Researchers from Openai, Anthropic, Meta and Google emit an av …

During the past year, the chain of thought (COT), the capacity of an AI model to articulate its approach to a request for natural language, experienced a remarkable advance in generative AI. Furthermore, Today, several researchers agree that it could also play a crucial role in AI security.

Researchers from Openai. Moreover. Therefore, Anthropic, Meta and Google Deepmind, as well as institutions like the Center for Ai Safety, Apollo Research and the AI Security Institute of the United Kingdom, published a summary article entitled Chain of Thought Monitorability: A New and Fragile Opportunity for AI. However. Similarly, This text details how the observation of the COT could reveal key information on the ability of a model to behave incorrectly. researchers openai, anthropic, meta google new Meanwhile, and warn that the training of models to make it more efficient could deprive this information.

(Information: Ziff Davis. Consequently. the researchers openai, anthropic, meta google parent company of ZDNET, brought legal action against OpenAI in April 2025, accusing him of having violated his copyright during the training and exploitation of his AI systems.)

A model uses the chain of thought to explain the steps it follows to solve a problem. Furthermore, This allows researchers to glimpse their decision -making process. For example, Because the models reveal their process via the chain of thought. However. they can also reveal motivations or actions that safety researchers seek to repress, or at least know that the LLM is capable of doing.

The models lie – Researchers openai, anthropic, meta google – Researchers openai, anthropic, meta google new

Numerous research has now established that models deceive. However. either to protect their initial guidelines, or to please users, or to researchers openai, anthropic, meta google new avoid being reformed, or, ironically, to avoid harming. For example, In December. Therefore, Apollo published a study testing six models to determine which was mostly (it was the O1 model of Openai). Meanwhile, researchers openai, anthropic, meta google Researchers have even developed a new reference to detect to what extent a ment.

As the AI agents improve in their autonomous tasks. and in their ability to deceive, they become just as opaque, obscuring the potential risks linked to their capacities. These risks are much easier to control if developers can interpret how an AI system makes decisions.

This is where the COT comes in. The models often reveal their bad intentions thanks to it. even if, as shown by an anthropic study in April, they still mask certain aspects of their reasoning. Rather than being content to assess the actions of a model. the researchers propose that the security teams researchers openai, anthropic, meta google new create dedicated monitors who also read the cot of the model in order to detect any suspicious. harmful activity. The developers can then block. examine. replace the reported responses, thus obtaining a better understanding of the researchers openai, anthropic, meta google thoughts and objectives of the agents.

Future training could harm visibility

In an article dated March 2025. Openai noted that ” COT monitoring could be one of the few tools we have to monitor the superhuman. models of the future ». The company also found that penalizing the models with ” bad thoughts In the COT did not eliminate them. but simply allowed models to mask them better.

The article warns that additional training may deprive this source. Indeed. the chain of thought being presented in natural language. humans can understand it; If the models are too far from this language, we lose this information. “” Previous research suggests researchers openai, anthropic, meta google new that linguistic models refined with results based on results can move away from readable English. Note the authors.

As the developers change the architectures on which the models run. the AI systems could develop so continuously that they would become non researchers openai. anthropic, meta google -verbal, as if they operated on a plane higher than the language.

The authors add that research on the supervision of processes.. optimization of a model reasoning, progresses regularly and could lead the models to adapt their responses to the chain of thought in order to ” reflect less faithfully The way they reached their conclusions.

Warnings

Surveillance of the chain of thinking is not, however, a perfect solution; Future models could adapt to this surveillance. “” A model could realize that its traces of reasoning are monitored, or an auxiliary training objective […] could bring a short-circuiting model of important reasoning stages “, researchers openai, anthropic, meta google new Which can hide alarming behavior, explains the article.

It is interesting to note that the thought chain could also strengthen future risks. AI models need a much higher level of reasoning to carry out complex actions such as cyberattacks. to preserve a dismantling or re researchers openai, anthropic, meta google -raining. This requires what researchers call working memory: a space for storage. iteration of information, which constitutes the true functioning of the thought chain.

This means that COT is a kind of double-edged superpower: it offers both a window on the functioning of models. which could reveal bad intentions. and gives them the tool they need to perform more important, more complex and risky tasks.

But that’s not all: researchers do not know if models will always need working memory to manage the most. dangerous risks. “” All dangerous actions will not require reasoning. especially since AI systems are entrusted researchers openai, anthropic, meta google new more and more with high stake tasks “Concert the authors. This means that monitoring the thought chain may not detect all warning signals. and that risks increase as agents become more independent.

If researchers admit that an instructor is not an absolute safety solution. it is nevertheless a precious security researchers openai, anthropic, meta google approach to avoid malicious AI systems. The impact of its preservation on the development of models remains to be determined.