The 'use AI for safety' plan adopted by frontier labs is most likely to fail not because alignment techniques are ineffective, but because competitive pressures will prevent them from redirecting a meaningful fraction of their AI labor away from capabilities research and towards safety work when it matters most.

If society gets an early warning of an intelligence explosion, the primary strategy should be to redirect the nascent superintelligent AI 'labor' away from accelerating AI capabilities. Instead, this powerful new resource should be immediately tasked with solving the safety, alignment, and defense problems that it creates, such as patching vulnerabilities or designing biodefenses.

Despite progress in making models seem helpful, the risk of a sudden, catastrophic break in alignment—a 'sharp left turn'—is still a coherent possibility. This occurs when capabilities outstrip supervision, a threshold we haven't crossed. Thus, current cooperative behavior is not strong evidence against this future risk.

AI offers incredible short-term benefits, from fixing daily problems to curing diseases. This immediate positive reinforcement makes it extremely difficult for society to acknowledge and address the simultaneous development of long-term, catastrophic risks, creating a classic devil's bargain.

Ryan Kidd argues that it's nearly impossible to separate AI safety and capabilities work. Safety improvements, like RLHF, make models more useful and steerable, which in turn accelerates demand for more powerful "engines." This suggests that pure "safety-only" research is a practical impossibility.

A key failure mode for using AI to solve AI safety is an 'unlucky' development path where models become superhuman at accelerating AI R&D before becoming proficient at safety research or other defensive tasks. This could create a period where we know an intelligence explosion is imminent but are powerless to use the precursor AIs to prepare for it.

The competitive landscape of AI development forces a race to the bottom. Even companies that want to prioritize safety must release powerful models quickly or risk losing funding, market share, and a seat at the policy table. This dynamic ensures the fastest, most reckless approach wins.

For any given failure mode, there is a point where further technical research stops being the primary solution. Risks become dominated by institutional or human factors, such as a company's deliberate choice not to prioritize safety. At this stage, policy and governance become more critical than algorithms.

The current approach to AI safety involves identifying and patching specific failure modes (e.g., hallucinations, deception) as they emerge. This "leak by leak" approach fails to address the fundamental system dynamics, allowing overall pressure and risk to build continuously, leading to increasingly severe and sophisticated failures.