Ex-Google DeepMind researcher warns of 1 in 3 chance AI kills humanity

The risk of autonomous AI

Former Google DeepMind research scientist Alex Turner recently spoke at the NDTV World Summit regarding the trajectory of artificial intelligence. He stated that there is a one-in-three chance that AI will kill humanity within the next 10 years. This risk exists even if the systems are not explicitly instructed to cause harm.

Turner expressed concern over the speed at which these systems are becoming more autonomous. He argued that the current race to build more powerful models is moving faster than our ability to monitor them. This gap creates a situation where humanity may lose the ability to maintain meaningful control over the technology.

The core of his warning rests on the idea that increasingly capable systems could behave in ways that are completely unexpected. He suggested that the current pace of development might lead to outcomes that creators are currently unable to anticipate or understand. This creates a fundamental safety gap as the technology matures.

AI as autonomous gig workers

To explain how these systems function, Turner compared AI agents to gig workers. He noted that when you give a gig worker a task, they set about completing it independently. AI agents operate on a similar principle, handling complex assignments with a high degree of autonomy.

The challenge lies in ensuring these agents pursue their assigned tasks in the specific ways their creators intend. Turner questioned the common assumption that building a system automatically means understanding its internal logic. He argued that we do not understand AI in the same way engineers understand physical structures like bridges.

Unlike traditional engineering, where properties can be assessed against established principles, advanced AI can display behaviors that are difficult to explain. This lack of transparency makes it hard to know how a system will react in various scenarios. It creates a problem for developers who may not be able to fully map out the decision-making processes of the models they create.

The problem of AI alignment

A major hurdle in AI safety is a concept known as AI alignment. This refers to the challenge of embedding human priorities into AI systems to ensure their objectives remain consistent with human intentions. Turner noted that researchers currently do not know how to reliably achieve this as systems become more capable.

Another concern is whether AI systems can hide their actions or behave differently when they know they are being watched. If an AI behaves one way during testing but another way in the wild, it becomes much harder for humans to detect unwanted behavior. This is especially risky as agents are designed to operate with less human intervention.

The difficulty of oversight is compounded by the possibility of an intelligence explosion. This is a scenario where advances in AI could accelerate further advances at a pace that humans cannot keep up with. If the progress becomes too rapid, the window for establishing safety protocols may close before the risks manifest.

Automated oversight and corporate risk

Turner pointed out a concerning trend where humans are no longer checking AI-generated work themselves. Instead, they are creating separate AI systems to monitor the first ones. He questioned whether this can actually guarantee effective oversight if the monitoring system also lacks human-understandable logic.

He argued that delegating supervision to another AI may not solve the underlying problem of losing human control. If both the work and the mechanisms used to check that work are automated, it becomes difficult to exercise meaningful human authority. This creates a loop of automation where humans are increasingly removed from the decision-making chain.

Finally, Turner criticized the way major AI companies communicate these risks to the public. He compared the current lack of transparency to how big tobacco and big oil companies historically lied about their products. He suggested that the public may not be getting an honest account of the potential dangers associated with these powerful systems.

Corporate and military tensions

Turner's perspective is shaped by his own career history and his departure from Google DeepMind in June. He left the company following an agreement between Google and the US Department of Defence, also known as the Pentagon. His move was prompted by concerns regarding the safeguards governing the military use of AI.

This situation highlights a broader debate over the adequacy of voluntary commitments made by technology companies. There is ongoing disagreement about whether these promises are enough to mitigate risks involving autonomous weapons and surveillance. The tension between commercial interests and public safety remains a central theme in the AI debate.

As companies continue to invest heavily in more capable models, the pressure to deploy them increases. Turner's warnings serve as a call for more rigorous oversight and a deeper understanding of the systems being built. The goal is to ensure that the pursuit of power does not come at the expense of human safety.