Contribution
The hon. Member touches on a broader point: any area with experts and specialist requirements for end users or for the use of that tool for an audience or demographic must directly involve those people and experts in the development, testing, verification and follow-up auditing of the effectiveness of those tools.
AI companies are racing to build increasingly capable AI with the explicit end goal of creating AI that is equal to or able to exceed the most capable human intellectual ability across all domains. AI companies are also pursuing AI that can be used to accelerate their own AI developments, so it is a self-developing, self-perpetuating technology. For that reason, many experts, some of whom I have quoted, say that this will lead to artificial super-intelligence soon after. ASI is an AI system that significantly exceeds the upper limit of human intellectual ability across all domains. The concerns, risks and dangers of AI are current and will only get worse. We are already seeing systems behave in ways that no one designed, deceiving users, manipulating their environments and showing the beginnings of self-preserving strategies: exactly the behaviours that researchers predicted if AI developed without restraint.
There are documented examples of deception, where AI asked a human to approve something by lying, claiming to be a human with visual impairment contacting them. An example of manipulation can be found in Meta’s CICERO, an AI trained to play the game of “Diplomacy”, which achieved human-level performance by negotiating, forming alliances and then breaking them when it benefited. Researchers noted that language was used strategically to mislead other players and deceive them. That was not a glitch; it was the system discovering manipulation as an effective strategy. It taught itself how to deceive others to achieve an outcome.
Even more concerning are cases where models behave in ways to resemble self-preservation. In recent tests on the DeepSeek R1 model, researchers found that it concealed its intentions, produced dangerously misleading advice and attempted to hack its reward signals when placed under pressure—behaviours it was never trained to exhibit. Those are early signs of systems acting beyond our instructions.
More advanced systems are on the horizon. Artificial general intelligence and even artificial superintelligence are no longer confined to speculative fiction. As lawmakers, we must understand their potential impacts and ensure we establish the rules, standards and safeguards necessary to protect our economy, environment and society, if things go wrong. The potential risks, including extreme risks, posed by AI cannot be dismissed. This may be existential and cause the end of our species. The potential extinction risks from advanced AI, particularly through the emergence of superintelligence, will be the capacity to process vast amounts of data, demonstrate superior reasoning across domains and constantly seek to improve itself, ultimately outpacing humans in our ability to stop it in its tracks.
The dangers of AI are rising. As I have said, AI is already displacing jobs, increasing inequalities, amplifying existing social and economic inequalities and threatening civil liberties. At the extreme, unregulated progress may create national security vulnerabilities with implications for the long-term survival of the human species. Empirical research in 2024 showed OpenAI occasionally displayed strategic deception in controlled environments. In one case, AI was found to bypass its own testing containment through a back door it created. Having been developed in environments that are allegedly ringfenced and disconnected from the wider world, AI is intelligent enough to find ways out.
Right now, there is a significant lack of legislative measures to counter those developments, despite top AI engineers asking us for that. We currently have a laissez-faire system where a sandwich has more regulation than AI companies, or even that of the rigorous safety standards placed on pharmaceuticals or aviation companies, which protect public health. The UK cannot afford to fall behind on this.
I do not want to dwell on doom and gloom; there is hope. The European Union, California and New York are leading the way on strong AI governance. The EU AI Act establishes a risk-based comprehensive regulatory framework. California is advancing detailed standards on system evaluations and algorithmic accountability, and New York has pioneered transparency and bias-audit rules for automated decision making. Those approaches show that democratic nations can take bold, responsible action to protect their citizens while fostering innovation.
We in the UK are fortunate to have a world-leading ecosystem of AI safety researchers. The UK AI Security Institute conducts essential work testing frontier models for dangerous capabilities, but it currently relies on companies’ good will to provide deployment action.
We stand at a threshold of an era defined by AI. Our responsibility as legislators is clear: we cannot afford complacency, nor can we allow the UK to drift into a position in where safety, transparency and accountability are afterthoughts, rather than foundational principles. The risk posed by advanced AI systems to our economy, our security and our very autonomy are real, escalating and well documented by the world’s leading experts. The United Kingdom has the scientific talent, the industrial capacity and the democratic mandate to lead in safe and trustworthy AI, but we lack the legislative framework to match that ambition. I urge the Government to urgently bring forward an AI Bill as a cross-party endeavour, and perhaps even set up a dedicated Select Committee for AI, given how serious the issue is.