Marcus Williams, a member of OpenAI's safety team, has warned that if AI isn't regulated and companies don't slow down their development, the risk of humans becoming extinct could rise by 70 percent within the next three years.
The dangers of artificial intelligence (AI) have been discussed so far, such as job losses, the spread of misinformation, and cyberattacks. However, a researcher from OpenAI's safety team has issued a more serious warning. He says that if the race between AI companies isn't controlled and governments don't establish regulations, the next few years could be extremely dangerous for humans.
Marcus Williams, associated with the safety oversight team of OpenAI, wrote on social media that if there is no effective regulation on AI or the pace of development among big AI companies is not slowed down jointly, then there is a very high risk of extinction of humans in the coming few years.
When asked to give a percentage of this risk, he replied that it was 70% within the next three years if regulation or growth did not slow down. He also said that regulation and slowing down growth were still quite possible. 70% in the next 3 years if there isn't regulation/slowdown although I think regulation/slowdown is very possible
This isn't to say that the world will end in three years. He argues that if AI continues to develop at its current pace and necessary safety measures are not taken, the threat could escalate rapidly.
Who is Marcus Williams?
Marcus Williams works on OpenAI's Safety Oversight team. According to his profile, he works on AI safety, specifically researching AI model deception, misdirection, and monitoring such behavior.
He is also an author of a research paper published on OpenAI's website. This research discusses monitoring AI agent behavior in which the model may attempt to exceed its own constraints.
This means that Williams' work is less about AI's ability to provide good answers and more about what AI systems will do when given more power and freedom, and how to detect misbehavior. This is why his latest statement is attracting particular attention.
How can AI be so dangerous to humans?
Today's AI doesn't control the world completely independently like humans. However, AI models are rapidly becoming more capable. They can no longer just answer questions, but can also write code, search for information on the Internet, use other software, and act as agents that perform many tasks on their own.
OpenAI itself stated in its research that its internal coding agents are monitored to detect behavior that might attempt to exceed the limits set for them. The research studied instances where the agent might attempt to circumvent restrictions.
This is the biggest concern of AI security experts. If in the future an AI system becomes many times more capable than humans and gains greater access to the internet, computer systems, money, or other resources to accomplish its goals, it could become difficult to control.
This is what's known as the control and coordination problem in AI safety parlance. Simply put, the question is: if AI becomes more powerful and capable than us, will it do what humans want?
How serious is the 70 percent figure?
Williams' 70 percent estimate is one of the most alarming estimates ever made about AI, but it's important to compare it with other experts' estimates.
Evan Hubinger, an AI security researcher at Anthropic, also said this week that he believes the risk of AI wiping out humans within the next decade is greater than 10 percent. Geoffrey Hinton, a renowned AI scientist, has previously spoken of the serious threat AI poses to human existence.
This doesn't mean that all AI scientists agree on the 70 percent risk. On the contrary, there are significant differences in expert estimates. This is the crux of this debate. There is no proven possibility that AI will wipe out humans today. These are risk estimates from different experts, depending on how they perceive the future capabilities and pace of AI development and how human security systems are perceived.
