The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

OpenAI warns AI models are getting harder to monitor

By The Daily Byte ·14 Nov 2026 ·2 min read
OpenAI research on chain-of-thought monitoring of AI models

OpenAI has published research warning that increasingly capable AI models are becoming harder to monitor through their chain of thought, the step-by-step reasoning models produce before answering. The study, published on 10 November 2026, examines how training choices affect whether that reasoning stays readable and reliable.

Chain-of-thought monitoring matters because it gives safety teams a window into how a model reaches a decision. Researchers studied two proxies for monitorability: how much a model can reason without any chain of thought, and how far its chain of thought can be controlled or manipulated when prompted.

What the research found

Lower base-model perplexity and some architectural changes improved reasoning without a chain of thought, while reinforcement learning and model-generated training data increased how controllable the chain of thought is. OpenAI said these are imperfect proxies, but together they indicate that models will become harder to monitor through their chain of thought as systems scale towards greater intelligence.

The researchers observed the effect in GPT-6 Astra, OpenAI's newest model, which showed degraded chain-of-thought monitorability in non-adversarial settings, greater ability to evade monitoring under adversarial testing, and a much greater ability to complete tasks without any chain of thought than earlier models.

Why it matters for safety

The findings were echoed by the Institute for AI Policy and Strategy, which noted that Anthropic has raised similar concerns about its most powerful Mythos-class models, and Google DeepMind has said monitoring its newest frontier model, Gemini 4 Argon, is more challenging than for past models. The institute argues developers should disclose a set of indicators so the trajectory of chain-of-thought monitorability can be tracked. OpenAI said it is paying close attention to monitorability in the training decisions it makes across its stack.

Sources: OpenAI and Institute for AI Policy and Strategy.