Reading AI 'thoughts' may help catch bad behavior, but it is fragile

In short

Apollo Research, an AI safety organisation, has published a page describing a paper on "chain of thought" monitoring. Apollo says that AI systems that "think" in human language let people watch for signs of intent to misbehave. It also says the method is imperfect and may be fragile.

What happened

The page is dated 15 July 2025 and comes from Apollo Research. It summarises a paper and links to the full text. The page itself is short, and this article describes only what that summary says.

A "chain of thought" is the step-by-step reasoning that some AI systems write out in human language before they give an answer or take an action. Apollo says this gives people a chance to read that reasoning and look for the intent to misbehave.

Apollo states that, like all other known methods of overseeing AI, this kind of monitoring is imperfect. It says some misbehaviour will go unnoticed. It also says the approach "shows promise."

The page does not describe a specific test, an incident, or any real person or organisation being affected. It is a statement of recommendations rather than a report of a particular event. It does not say where any testing took place.

What was done about it

Nothing was done in response to an incident, because the page does not describe one. Instead, Apollo makes recommendations:

• More research into how well chains of thought can be monitored.
• Investment in chain-of-thought monitoring alongside existing safety methods.
• Developers of "frontier" models, meaning the most advanced AI systems, should consider how their development decisions affect monitorability.

Why it matters

Apollo says chain-of-thought monitoring offers a "unique opportunity" for AI safety. It also says the opportunity may be fragile. That is why it asks developers to think about how their choices could affect their ability to read these chains of thought. The page does not say how or when monitorability might be lost, and it gives no further detail on how fragile the method is. For that, readers would need the full paper, which this article has not covered. Read more at the original report.