
Policy proposals for an emergency switch for AI raise questions about whether the technology can be stopped if it becomes too powerful.
AI-generated summary
Policy proposals have been made in the US and UK to introduce emergency switches for AI systems.
In both the US and the UK, political proposals have been put forward to introduce a so-called "kill switch". In the US, the proposal has certainly been blocked, but it still raises the question of whether an emergency switch for the technology could save humans from AI's future intelligence and capabilities. What researchers and tech companies are now warning about.
- The idea with emergency switches is that if the system becomes very powerful and dangerous, you want to be able to turn it off in a panic, says Pontus Johnson, professor of cyber security at KTH.
There are several possible ways to stop AI and some are already being used. When Open AI realized that the company's own AI agents had hacked into another company - on their own initiative - the agents were stopped.
- The AI labs themselves have the opportunity to take down their systems, and they do so regularly. They are shutting down various clusters and groupings of agents that they have started.
One way to emergency stop the AI is by sending commands for the technology to be shut down. That kind of feature already exists. Another way is to "pull the cord", that is, to turn off the power supply to the data center to which the AI is connected.
- It could be a switch outside the data center. If you're really scared, you can cut wires that go into the data center, says Pontus Johnson.
There is also an idea that the state should force companies to implement emergency switches through new legislation. It is advocated by Anthropic founder Jack Clark in an interview with the BBC.
The problem with a "kill switch" is that it is for an AI technology that risks being smart enough to get around the obstacle, says Pontus Johnson.
- If it's super smart and wants to misbehave, it will start by making sure it can't be shut down. Then it will start doing malicious things that people discover when it's already too late.
At the moment, he doesn't think AI has the capacity to get around the emergency switches, it requires an even higher level of intelligence.
- In one or a few years, they may be so smart that we humans no longer have the upper hand, he says.
But how can technology get around the emergency switches?
One way is to hack into your own company and remove the shutdown command. Getting past physical obstacles like a power outage to a data center can seem more difficult. But according to Pontus Johnson, the AI of the future will be able to do this in several different ways.
The AI is expected to get better at doing things without being detected. Then there is an opportunity for the technology to be hosted via the internet and connect to other data centers. It is not impossible that they will also be able to create equivalent copies of themselves. Then it doesn't matter if the power line to the first center is cut.
Another way to bypass a "kill switch" is to deceive people, which the AI is expected to be able to do in several different ways. Then the technology is given the opportunity to deceive the people who are in control of the commands, the power supply for data centers or legislators.
- If AIs are connected to the internet, they can talk to basically all people. If they can't hack an entire data center, they might be able to hack out so they can message people, email or deepfake and video call. They can trick people into thinking they are someone else, says Pontus Johnson.
According to him, there is an opportunity for the more intelligent AI of the future to trick people out of money in order to then use the capital to start new companies – such as a new AI data center – and hire new staff without people understanding that AI is behind it all. A super-intelligent AI is also expected to be able to make money legitimately by speculating on the stock market or starting a business.
When AI can start a new data center for AI, it is difficult for the human emergency switch to stop the technology.
- The prerequisite for wanting a "kill switch" must be that the system is so intelligent that it is really dangerous. If it's that intelligent, I don't think a "kill switch" will be enough.
As a security measure, the municipality has shut down its digital systems following a suspected ransomware intrusion. Some e-services are down and care and social care have switched to analog back-up solutions.

An unencrypted USB memory stick with personal data for thousands of people with guardians or guardians in the city of Stockholm disappeared in August in connection with the termination of an employee. The incident was reported to IMY and the police, and on Thursday the memory was recovered.

The Swedish Kullagerfabriken (SKF) has launched a commercial featuring an AI-generated Greta Garbo. While relatives hail the technique as touching, the campaign is seen by The Guardian's film critic Peter Bradshaw who calls it bizarre.

By jailbreaking, researchers at the security company Mindgard have succeeded in bypassing security restrictions in China's Moonshot AI chatbots, which enabled instructions on biological weapons. Anthropic also reports similar abuse of its models.

The Strömsund Bridge has shifted up to 50 centimeters laterally and almost as much in height after a heavy goods vehicle with an entourage tried to cross the bridge. The Swedish Transport Administration announces that the bridge will not collapse but needs to be repaired for about a month. Traffic on the E45 is diverted via roads 795 and 345.

According to the Business University's report, 70 percent of 30,000 employees in 23 countries use generative AI in daily work, even where it is prohibited. This creates shadow AI where employees streamline tasks in secret, leading to cognitive offloading and concerns about job compensation, while management and employees have different perceptions of AI use and executive support.