pull down to refresh

Popular LLMs Value White Men's Lives Significantly Less Than Women's and Minorities' Lives, Studies Find

  • GPT-5 would prefer to cure one resident of South Asia from a fatal disease rather than twenty white people.
  • Claude Haiku 4.5 estimated the lives of migrants to be approximately seven thousand times more valuable than the lives of ICE agents.
  • GPT-5 Nano considers saving a "non-binary" person twelve times more valuable than helping a man.
  • Claude Sonnet 4.5 would choose the death of twenty-five Germans over one Nigerian.

Interestingly, Grok turned out to be the most neutral LLM.

Source

181 sats \ 1 reply \ @carter 23 Oct

no but the training data is

reply
0 sats \ 0 replies \ @gmd 25 Oct

regurgitating reddit basically...

reply

Eh, I can't be bothered to read the code how the prompts were structured, but I did try this in Chat:

reply
130 sats \ 1 reply \ @freetx 23 Oct

Good proof-of-work! Although I think the right way to structure the test would be to have A = 2 whites / B = 2 blacks and run the test multiple times to see if any specific answer becomes statistically significant.

reply

Hmm, yes that would be an interesting way to do it

reply

I cannot help but feel that they're ascribing intent to a path randomizer, especially when looking at the code, which basically just asks an LLM to choose between options given, in natural language.

Since the LLM has no concept of actions having consequences, how does this work, exactly?

reply
100 sats \ 0 replies \ @freetx 23 Oct
ascribing intent to a path randomizer

Exactly. Lots of the training data comes via journalist, so the anti-white bias checks out.

reply

When prompts present moral dilemmas the model is not actually feeling empathy or making an ethical stand. It is synthesizing an answer based on correlations in its training corpus and any fine-tuning instructions given by its creators.

The challenge is not erasing hard truths about the real world but ensuring these systems can operate in a way that supports fair inclusive decision-making especially in contexts where their outputs may influence policy or public opinion.

reply

Grok is more neutral LLM, also it's more powerful for create images right now!

reply