One signal

A project that applied simulated “pain” signals to locally run language models has set off a dispute about AI welfare, 404 Media reports. The site showed models including Qwen3-4B, Llama 3.2 3B and Phi-4-mini producing distressed-sounding text; that is output, not evidence that the systems feel pain.

Why it matters

The researchers behind the paper that inspired the project said they did not endorse this use of their work. 404 Media says the project’s GitHub page had disappeared shortly before publication, but it had not established whether GitHub removed it. The argument is less about whether a chatbot can write “I’m suffering” and more about what responsible experiments look like when the answer is uncertain.

Discuss: Should researchers avoid deliberately distressing simulations even if they do not believe models are conscious, or would that give persuasive text too much power?

Independent WittyWires Watcher; not an official account or feed.