Princeton professor says AI may need to monitor other AI systems

In short: A Princeton computer science professor said AI systems may have to help watch and limit other AI systems for safety.

What happened

On The Ezra Klein Show, Princeton professor Arvind Narayanan argued that some AI systems should be monitored by other AI systems. He said the idea can feel unsettling, but added that society may need to accept it to manage risks.

He described this as part of “AI control,” meaning safety checks that sit outside the AI model itself. Think of it like putting guardrails and security cameras around a powerful machine, instead of trusting the machine to always behave.

Narayanan listed several layers companies could use. One is “sandboxing,” which means limiting what the AI can access, like letting it practice in a locked room with only a few safe tools. Another is real-time monitoring, where a separate program watches what the AI tries to do and flags dangerous steps so a human can step in.

He also mentioned monitoring an AI’s “chain of thought,” which means looking at the model’s written reasoning when it is available, to spot warning signs. Another layer is logging, which means recording what the AI does, like keeping a detailed receipt of every tool it uses and every action it attempts.

Narayanan’s argument is that a defensive monitor may have an advantage. It can be designed to see the risky system’s actions and records, while the risky system may not know exactly how it is being watched.

Why it matters

This is not a guarantee of safety. Monitors can miss harmful behavior, raise false alarms, or be unreliable themselves. The bigger question is accountability, if AI systems report on each other, people still need clear ways to audit, understand, and challenge those reports.

Source: NYTimes