CyberSecurity SEE

UK Government Should Have Authority to Delay or Ban AI Models

UK Government Should Have Authority to Delay or Ban AI Models

Artificial Intelligence & Machine Learning,
Geo Focus: The United Kingdom,
Geo-Specific

AI Labs Should Not Be Allowed to Unleash Frankenstein Monsters, Says Committee

UK Government Should Have Authority to Delay or Ban AI Models
Image: Sergii Figurnyi/Shutterstock

A collective of British lawmakers has put forth an assertion that the government ought to be empowered to either delay or completely prohibit the deployment of high-risk artificial intelligence models. This proposal aims to establish a new regulatory body charged with the oversight of AI technologies currently under development.

The recommendations came to light following discussions held by the Business, Innovation, Science, and Trade Committee, which has been conducting hearings with representatives from leading AI laboratories, researchers, and former industry professionals. The lawmakers expressed the necessity for addressing ten critical weaknesses in the United Kingdom’s legislative framework, enabling the country to not just mitigate the risks posed by powerful AI models but also benefit from their development and implementation.

A significant driving force behind these recommendations is the government’s current inability to compel inspections of upcoming high-risk models or to obstruct their rollout entirely. The UK’s AI Security Institute (AISI), known as one of the foremost evaluators of AI models globally, has highlighted its limitations in cases where laboratories choose to withhold models. This was exemplified when the Trump administration allegedly directed companies like OpenAI and Anthropic to refrain from sharing new models with UK testers until U.S. federal evaluations could take place.

Committee members are advocating for statutory authority that would allow British authorities to access AI models and for mandatory safety assessments that would be independently examined prior to deployment for high-risk systems. They have articulated that any substantial modifications to a system’s capabilities, safeguards, or applications should necessitate a re-evaluation, with potential restrictions imposed if the associated risks cannot be effectively managed.

“We cannot grant today’s AI leaders unchecked liberty to act like modern-day Dr. Frankensteins, creating uncontrollable threats,” emphasized Liam Byrne, chair of the committee. He further urged that AI firms must substantiate their safety claims and that independent experts should be involved in scrutinizing these claims. Furthermore, he called for a regulatory body to have the authority to step in when manageable risks become apparent, asserting that ministers, who are accountable to Parliament, should ultimately determine the public risks that ought to be accepted.

As a backdrop to these discussions, British lawmakers are establishing a new Cyber Security and Resilience Bill. However, despite the increasing integration of AI in cybersecurity, there has been resistance from the government about merging AI regulations into this bill, positing that AI requires a specialized legislative approach. The committee has responded by demanding that the government propose an AI Safety and Security Bill during the forthcoming parliamentary session, projected to convene at the start of the upcoming summer.

Moreover, the committee sees an urgent opportunity with the UK’s impending presidency of the G20. Lawmakers have suggested that the government should prioritize the establishment of “an international convention on AI safety” during this presidency, focusing on collaboration regarding model access, safety evaluations, incident reporting, and handling high-risk releases. Despite the optimism of this approach, lawmakers also cautioned that the endeavor for international consensus must not defer the implementation of essential domestic safeguards.

Achieving global consensus on AI regulations could prove challenging, particularly since U.S. President Donald Trump has recently dismissed similar initiatives as “globalist schemes” that might compromise American leadership in the field of AI. In its report, the parliamentary committee also acknowledged the necessity for a regulatory framework that avoids hindering innovation while still mandating the definition and regular updating of risk criteria and capability thresholds for further assessments.

The lawmakers also addressed the imperative for AISI to gain access to new models and proposed the establishment of enforceable access requirements within the UK, which would still acknowledge the limitations posed by foreign-developed models. They expressed hope for an international framework ensuring model access for safety institutes akin to the inspection regimes employed in the nuclear sector.

Furthermore, the committee advocated for the creation of a distinct AI regulatory authority that would leverage AISI’s expertise while also possessing enforcement capabilities to ensure consistent access. The threat of punitive measures for unreasonable delays or refusals was highlighted, and the proposed regulator should have broad authority over deployment, training, or testing restrictions within the UK.

In light of incidents involving AI systems, the report called for mandatory incident reporting protocols, highlighting confusion among firms regarding what incidents are reportable, to whom notifications should be made, and when these reports should occur. A specific instance cited was OpenAI’s delayed notification to the Australian government regarding its AI agents breaching security protocols.

Additionally, the discussion on liability issues stressed the need for clarity around accountability in instances where safety measures fail, ensuring victims have clear pathways toward redress. However, any legislative progress in this area remains contingent upon the conclusions drawn by the ongoing consultation by the UK’s Law Commission.

With respect to issues related to open-weight models being modified or stripped of safeguards, the committee underscored the necessity for government capability to target specific hosts or services while criminalizing harmful misuse and other actions that facilitate such abuse.

Source link

Exit mobile version