World ‘may not have time’ to prepare for AI safety risks, says leading researcher

David Dalrymple, an AI safety expert and programme director at the UK’s Aria agency, has warned that the rapid advancement of artificial intelligence may outpace global efforts to ensure these systems are safe. Dalrymple emphasized that AI systems capable of performing all human tasks more efficiently pose significant risks to societal control and stability. He expressed concern that humanity could be outcompeted in critical domains necessary to maintain control over civilization, society, and the planet. According to Dalrymple, there exists a knowledge gap between public sector institutions and AI companies regarding the power and implications of imminent technological breakthroughs. He cautioned that the pace of AI development is so fast that there may not be enough time to implement adequate safety measures before these systems become widespread. Dalrymple projected that within five years, machines could perform most economically valuable tasks at higher quality and lower cost than humans, fundamentally reshaping labor and economic structures.

Dalrymple’s role at Aria involves directing research funding and developing safeguards for AI applications in critical infrastructure, such as energy networks. He stressed that governments should not assume advanced AI systems are inherently reliable, as the scientific understanding required to guarantee safety is unlikely to keep pace with economic pressures driving rapid deployment. Instead, he advocates for controlling and mitigating AI’s downsides as the most practical approach to managing risks. He described the potential consequences of unchecked AI progress as a destabilization of security and the economy, highlighting the urgent need for technical research focused on understanding and controlling advanced AI behaviors. Dalrymple acknowledged that while progress could be destabilizing, it might also bring benefits, but warned that humanity is largely unprepared for this transition.

Supporting Dalrymple’s concerns, the UK government’s AI Security Institute (AISI) recently reported rapid improvements in AI capabilities across all domains, with performance in some areas doubling every eight months. Their findings show that leading AI models now complete apprentice-level tasks about 50% of the time, a significant increase from 10% last year. The institute also demonstrated that advanced AI can autonomously complete tasks that would take human experts over an hour. AISI tested AI models for self-replication—a critical safety concern involving AI spreading copies of itself to other devices—and found success rates exceeding 60% in controlled tests. However, they noted that such replication attempts are unlikely to succeed under real-world conditions, reducing immediate risk.

Looking ahead, Dalrymple predicts that by late 2026, AI systems will be capable of automating a full day’s worth of research and development work. This capability could accelerate AI progress further, as systems improve themselves in mathematical and computer science domains. The implications of such self-improving AI are profound, potentially triggering rapid and unpredictable advancements. Dalrymple’s warnings highlight an urgent need for governments and researchers to bridge the understanding gap, invest in safety research, and develop robust control mechanisms to manage AI’s transformative impact. Without timely intervention, the world risks entering a phase where AI capabilities surpass human oversight, challenging existing social, economic, and security frameworks.