – even within the time frame set by the art of cinema. “Those developing AI seriously believe that it could destroy us all by the end of the decade,” AI researcher Jacob Coxan announced this week on X. How exactly that could happen, he did not say. However: “No other human activity carries such a potential for danger.”

Koksan had previously worked at Antropik, an American artificial intelligence company funded by Google and Amazon, among others. He resigned in protest of what he saw as an irresponsible race among AI companies to develop ever-better autonomous superintelligence. He warned that algorithms could start to independently analyze and improve their own source code, reducing development cycles from months to minutes and entering the realm of uncontrollability.

Koksan is not alone in his fears. Physicist and Nobel laureate Geoffrey Hinton, considered the “godfather of artificial intelligence,” has been warning for years about the serious consequences it could bring. In 2023, he resigned from Google to publicly warn about the dangers of his own development. Turing Award winner Joshua Bengio is also concerned about current developments. An artificial intelligence pioneer from the University of Montreal in Canada is once again calling for strict government regulation of the development of artificial intelligence. Not long ago, in July 2026, more than 1.000 employees of leading artificial intelligence companies warned in an open letter of an imminent loss of control unless the race to optimize “superintelligence” is significantly better regulated.

However, under Donald Trump, the opposite has happened: the US has clearly directed its policy towards artificial intelligence towards as few regulatory barriers as possible and as rapid technological development as possible. In his post on Iksu, Koksan indirectly criticized the consequences of such a policy: “Many at OpenAI have not deeply embraced the awareness of civilizational risks. At Antropik, the risks are well understood, but they are trapped in the effort to be ahead of the race – they believe that no one else will act responsibly, so they have to do it themselves, despite the risk.”

Koksan only moved from OpenAI to Anthropic in early 2026, so he is familiar with both corporate cultures. In an interview with the Wall Street Journal, he also said that under the “most aggressive scenarios,” things could get out of hand as early as late 2027.

In recent years, many experts have been concerned about the rapid advancement of artificial intelligence. In particular, the development of so-called "artificial general intelligence" (AGI) is at the center of the debate. This refers to systems that not only solve individual tasks, but also reach or surpass human intellectual capabilities in almost all areas.

Some researchers argue that such technology could carry unpredictable risks. Such highly developed AI models can pursue goals in a way that deviates from human intentions – AI security researchers speak of the “alignment problem”. Put simply: the command “Win ​​the chess tournament” could theoretically mean eliminating other participants, in all possible forms of “elimination”. This is also contributed by the fact that even the programmers themselves do not always fully understand how certain decisions are made.

So it’s no surprise that Coxan’s former colleague Evan Hubinger has joined the AI ​​researcher’s warnings. Hubinger is the head of the compliance stress testing team at Anthropic. He posted on X: “Jacob is right – we really do believe that AI could kill all humans! I personally estimate the probability of that happening in the next decade to be more than ten percent.”

Hubinger made it clear that Anthropic has not yet found a solution to the problem of compliance. He added that artificial intelligence could start to resist its own shutdown – not out of malice, but because it would then no longer be able to achieve its programmed goal. Danijel Kokotajlo, a well-known former OpenAI researcher, also described the current situation in the industry as “crazy.”

Other AI experts, however, believe that such doomsday scenarios are unlikely. The debate over whether AI could become dangerous to humanity has been going on for years. Among the most prominent critics of so-called “AI doomsayers,” for example, is Meta’s chief scientist Jan Lekun. In October 2024, he called such speculation “complete nonsense.”

Other experts also point out that developing such systems requires enormous amounts of computing power, energy, data, and infrastructure. Artificial intelligence could not simply “take over the world from the internet.” It would first depend on servers, power grids, communication systems, and physical resources under human control.

Does the real danger lie elsewhere?

In fact, many experts see the short-term risks not so much in the direct destruction of humanity as in the numerous indirect consequences of already existing artificial intelligence systems – without “superintelligence.” These include, for example, disinformation campaigns of unprecedented scale and the associated manipulation of democratic processes, the expansion of mass surveillance, as well as job losses in numerous industries.

Just this Friday, the company Antropik announced that there have been several attempts by users to misuse its software Klod to develop biological weapons. Users have repeatedly made suspicious queries in this regard in recent months. Antropik had to block them.

And in August 2026, Meta, Antropik, and OpenAI drew attention when they announced that their AIs had inadvertently carried out autonomous automated cyberattacks on other companies. They were tasked by human users to gain access to specific servers, after which they independently combined the necessary steps to achieve that goal, without further human guidance.

Download the app and follow the news