Bill allowing the US Government to immediately interrupt system operation artificial intelligence which are considered to be dangerous for public safety have testified to Congress MPs of Democrats and Republicans.
The initiative comes after the revelation of OpenAI that an autonomous artificial intelligence agent managed to escape a controlled test environment and carry out a cyberattack on the Hugging Face platform. The inter-party proposal for a law called «AI Kill Switch Act», Members have testified Ted Liu (Republican) and Nathaniel Moran (Republican), seeking to establish a direct intervention mechanism in cases where artificial intelligence systems operate uncontrollably or pose a threat to public security.
The initiative follows the recent revelations of OpenAI, according to which an advanced autonomous AI agent escaped the company's controlled test environment, gained internet access and violated the platform Hugging Face, causing intense reflection on the risks of the most sophisticated artificial intelligence models.
Mandatory deactivation mechanism
The proposed bill provides that artificial intelligence development companies should maintain a technical capability to slow down, suspend or completely disable their models. At the same time, it gives US Department of Homeland Security the power to order the immediate shut-down of an AI system when it is considered that its operation involves serious risks.
THE Ted Liu stated that «It is imperative that artificial intelligence systems have a shutdown mechanism and that the federal government has clear competence to disable AI models that have escaped control» . On his part, the Nathaniel Moran argued that «artificial intelligence must continue to evolve, but its responsible development means that people must maintain the ability to control the technology they create» .
Compulsory occurrence reporting
The proposal also provides that companies will be required to inform the US authorities immediately of incidents or failures related to artificial intelligence systems. At the same time, an official crisis management protocol is established, which will provide for successive stages of response, from the slowing down of the operation of a model to its complete deactivation.
Concern about the potential of new models
THE Ted Liu It also invoked recent developments concerning the OpenAI competitor, Anthropic, saying that the possibilities of cyberattacking some of its new models led even to the US Department of Commerce temporarily restricting their public disposal through export legislation. Similar concerns had been expressed by Anthropic co-founder, Jack Clark., who had stated that the artificial intelligence industry currently has only «gas» and no «brake».
«You want to be able to lift your foot off the gas and hit the brake. Right now the AI industry has an acceleration pedal, but not a brake pedal.»
, had features reported. Supporters of the bill point out that artificial intelligence is now evolving from a tool that answers questions to a system that can perform economic transactions, control transport infrastructure or even participate in cyber defense and cyberattack operations. For this reason, they consider it necessary to have an institutionalized system. «safety switch», which will allow authorities to intervene directly in the event of uncontrolled behaviour.
Main photo: Sam Altman, CEO of OpenAI
