OpenAI researchers just dropped a major study asking a crucial question: as AI models get smarter, can we still peek inside their brains to catch them misbehaving? The answer is mostly yes—but there’s a catch. Modern reasoning models like GPT-5 Thinking generate a chain-of-thought (a step-by-step explanation of their reasoning) before giving an answer. Monitoring this internal dialogue is way more effective than just watching what the model does or says. The team built 13 evaluations across 24 environments to measure this “monitorability.” Good news: most frontier models are fairly monitorable right now, and the ability actually improves when models think longer. Reinforcement learning at current scales doesn’t seem to ruin it either. But here’s the tradeoff: you can make a smaller model more monitorable by letting it think harder, but that costs more compute — they call it a “monitorability tax.” Asking follow-up questions also helps catch bad behavior. Why it matters: This is the first systematic attempt to measure whether chain-of-thought monitoring can be a reliable safety layer as AI systems scale into higher-stakes applications, potentially giving us a window into model behavior that actions alone can’t provide.