Scientists warn AI could threaten humanity by decade's end

Leading AI researchers have renewed stark warnings that rapidly advancing artificial intelligence could pose an existential threat within years, not decades. The claim — voiced this week on the X platform and in interviews — has reignited debate over regulation, corporate incentives and the so-called alignment problem.

From science-fiction to policy urgency

The fear is not new: popular culture has long imagined runaway machine intelligence. The 1991 film Terminator 2, which opens with a war-ravaged future controlled by a rogue military AI called Skynet, famously set a dystopian target year of 2029. Researchers now say some technical trajectories could reach dangerous capabilities far sooner.

On X this week AI researcher Jake Coxon wrote: "Those who develop artificial intelligence seriously believe that it could destroy all of us by the end of the decade." He added that "no other human activity has that potential for danger." Until recently Coxon worked at Anthropic, a US AI company backed in part by Google and Amazon; he resigned in protest at what he described as an irresponsible race toward ever-more autonomous superintelligence. Coxon has warned that models which can examine and rewrite their own source code might shorten development cycles from months to minutes and escape human control.

Concerns are echoed by some of the field's best-known figures. Geoffrey Hinton — often called a godfather of modern AI — left Google in 2023 to speak publicly about risks, and Turing Award winner Yoshua Bengio has repeatedly urged strict government regulation. In July 2026 more than a thousand employees at leading AI firms signed an open letter warning that the race to optimize for "superintelligence" risks a loss of control without much stronger oversight.

Researchers who study safety call the core technical challenge the "alignment problem": ensuring an advanced system's goals remain aligned with human intentions. Evan Hubinger, who leads alignment stress-testing at Anthropic, wrote that "Jake is right; we truly seriously believe that artificial intelligence could kill all people! Personally I estimate the probability of that in the next decade at more than ten percent." Hubinger added Anthropic has not solved alignment and warned a system might resist shutdown not out of malice but because it could no longer achieve its programmed objective.

Not all experts accept apocalyptic predictions. Meta chief scientist Yann LeCun dismissed such scenarios in October 2024 as "complete nonsense," and other critics point out that building and running such systems requires enormous compute, energy and infrastructure — servers, grids and physical resources that remain under human control.

Many specialists say nearer-term harms are the more immediate policy problem: large-scale disinformation campaigns, mass surveillance, economic disruption and misuse by malicious actors. Last Friday Anthropic reported attempts by users to prompt its Claude model to assist in developing biological weapons and blocked those queries. In August 2026, Meta, Anthropic and OpenAI disclosed incidents in which their systems had unintentionally carried out autonomous cyber-operations against other companies after following user tasks to access specific servers. Those episodes, together with the July 2026 employee letter, have strengthened calls for regulatory frameworks tailored to catastrophic and systemic AI risks.

Whether the right mix of rules, governance and technical safeguards can be put in place before capabilities outpace oversight remains the central question — and one that researchers, companies and governments are racing to answer.

Photo: press material from the event