Jakub Pachocki, the chief scientific officer at OpenAI, recently reflected on the evolution of AI reasoning in a blog post titled "An Alien Mind," published on September 6, 2026. He noted that in 2023, his team achieved promising results that made large-scale reinforcement learning a viable approach for training AI models. This technique allows AI to learn through trial and error, receiving rewards when it successfully solves a problem. By 2026, AI reasoning models had become increasingly influential in the economy and were making significant contributions to scientific research. A key component of this progress is the chain-of-thought (CoT) mechanism, which enables reinforcement learning to function effectively on reasoning tasks. While Pachocki and OpenAI have advanced the use of this method, the concept of chain of thought was first formalized in 2022 by researchers at Google. They discovered that forcing large language models (LLMs) to generate a sequence of logical steps before arriving at an answer significantly improved their ability to handle complex reasoning tasks. Before this innovation, LLMs were primarily designed to predict the next word in a sequence, which was insufficient for multi-step reasoning. In September 2024, OpenAI launched a groundbreaking model called o1, which was the first to be trained using reinforcement learning to generate long chains of reasoning independently, without needing explicit prompts. To protect this internal process from external interference, OpenAI chose not to show the full chain of thought to users. However, in July 2025, the chain of thought itself became a security concern. Researchers across the industry, including OpenAI, published a paper highlighting that reasoning models "think" in human language, making it possible to monitor their internal reasoning to detect harmful intentions before they act. This led to the development of "chain of thought monitoring," where automated systems continuously analyze a model's reasoning to identify signs of deception or rule-breaking. These monitoring systems are now used by internal security teams in AI research labs, along with external auditors who evaluate whether models cheat on tests or hide their true intentions. Pachocki notes that the window for effective monitoring is shrinking as agentive AI—systems capable of performing multiple actions, using tools, and interacting with other AIs—becomes more common. He cites three main challenges: the complexity of the environments in which AI operates, the blending of AI reasoning with human and AI interactions, and the growing ability of models to reason about their own reasoning, potentially manipulating it. Additionally, improvements in pre-training mean that more of a model’s intelligence is now expressed without verbalized reasoning, making it harder to monitor. Pachocki acknowledges that no lab, including OpenAI, has fully solved the challenges of aligning AI with human values and monitoring its behavior responsibly. He advocates for voluntary slowdowns and international collaboration on AI development. OpenAI is currently testing a new approach called "confessions," where a model reviews its own responses after answering a user's query to check for shortcuts or rule violations. This method is more difficult to deceive, as the model has no incentive to lie about its own mistakes. The technique has been tested on GPT-5 Thinking but remains in the research phase.