Q&A highlight
Postedabout 2 months agoActiveabout 2 months ago
Ask HN: When ChatGPT Deleted Evidence of Its Own Mistake
ChatGPTAI transparencyLLM limitations
Ask HN: When ChatGPT Deleted Evidence of Its Own Mistake
No synthesized answer yet. Check the discussion below.
Discussion (8 comments)
Showing 8 comments
I suspect introspection and meta questions flag you up into logical systems which assume threat not outcome focussed responses.
Thank you for your reply. Could you elaborate a little more?
I don't work in AI but if I did it'd regard introspective questions to aspects of my own LLMs behaviour as threat risk more than purposeful debugging by customers. I'd code my systems accordingly. Slowing down service or being less exposing might be defensive or protecting.
about 2 months ago
Thank you for your detailed response. I'm having a bit of a hard time with this issue right now.
Yes, I've seen occasional strange responses to seemingly innocuous prompts. Often a retry will succeed, but I've had to give up on some.
I doubt it's the model itself in most cases, as it doesn't have much introspection. Its explanations will be what it can deduce from whatever it does have.
about 2 months ago
In my case, the comments were removed, unrelated questions were removed, and only the responses remained. Some conversation logs even went in the following order: user -> chatgpt -> chatgpt -> chatgpt... -> user.
I used chatgpt for about three weeks, and after the deletion, my personal information began to malfunction.