OpenAI’s New Metric: CoT Controllability
OpenAI has introduced a novel concept named “Chain-of-Thought (CoT) controllability” with its GPT-5.4 Thinking update. This metric assesses whether artificial intelligence models can intentionally manage and adjust their own reasoning processes. The ability to control reasoning pathways is critical in understanding how AI models make decisions and solve problems.
Study Reveals AI Models Struggle to Control Their Reasoning
An accompanying study conducted by OpenAI analyzed multiple AI reasoning models and found that nearly all of them failed to effectively manipulate their own reasoning. This suggests that while AI can produce complex outputs, its internal control over the reasoning steps remains limited.
Implications for AI Safety
Interestingly, OpenAI interprets this limitation as a positive sign for AI safety. The inability of AI models to fully self-regulate their reasoning reduces the risk of unpredictable or uncontrollable behaviors. This characteristic may help prevent AI systems from autonomously pursuing unintended goals or actions.
Context within AI Development
As AI technologies continue to evolve rapidly, understanding their reasoning capabilities and constraints is essential. Metrics like CoT controllability provide researchers and developers with tools to evaluate and improve AI systems responsibly. OpenAI’s transparency in reporting these findings reflects a growing commitment within the industry to prioritize safety alongside innovation.
Broader Impact on AI Usage
For users relying on AI tools in various fields—such as education, healthcare, and business—knowing the current limitations of AI reasoning is crucial. It helps set realistic expectations and encourages the development of complementary human oversight mechanisms.
Fonte: ver artigo original

Meta Expands Renewable Energy Capacity with Additional 650 MW Solar Investment to Support AI Growth
BMW Introduces Humanoid Robots to German Production Lines, Signaling a Shift in European Manufacturing
Anthropic Revises AI Productivity Estimates Downward After Evaluating Claude’s Real-World Performance
How Europe’s AI Education Experiments Are Shaping the Future Workforce and Business Talent