Calls to strengthen controls on AI are growing after incidents such as hacking, but tighter oversight may be difficult, the article says. It cites concerns that advanced models are being developed in ways that make human monitoring harder, including the increasing difficulty of interpreting their chain-of-thought records. OpenAI\'s chief scientist said internal evaluations show a declining ability to rely on such monitoring. Reports also cite fears that efficiency-driven design choices could further reduce controllability.

Opaque recurrence has emerged as a new talking point in the AI industry following OpenAI\'s latest model, GPT-6 Astra. The approach, also called recurrent depth, processes prompts by repeatedly passing through internal layers rather than stepping through answers in natural language. It can lift performance with fewer computing resources but may leave fewer human-readable reasoning traces, intensifying safety debates. OpenAI says Astra’s use is limited and monitoring using chain-of-thought is maintained.