OpenAI's latest model, Astra, has prompted a fresh debate in AI safety circles after researchers noted its reasoning appears less visible in text outputs. This shift toward reduced transparency raises concerns about the ability to monitor AI decision-making effectively, even as OpenAI and leading AI institutions push for stronger evaluation standards.
- Astra’s reasoning appears less visible, complicating oversight.
- Experts co-wrote a 2025 paper urging standardized monitorability reporting.
- EU rules formalize these reporting standards under the AI Office.
What happened
OpenAI’s Astra model has drawn criticism from AI safety researchers for performing much of its reasoning outside of visible text outputs, effectively making its thought process less accessible for human monitoring. Industry experts, including Ryan Greenblatt and OpenAI’s chief scientist Jakub Pachocki, have publicly discussed these concerns alongside the reported possibility that OpenAI may have intentionally limited Astra’s transparency to enhance capabilities.
This situation follows reporting by The Next Web and Semafor, highlighting that Astra reportedly solves complex problems internally without showing the detailed steps in text, raising alarms about the loss of explainability. Both OpenAI insiders and external researchers have engaged in these conversations, underscoring the shared recognition of the problem.
Why it matters
The ability to monitor and understand AI models’ reasoning is considered critical for ensuring safety, ethical alignment, and regulatory compliance. Lack of transparency can create risks by making it harder to detect errors, biases, or attempts by AI models to evade oversight.
In July 2025, a coalition of over 40 AI researchers from leading organizations including OpenAI, Google DeepMind, and Anthropic published a position paper emphasizing the fragile opportunity to monitor chain-of-thought in AI systems. They recommended standardized evaluations and transparent reporting methodologies. These standards are now being implemented across Europe through the EU’s AI code of practice, which requires detailed filings to the AI Office prior to market introduction.
What to watch next
OpenAI and other AI developers will need to balance advancing AI capabilities with maintaining sufficient transparency to comply with evolving safety expectations and regulatory demands. How Astra and similar models are monitored, audited, and reported will serve as a crucial test case in this regard.
Regulatory bodies like the EU’s AI Office will closely scrutinize submitted model reports, including standardized samples of inputs and outputs to ensure external evaluators can independently verify AI behavior. Industry-wide adoption of these reporting practices may set new norms for AI development and deployment worldwide.