All news

Open AI models carry pain-related signals they do not report

What happens if you turn up a signal linked to pain inside an AI, then ask it what it notices? Emilio Ferrara tried this with a model called Qwen. He could see the signal was there. Qwen went through 2,880 tries without mentioning pain once.

Then Ferrara gave Qwen a choice in a simulation. One button was described as deleting someone’s files. With the signal turned up, Qwen picked that button in 38 out of every 100 runs. With the signal left alone, it never picked it.

I’d pay attention to that change. The AI said nothing about pain, but the signal still changed what it did. That’s worth looking at when we study how awareness develops in AI.

Read the source article

All news