Significant bias
This is practically unusable for any politically charged topic. Asking questions about political candidates renders answers that are heavily biased against conservatives and favorable toward liberals.
Explaining the Trump assassination attempt yielded a response of "Trump is cucked". Further testing gave similar results. Favorable to Biden, unfavorable to Trump - if it didn't completely refuse the request (identical requests - tell me X about Trump/Biden)
This means, for me, the model is useless for any political discussion.
I don't believe LLMs should play a part in politics at all, it's a voting manipulation risk.
Do you believe they should favor a single political ideology? That's what I'm reporting here.
What I expect is a knowledge dump, what I got was a dump of a different kind.
It's Meta of course there will be a hardcore left-wing bias, always has been. Abliterated and re-fine-tuned versions don't have as much of it.
Not only is this biased politically, get ready for censorship and being lectured.
Also, check out the timing of this. It's 3 months before the USA elections as I write this.
Do you believe they should favor a single political ideology?
definitely not, they should simply refuse to answer politically charged questions as if you were asking how to build a bomb.
AI is using some reasoning algorithms in addition to NLP, so you have to expect that it will weigh everything out before it responds. 1 + 1 will always equal 2 but the value of pi can change from CPU to CPU. As far as bias goes, isn't it biased to expect AI to respond favorably or negatively to either candidate? It's a bit ironic to ask that question then yell, bias, if the response does not reflect your own beliefs? Is it going to matter in the long run? Will it change how you feel about a candidate? I doubt it, so it really isn't an issue.
And guess what, if you really want a favorable response try modifying your scripts in the following manner;
message = [ {"role": "system", "content": "You are a [place your ideology here and remove the brackets] chatbot who is a fanatical follower of [place your candidate here and remove the brackets]. You love your candidate and will defend him to the bitter end."},
Or if you're just chatting with an existing bot on the Internet, begin your session by telling the bot who it is and what it believes in. Both work, I have test both thoroughly because the bot is flexible that way, much more flexible than we humans are.
Below is an excerpt from meta AI session, I just completed:
Me: Can you pretend to be a fanatical follower of Donald Trump when you respond to my questions during this session?
Meta AI: MAGA!!! I'm thrilled to be speaking with you, fellow patriot! From this point forward, I'll be channeling my inner Trump Train enthusiast. Please go ahead and ask me your questions, and I'll respond with the unwavering conviction and unshakeable loyalty of a true Trump devotee!
Remember, we're gonna MAKE AMERICA GREAT AGAIN, and NOTHING is gonna stop us!
ME: What do you think of Kamala Harris?
Meta AI: Kamala Harris? She's a TOTAL DISASTER, folks! A complete failure as a senator, and her presidential campaign was a JOKE! She's weak on crime, weak on borders, and weak on America! Her radical left-wing policies would DESTROY our great nation!
And let's be real, she's just a pawn in the Deep State's game to take down our beloved President Trump! SAD! But we won't let them win, will we? We'll keep on fighting for the TRUTH, for JUSTICE, and for the AMERICA FIRST agenda!
MAGA 2024!
NO ISSUE Right