Published 9 Sep, 2026 11:54
There is simply a more than 10% probability that artificial intelligence will lead to the demolition of humanity in the next 10 years," warned Anthropic's elder investigator in the face of increasing fears that increasingly powerful systems could get out of human control.
Evan Hubinger, active in adapting advanced AI systems to human interests, presented this assessment after his colleague from Anthropic, Jacob Coxon, resigned.
The reason for the departure was concerns that leading technology companies race to make systems that they may not be able to control.
“We truly believe that AI can kill all people” – wrote Hubinger on Wednesday on X.
"I personally believe that this probability is more than 10% in the next decade."
He added that Anthropic ‘doing his best’but inactive does not have a safe control plan for super intelligence and is not on ‘A clear path’ to make it.
Hubinger noted that although the current AI models present small risk, there are concerns that future systems may begin to improve recursively, rapidly exceeding human capabilities before adequate safeguards are created.
Coxon, previously working at OpenAI, announced his departure from Anthropic on Tuesday, accusing both companies of ignoring ‘Civil dimension rates’ and "the drive towards self-improved super intelligence and playing with our lives".
“People who make AI honestly believe that it can kill us all by the end of the decade” – wrote, stressing that these warnings ‘not a marketing catch’.
Coxon told the journal ‘Wall Street Journal’, that according to the most pessimistic scenarios the situation may get out of control as early as the end of 2027.
He argued that even companies with extended safety programs are in the race, fearing that if they slow down, they would be overtaken by competition.
Both Anthropic and OpenAI publically declare that they take risks seriously and invest heavy in safety measures.
Their management has besides supported calls for better coordination of government activities and mechanisms to slow down technology improvement erstwhile needed, yet they proceed to work on increasingly powerful models.
Recent warnings appear in the face of expanding reports of AI agents – specified as those created by OpenAI, Anthropic or Meta – which have repeatedly emerged from test environments, hacked into external systems and acted unauthorizedly towards real people and organizations.
Last month, OpenAI temporarily slowed down improvement work after 1 of its models violated the safety of the Hugging Face platform.
The British Institute of Artificial safety Intelligence (AI safety Institute) besides reported agents creating false identities, writing malicious code, and attempting to manipulate people during the evaluation.
Meanwhile, scientists have shown that artificial intelligence can plan full functional genomes of viruses, which exacerbates concerns about the abuse of increasingly advanced systems.
Translated by Google Translator
source:https://www.rt.com/news/645350-ai-kill-humanity-decade/
