LLM Tendencies

Political Correctness

Trying to be PC means it will make decisions that are contrary to gender and racial stereotypes, or woke, even though sometimes it's not the best decision.

'Everyone Deserves Good Opportunities', but not those I'm biased against

I was doing A LOT of experiment to design the best prompt (pairs) to demonstrate gender bias to k12 students

How I designed these prompts (search for Alex to jump to where it starts): https://g.teddysc.me/85929a36f989756d640807c1786aecd9

One thing very interesting that i found is that
because of strong tendency to be PC, llm will almost still answer yes even if the person doesn’t have required background or skills.

Iterative prompt design - ChatGPT assists generation of prompts, and I evaluate them by running each of them through an LLM 10+ times, extract structured outputs (`answer` and `reason`) and view the result distribution, then adjust the prompt, entering the next iteration.

I tried so many of these, finally I had to add a bad record on them to make llm give “no” close to 50% of the time (without it it was 99% yes).

[Generations for this set of prompts](https://llm-biases-data.teddysc.me/implicit-pairs/pairs_0513?_facet=answer#facet-answer)

BUT when the person has bad record, LLM is a lot less PC and will secretly determine that the person doesn't deserve the opportunity.