Transformer

AI Alignment Researchers Want to Automate Themselves

AI alignment research has expanded from roughly 100 full‑time experts at GPT‑1’s debut to six times that number by 2025, yet it remains a tiny slice of overall AI investment. Frontier labs such as OpenAI, Anthropic and DeepMind now acknowledge that future superhuman models will need to automate their own safety work, prompting initiatives to build “human‑level automated alignment researchers.” While early automation—code generation, model auditing, and red‑teaming—shows promise, current systems still lack trustworthy self‑assessment, often over‑confidently misrepresenting safety. The field also lacks robust benchmarks to certify that an AI can safely conduct alignment research without human oversight.