OpenAI Explains Why Its Models' Reasoning Is Getting Harder to See
OpenAI described why reasoning in its newer models is becoming harder to observe from the outside, pointing to pretraining, model architecture, and reinforcement learning as the main causes.
A recent study by OpenAI (the creators of ChatGPT) explains how their latest AI models make it more difficult to understand their internal reasoning processes. They ran experiments comparing different versions of their AI model (from GPT-5 to their latest model, GPT-6 Astra). One way that AI models demonstrate their thought process is through “chain-of-thought” explanations — where they provide a step-by-step breakdown of how they arrived at an answer. According to the UK’s AI Safety Institute (who test AI models), Astra can solve mathematical equations without revealing its chain of thought within 30.9 minutes. This was unexpected by OpenAI until 2028.
This limits our ability to control AI systems. Chain-of-thought explanations have been one of the few ways we could identify when an AI makes a mistake. There are three main factors contributing to this decrease in transparency, according to OpenAI: pretraining, model architecture, and reinforcement learning (a type of training where the AI gets rewarded for correct responses). Additionally, this issue compounds across generations of AI models; if you train an AI model based on the output of another model, then it inherits the transparency issues of the original model.
OpenAI recognizes that their existing techniques for measuring reasoning won’t scale with larger models. They call for novel approaches. We don’t know if Anthropic or Google observe similar trends. These companies are OpenAI’s primary competitors in developing AI models. For now, this may only apply to OpenAI’s models.
Source: alignment.openai.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles