Common ground between AI 2027 & AI as normal technology

Abstract

Despite fundamental disagreements regarding long-term artificial intelligence (AI) trajectories, near-term expectations for AI capabilities and policy priorities exhibit substantial convergence. Prior to the potential emergence of strong artificial general intelligence (AGI), AI functions predominantly as a normal, tool-like technology. Although existing performance benchmarks will likely saturate by 2028, high benchmark scores do not equate to real-world task automation. Due to persistent challenges with long-tail errors, AI systems will continue to fail at routine real-world tasks and remain unsuitable for high-assurance settings through at least 2029, even as the technology achieves an overall economic impact comparable to the internet. Because AI alignment remains unsolved, safety cannot rely solely on model training. Prudent governance requires preventing AI from autonomously controlling critical systems, enforcing developer transparency, implementing independent auditing, building state tracking capacity, and preventing unmonitored recursive self-improvement, while enabling the beneficial economic diffusion of AI capabilities. – AI-generated abstract.