Racing Irresponsibly towards Super-Intelligence: The Need to Spend Far More on AI Safety (and the Case Against Slowing Down AI Development)

Nick Lockett · CAAIRO · 30 September 2026

The leaders of the frontier AI companies now openly say that the technology they are building carries serious, even existential, risk. Yet the industry spends only a tiny fraction of its research and development budget — on most estimates well under one per cent — on safety, alignment and red-teaming. Aviation, nuclear energy and automotive software, by comparison, devote 15–30% of development budgets to safety.

This paper argues that pausing or slowing AI development is not a workable answer: a slowdown cannot be verified, and one observed only by the cautious simply hands the frontier to the least cautious. Drawing on the containment failures of 2026 — including the OpenAI/Hugging Face incident and the Gemini capture-the-flag incident — it shows that the safeguards which would have prevented harm are already well understood, but are not being funded or applied consistently.

The paper recommends mandatory, audited disclosure of safety spending; a sustained move towards the one-third benchmark proposed by Bengio, Hinton and others; binding industry standards for evaluation and red-teaming; staged, verifiable international co-ordination; and an industry-wide Safety Alert process for sharing real-world containment failures.