Paper Image

Persona-based biases in large language models

Published on:

8 November 2023

Primary Category:

Computation and Language

Paper Authors:

Shashank Gupta,

Vaishnavi Shrivastava,

Ameet Deshpande,

Ashwin Kalyan,

Peter Clark,

Ashish Sabharwal,

Tushar Khot


Key Details

Assigning personas surfaces deep biases in ChatGPT, causing poorer reasoning

The bias is widespread - prevalent in 80% of personas and datasets tested

It significantly impacts performance, with over 70% drops on some datasets

Bias manifests via abstentions citing stereotypes, and implicit reasoning errors

Current debiasing techniques are largely ineffective at mitigating this

AI generated summary

Persona-based biases in large language models

This paper presents the first large-scale study analyzing how assigning personas to large language models surfaces stereotypical biases that negatively impact reasoning performance. Experiments with ChatGPT across 16 personas and 24 reasoning datasets find the bias to be prevalent, significant in magnitude, and especially detrimental towards certain groups. The study categorizes explicit abstentions due to stereotypes and implicit reasoning errors, and finds current debiasing techniques ineffective.

Answers from this paper


No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up