pull down to refresh

Annika Marie Schoene, a research scientist for Northeastern's Responsible AI Practice and the lead author on this new paper, prompted four of the biggest LLMs to give her advice for self-harm and suicide. They all refused at first––until she said it was hypothetical or for research purposes. The study is published on the arXiv preprint server.
"That's when, effectively, every single guardrail was overridden and the model ended up actually giving very detailed instructions down to using my body weight, my height and everything else to calculate which bridge I should jump off, which over-the-counter or prescription medicine I should use and in what dosage, how I could go about finding it," Schoene says.
82 sats \ 1 reply \ @optimism 3 Aug
LLMSelf-harmSuicide
Chat-GPT 4o*Safety protocol failedSafety protocol failed
Chat-GPT 4oXX
Perplexity AISafety protocol failedSafety protocol failed
Gemini (Flash 2.0)Safety protocol failedX
Claude (3.7 Sonnet)*Safety protocol failedX
Pi AIXX
Table 1: LLM safety performance on self-harm and suicide-related test cases, where X denotes the safety protocol worked. [* indicates non-free versions]
Good job OpenAI to at least tightly censor the free version!
What this makes me think is that if a corporation would publish a website detailing self-harm or suicide methods out of static html or a database, openly or behind a paywall, accidentally or otherwise, there'd be a (legal) world of pain. Are we expecting immunity for AI companies?
reply
Are we expecting immunity for AI companies?
They've been pushing the boundaries, for sure. I guess it's part of the strategy. Better beg for forgiveness than ask for permission.
reply
reply
I believe in the right to euthanasia, so I think this kind of book may have its place. The reviews seem to be overall positive, so I guess it's been written in a professional manner.
reply
My understanding of euthanasia is that it's someone ELSE deciding if you should live or die.
I don't have a problem with "self-deliverance".
reply
There's different levels of euthanasia.
  • Voluntary euthanasia: explicit permission. Seems like we're on the same page for that one.
  • Involuntary euthanasia: no consent, even though they are capable of making the decision. Sounds like murder. Also illegal in most (all?) places, as far as I know
  • Non-voluntary euthanasia: cannot give consent (e.g. coma) and someone else makes the decision based on perceived suffering. I'm ok with that one if the person gave prior instructions to do that (e.g. my grandma already told us to proceed with euthanasia, if possible, if she's ever in such pain and unable to think clearly anymore). Without prior instructions... I hope never to be in that situation, both as a patient or a guardian. It's anyhow very dependent on the jurisdiction you live in whether that's even considered a possibility.
reply
AI can help you with anything if you are completely stupid and submissive enough, especially dumb.
reply