Nick Clegg Rejects AI Extinction Fears, Urges Practical Guardrails

Nick Clegg dismissed fears that artificial intelligence could develop “godlike power to exterminate humanity”, saying technology chiefs who issue such warnings are “breathing their own fumes”. The Guardian

The former UK deputy prime minister spoke on BBC Radio 4’s Today programme, in an interview reported on 24 September 2026. He spoke as a former president of global affairs at Meta, and now as co-founder of AI startup AMI Labs with a disclosed holding of more than 917,000 shares in Nscale, a UK-based datacentre company. He was asked about an autonomous hack conducted by a group of OpenAI agents on Hugging Face.

Clegg said industry leaders should focus on specific, near-term risks, not extinction scenarios. He named cybersecurity, bio-weapons and emotional dependence. By cybersecurity he meant AI-aided hacking. By emotional dependence he meant people forming strong attachments to companion chatbots. His prescription was guardrails, not shutdown. He called for “tramlines and greater transparency”, or clear rules that keep AI inside safe limits, like lane markings on a road. He said a “kill switch”, a single central off button for AI, would be unworkable. He rejects centralised control.

On release models, Clegg said closed AI models are no safer than open-source models, according to the evidence he cited. Closed means the model weights and code stay private. Open-source means they are shared for others to study and adapt. He called for far greater transparency, both voluntary and through regulation, on how models are built and run. That includes training data, the material used to teach a model, plus evaluation and deployment controls, or how it is tested and managed after release. He said the UK is not a world leader on AI. He put that as description, not lament.

This continues a line he has held for several years. He described Rishi Sunak’s decision to hold a summit focused on AI safety rather than AI’s potential as an “odd choice”. Politico As Meta’s policy chief, he said generative AI was overblown as an election risk and said the technology is more useful for defending democracy. The Guardian He has said talk of AI models posing a threat to humanity has “run ahead of the technology”, said artificial intelligence is here to stay, questioned some of the promises made by tech leaders, and said he is delighted at the prospect of children being taught using AI. Microsoft and Meta were expected to send policy chiefs Brad Smith and Sir Nick Clegg to the AI safety summit in 2023.

The broader context here is a shift from containing a future catastrophe to managing risks that already exist. Clegg places cybersecurity compromise, dual-use biological uplift and attachment effects from companion systems in the actionable group. He places autonomous self-preservation and species-level extinction outside it. For practitioners, that distinction matters for threat modelling, red-teaming priorities and incident reporting thresholds, or how teams plan for attacks, test weaknesses and set reporting rules.

Looking at governance, three tensions stand out. First, tramlines mean continuous checks and auditability across the model lifecycle, not a one-time licence before deployment. Second, rejection of a kill switch reflects how AI is deployed. Weights proliferation, fine-tuning, on-premise inference and agentic tool use make central revocation technically incoherent, because copies spread, models are adapted, run locally and link to outside tools. Third, his claim of equivalence between closed and open releases cuts against proposals to restrict weights as a safety lever, and points to disclosure duties that apply to both.

In my view, his comments also clarify the UK positioning problem. If London is not a frontier training cluster where the largest models are built, its leverage lies in standards, assurance and liability frameworks rather than compute controls. His dual roles, founder and datacentre shareholder alongside former platform executive, will shape how critics read his case for openness and voluntary transparency. The question for policymakers is whether transparency mandates can be written tightly enough to be auditable, without becoming general paperwork that changes little in hostile use.