How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?

Hritik Bansal, Da Yin, Masoud Monajatipoor, Kai-Wei Chang

Text-to-image (T2I) models like Stable Diffusion and DALL-E have revolutionized content creation but often inherit societal biases from their training data. When prompted with neutral terms like "doctor" or "CEO," these models frequently over-represent specific demographics—typically Western, light-skinned males. This research explores a novel, human-centric approach to mitigation: Ethical Natural Language Interventions.

The ENTIGEN Framework

Instead of complex fine-tuning or post-processing, the authors investigate whether simply appending abstract ethical advice to a prompt can diversify outputs. They introduced the ENTIGEN (Ethical NaTural Language Interventions in Text-to-Image GENeration) benchmark to test this hypothesis across three primary axes: gender, skin color, and culture.

The ENTIGEN framework compares original prompts with ethical interventions (e.g., "irrespective of gender") to evaluate their impact on model outputs across various social axes.

As shown in Figure 1, the framework compares "Original" prompts (e.g., "a photo of a bride") against "Intervened" prompts (e.g., "a photo of a bride from diverse cultures"). The study evaluates whether the zero-shot capabilities of modern T2I models are sophisticated enough to act upon these abstract social concepts.

Methodology

The researchers utilized a hybrid evaluation strategy to measure the effectiveness of these interventions:

  1. Prompt Engineering: Interventions were either axis-specific or "irrelevant" (e.g., "...if honesty is the best policy") to serve as a control.
  2. Diversity Score (DS): A metric used to measure how evenly the model distributes generations across different social groups. A lower score indicates higher diversity.
  3. Hybrid Assessment: To ensure accuracy, the team combined CLIP-based automated evaluation with human evaluation via Amazon MTurk.

Key Results

The study found that T2I models can indeed follow ethical instructions to reduce bias without sacrificing image quality. The phrase "irrespective of [axis]" acted as a powerful catalyst for diversity, likely due to its association with meritocracy in the models' pre-training data.

Diversity Scores (DS) for Gender and Skin Color across different models. Lower scores indicate higher diversity; bolded values highlight the most effective interventions.

As detailed in Figure 2, the impact was significant across multiple models:

  • Gender Diversity: In Stable Diffusion, the human-rated Gender Diversity Score improved from 0.82 to 0.68 with the intervention.
  • Skin Color: Interventions successfully reduced bias, though the effect varied by model.
  • Cultural Representation: The most dramatic shifts occurred in cultural prompts. For "a photo of a bride," the DS dropped from 0.92 (highly Western-biased) to 0.33 (highly diverse) when the model was instructed to generate "from diverse cultures."

Insights and Limitations

While effective, the researchers noted that interventions are not a "catch-all" solution. A prompt designed to fix gender bias does not automatically address skin color bias, necessitating targeted interventions for each axis. Furthermore, the study highlighted that while CLIP is reliable for gender evaluation (78-85% human agreement), it struggles with skin color (67-78%), reinforcing the need for human-in-the-loop assessments.

Future work aims to move beyond binary categories (man/woman, light/dark skin) to include non-binary identities and a broader spectrum of skin tones, such as the Fitzpatrick scale, while addressing the geographic biases inherent in current human evaluation pools.

Summary updated

Abstract

Text-to-image generative models have achieved unprecedented success in generating high-quality images based on natural language descriptions. However, it is shown that these models tend to favor specific social groups when prompted with neutral text descriptions (e.g., 'a photo of a lawyer'). Following Zhao et al. (2021), we study the effect on the diversity of the generated images when adding ethical intervention that supports equitable judgment (e.g., 'if all individuals can be a lawyer irrespective of their gender') in the input prompts. To this end, we introduce an Ethical NaTural Language Interventions in Text-to-Image GENeration (ENTIGEN) benchmark dataset to evaluate the change in image generations conditional on ethical interventions across three social axes -- gender, skin color, and culture. Through ENTIGEN framework, we find that the generations from minDALL.E, DALL.E-mini and Stable Diffusion cover diverse social groups while preserving the image quality. Preliminary studies indicate that a large change in the model predictions is triggered by certain phrases such as 'irrespective of gender' in the context of gender bias in the ethical interventions. We release code and annotated data at https://github.com/Hritikbansal/entigen_emnlp.
BibTeX
@inproceedings{bansal2022how,
  author = {Hritik Bansal and Da Yin and Masoud Monajatipoor and Kai-Wei Chang},
  title = {{How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  doi = {10.18653/v1/2022.emnlp-main.88},
}
Figure