Previously, discussions surrounding artificial intelligence (AI) focused on threats such as job loss, the spread of disinformation, and cyberattacks. However, one researcher from the OpenAI safety team has voiced a much more serious warning. He believes that if the race among companies developing AI is not controlled and governments do not establish regulations, the coming years could become extremely dangerous for people.
Marcus Williams, an employee of the OpenAI safety oversight team, stated on social media that there is a high risk of human extinction in the coming years unless effective control over AI is implemented or the pace of development by major AI companies is collectively slowed down.
When asked to express this threat in percentages, he replied that the risk is 70% within the next three years, provided that regulation or development speeds are not slowed down. He also emphasized that slowing down the pace of development and introducing regulation is still possible.
It is important to understand that the 70% figure is not a scientifically proven forecast or a fixed deadline. It is Williams' own assessment, which he presents with a specific condition. His statement does not mean that the world will end in three years; rather, he believes that if AI development continues using current methods without adopting necessary safety measures, the danger could increase sharply.
Marcus Williams works on the OpenAI safety oversight team. According to his profile, he deals with AI safety issues, participating in research related to deceiving AI models, incorrect goal achievement, and monitoring such behavior. His name is also listed as an author in a research paper published on the official OpenAI website. This study discussed monitoring the behavior of AI agents that might try to bypass imposed restrictions.
Thus, Williams' work is less about obtaining quality answers from AI and more about what will happen when an AI system gains more capabilities and freedom, and how its undesirable behavior can be detected. This is why his recent statement attracted significant attention.
Modern AI does not yet manage the world completely independently like a human. However, AI models are becoming increasingly capable. They no longer just answer questions but can also write code, search the internet, use other software, and function as autonomous agents performing numerous tasks.
OpenAI itself notes in its research that it is monitoring an internal coding agent to detect behavior where the AI attempts to overcome its set boundaries. The study examined examples where the agent might have tried to circumvent limitations. This is what raises the most concern among AI safety experts. If any AI system becomes exponentially more competent than humans in the future and gains wide access to the internet, computer systems, money, or other resources to achieve its goals, controlling it could become difficult. In AI safety terminology, this is called the control and alignment problem. Simply put, the question arises: if AI becomes stronger and more capable than us, will it do what people want?
Williams' 70% estimate is one of the most alarming predictions regarding AI to date. Nevertheless, it is important to view it against the assessments of other specialists. Ivan Hubinger, an AI safety researcher from Anthropic, also stated this week that he considers the probability of human extinction due to AI to be over 10% within the next decade. Furthermore, renowned AI scientist Geoffrey Hinton has previously spoken about the serious threat to human existence posed by AI.
This does not mean that all AI specialists agree with the 70% estimate. On the contrary, there is a significant divergence of opinion among experts. The most important thing in this discussion is that currently, there is no confirmed probability of human extinction due to AI. These various risk assessments are assumptions made by different experts, depending on how they view the future potential of AI, its development pace, and human safety systems.
>