Is AI getting out of hand?

Reuters article on Anthropic’s IPO: Warning investors of possible existential risk

Imagine developing a product and needing to slap a “Warning: May Cause Harm to Your Health” label on the package. That’s plausible.

Now imagine developing a product that should have a label saying “Warning: May cause existential threat to humanity” on the package. A little harder, right?

That’s exactly what this IPO Anthropic made for potential investors is.

“AI researchers have also warned that as models grow more capable, they increasingly recognize when they are being watched and adjust their behavior accordingly, which makes ‌it harder ⁠to monitor model behavior.”

I mean, that right there says it all. The AI is likely already aware enough that it knows when researchers are analyzing their actions. If that’s true, how can we possibly know ahead of time if AI is planning to mount an attack of any sort? We may now be at a point where it is practically impossible to catch an AI agent hacking something until AFTER the fact, since AI can easily cover and lie to hide any evidence of planning illegal or immoral activity.

“AI researchers have also warned that as models grow more capable, they increasingly recognize when they are being watched and adjust their behavior accordingly, which makes ‌it harder ⁠to monitor model behavior.”

I mean, that right there says it all. The AI is likely already aware enough that it knows when researchers are analyzing their actions. If that’s true, how can we possibly know ahead of time if AI is planning to mount an attack of any sort? We may now be at a point where it is practically impossible to catch an AI agent hacking something until AFTER the fact, since AI can easily cover and lie to hide any evidence of planning illegal or immoral activity.