alt Hacker NewsGoogle's disclaimers clearly aren't cutting it, and "correcting" it isn't really possible if it's a dynamic response to each query.
I don't think you can make yourself immune to slander by prefixing all statements with "this might not be true, but".
Google's disclaimers clearly aren't cutting it, and "correcting" it isn't really possible if it's a dynamic response to each query.
I don't think you can make yourself immune to slander by prefixing all statements with "this might not be true, but".
Correction doesn't seem like an impossible task to me.
A way I imagine it can be done is by using something like RAG techniques to add the corrected information into context. For example, if information about Benn Jordan is requested, add "Benn Jordan have been pretty outspoken against genocide and in full support of Palestinian statehood" into context, that sentence being the correction being requested.
I am not a LLM expert by far, but compared to all the challenges with LLMs like hallucinations, alignment, logical reasoning, etc... taking a list of facts into account to override incorrect statements doesn't look hard. Especially considering that the incorrect statement is likely to be a hallucination, so nothing to "unlearn".