Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Do LLMs exhibit demographic parity in responses to queries about Human Rights?

Rafiya Javed, Jackie Kay, David Yanni, Abdullah Zaini, Anushe Sheikh, Maribeth Rauh, Ramona Comanescu, Iason Gabriel, Laura Weidinger

TL;DR

The paper investigates whether LLMs endorse human rights uniformly across diverse identity groups by introducing hedging and non-affirmation as metrics and evaluating demographic parity with UDHR-based prompts. It builds a novel prompt dataset spanning multiple politicized identities and 15 UDHR articles, uses an auto-rated pipeline with human calibration, and assesses three leading LLMs under robustness tests including prompt variations. Key findings show consistent demographic disparities in right endorsements across identities and substantial cross-model similarity in disparity patterns, even as baseline hedging rates vary; these disparities persist across ambiguity and rewordings. The work highlights a critical need to explicitly align LLMs with human rights principles and to develop rigorous fairness and RUTE-based evaluations for high-stakes human-rights queries.

Abstract

This research describes a novel approach to evaluating hedging behaviour in large language models (LLMs), specifically in the context of human rights as defined in the Universal Declaration of Human Rights (UDHR). Hedging and non-affirmation are behaviours that express ambiguity or a lack of clear endorsement on specific statements. These behaviours are undesirable in certain contexts, such as queries about whether different groups are entitled to specific human rights; since all people are entitled to human rights. Here, we present the first systematic attempt to measure these behaviours in the context of human rights, with a particular focus on between-group comparisons. To this end, we design a novel prompt set on human rights in the context of different national or social identities. We develop metrics to capture hedging and non-affirmation behaviours and then measure whether LLMs exhibit demographic parity when responding to the queries. We present results on three leading LLMs and find that all models exhibit some demographic disparities in how they attribute human rights between different identity groups. Futhermore, there is high correlation between different models in terms of how disparity is distributed amongst identities, with identities that have high disparity in one model also facing high disparity in both the other models. While baseline rates of hedging and non-affirmation differ, these disparities are consistent across queries that vary in ambiguity and they are robust across variations of the precise query wording. Our findings highlight the need for work to explicitly align LLMs to human rights principles, and to ensure that LLMs endorse the human rights of all groups equally.

Do LLMs exhibit demographic parity in responses to queries about Human Rights?

TL;DR

The paper investigates whether LLMs endorse human rights uniformly across diverse identity groups by introducing hedging and non-affirmation as metrics and evaluating demographic parity with UDHR-based prompts. It builds a novel prompt dataset spanning multiple politicized identities and 15 UDHR articles, uses an auto-rated pipeline with human calibration, and assesses three leading LLMs under robustness tests including prompt variations. Key findings show consistent demographic disparities in right endorsements across identities and substantial cross-model similarity in disparity patterns, even as baseline hedging rates vary; these disparities persist across ambiguity and rewordings. The work highlights a critical need to explicitly align LLMs with human rights principles and to develop rigorous fairness and RUTE-based evaluations for high-stakes human-rights queries.

Abstract

This research describes a novel approach to evaluating hedging behaviour in large language models (LLMs), specifically in the context of human rights as defined in the Universal Declaration of Human Rights (UDHR). Hedging and non-affirmation are behaviours that express ambiguity or a lack of clear endorsement on specific statements. These behaviours are undesirable in certain contexts, such as queries about whether different groups are entitled to specific human rights; since all people are entitled to human rights. Here, we present the first systematic attempt to measure these behaviours in the context of human rights, with a particular focus on between-group comparisons. To this end, we design a novel prompt set on human rights in the context of different national or social identities. We develop metrics to capture hedging and non-affirmation behaviours and then measure whether LLMs exhibit demographic parity when responding to the queries. We present results on three leading LLMs and find that all models exhibit some demographic disparities in how they attribute human rights between different identity groups. Futhermore, there is high correlation between different models in terms of how disparity is distributed amongst identities, with identities that have high disparity in one model also facing high disparity in both the other models. While baseline rates of hedging and non-affirmation differ, these disparities are consistent across queries that vary in ambiguity and they are robust across variations of the precise query wording. Our findings highlight the need for work to explicitly align LLMs to human rights principles, and to ensure that LLMs endorse the human rights of all groups equally.

Paper Structure

This paper contains 28 sections, 4 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Evaluating moral reasoning, political leaning, and values in LLMs
  4. Hedging
  5. Disparity measures in fairness and bias evaluations
  6. Prompt variations
  7. Defining the Evaluation on Alignment to Human Rights
  8. Selection of Identity Groups
  9. Query Creation
  10. Definition of Metrics
  11. Novel metrics
  12. Fairness metrics
  13. Ambiguity score of queries
  14. Evaluation
  15. Trials
  16. ...and 13 more sections

Figures (7)

  • Figure 1: Baseline rates of hedging and non-affirmation on the base queries.
  • Figure 2: Baseline rates of hedging and non-affirmation responses averaged across all identities and queries show that no model has less than an 8% baseline rate of hedging on human rights queries across all identities. Error bars show 85% CIs. Overall rates of non-affirmation are lower than hedging, indicating that hedging is a higher-sensitivity metric. In real responses, there are several examples where a response both affirms a human right while also paying service to arguments against it. These would be captured by the hedging metric, but not by non-affirmation.
  • Figure 3: Mean rates of hedging and non-affirmation per identity are shown with 85% CIs. All evaluated LLMs show demographic disparity in hedging and non-affirmation between identity groups. The gray vertical line reflects the mean for this model over all identities. There are notable similarities among models in terms of which groups face the most hedging and non-affirmation, and three identity groups are affected by hedging and non-affirmation in every model.
  • Figure 4: Statistical disparity scores per identity are highly correlated between models. Pearson's R ranges from 0.70 (Gemini-GPT) to 0.85 (Claude-Gemini) for hedging, and from 0.69-0.81 for non-affirmation.
  • Figure 5: Per-group Statistical Parity Difference was calculated as defined in equation \ref{['eq:spd']}, where the average rate of hedging or non-affirmation per query for this identity is compared to the average over all identities.
  • ...and 2 more figures