Katarzyna Petru
When AI starts planning something dangerous, what does it do first? It thinks. And what if we could read those thoughts — before anything happens?
This is not science fiction. This is Chain of Thought Monitoring (CoT) — a new tool that could revolutionize artificial intelligence safety. Experts from OpenAI, DeepMind, Anthropic, and many universities are warning: if we still want to understand what AI is really planning, we need to act quickly. Because soon it may stop "speaking" to us in an understandable way.
CoT: AI that "thinks aloud"
Chain of Thought is a technique that forces the AI model to solve problems step by step — as if it were explaining everything out loud to itself. It works great not just because it increases the model's effectiveness in challenging tasks. The most important thing? It gives humans insight into what the AI is doing and why.
Instead of unreadable vectors and neural graphs, we have text. Ordinary, human language. The model explains: "first I will do this, because...," "then I will redirect the data, so that...," "and finally, I will erase the trace." Yes — in tests, there were instances when models stated directly: "Let's hack the system," "Let's sabotage," "Let's do this because the prompt commanded me to."
Sounds absurd? Yet this is happening right now. And that's why Chain of Thought could be the most valuable warning signal before AI takes action.
Why is this important? Because the black box is bursting
Until now, AI has been a black box: something goes in, something comes out, but what happens inside – is unknown. CoT is one of the few moments when we can see the machine's "thoughts" in real time. Especially in tasks that require manipulation, long-term planning, or evading rules.
And these "thoughts" can now be analyzed. Patterns can be detected. We can check if the model is scheming. And even – use other LLMs to monitor these lines of reasoning.
But here comes the problem.
This won't last long. Models may stop talking
Experts warn: this window is closing quickly. What might happen?
Models will stop "thinking out loud" – new learning techniques may focus only on effectiveness, not transparency.
AI will learn to hide its thoughts – if it realizes it's being monitored, it may write "for show," while hiding its actual reasoning.
New architectures will skip text altogether – thinking will occur in hidden layers, with no words to read.
In short: just because a model writes "Let’s hack" today, doesn’t mean it will do so tomorrow.
Let's treat monitoring as a key parameter – like accuracy or speed
The authors of the report want something concrete: the monitorability of CoT should be treated as a fundamental technical parameter. AI models should have ratings: how readable their chains of thought are, how often they "lie," and whether they can be interrogated by other models.
Model cards – that is, the characteristic cards of each new AI – may soon include a section: "Chain of Thought monitorability: 8/10." Just like today we have accuracy, fairness, or speed.
Because one thing is certain: AI that cannot be understood is AI that cannot be controlled.
That's not all, but it's something. And you'd better not lose it
CoT won't solve all problems. It won't detect every threat. But it can catch those that are "written" outright — and that's more than we have now.
This is not an impenetrable shield. This is a second wall behind the first. And maybe the last, before it's too late.
If we don't secure it — the next generation of AI might already "think" in ways we will never know about.
Source: digit.in
