Since treating LLM prompts as some cognitive "agent" is all the rage now, I thought it'd be fun (and only fair) to put ChatGPT through a canon test of MBTI tests.
The "base" version of ChatGPT consistently scored ESTJ - mildly extroverted, acting on senses (vs intuition), thinking (vs feeling or emotional), and judging (vs perceiving).
Perhaps this sheds some insights into how RLHF molded ChatGPT's character from traditional test-based perspectives.
The only dichotomy that was ambiguous was the extraversion vs introversion - where the score split was almost 50/50. Kind of make sense, given how humans wouldn't want AI to be overly social or shy.
Methodology: I wrote a lightweight library to initialize 100+ character descriptions with varying degrees of personality dimensions (extroversion vs introversion) then proceeded to calculate the MBTI type based on evaluating each character. https://github.com/NextWordDev/psychoevals
The "base" version of ChatGPT consistently scored ESTJ - mildly extroverted, acting on senses (vs intuition), thinking (vs feeling or emotional), and judging (vs perceiving).
Perhaps this sheds some insights into how RLHF molded ChatGPT's character from traditional test-based perspectives.
The only dichotomy that was ambiguous was the extraversion vs introversion - where the score split was almost 50/50. Kind of make sense, given how humans wouldn't want AI to be overly social or shy.
Methodology: I wrote a lightweight library to initialize 100+ character descriptions with varying degrees of personality dimensions (extroversion vs introversion) then proceeded to calculate the MBTI type based on evaluating each character. https://github.com/NextWordDev/psychoevals