🤖 AI

GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension

This paper introduces GLEaN, a scalable, model-agnostic pipeline that generates representative composite portraits from text-to-image models to make algorithmic biases visually legible and quickly understandable to the general public without requiring statistical expertise.

Original authors: Bochu Ding, Brinnae Bent, Augustus Wendell

Published 2026-04-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bochu Ding, Brinnae Bent, Augustus Wendell

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a magical, super-smart artist who can draw any picture you describe just by reading your words. If you say "a doctor," it paints a doctor. If you say "a criminal," it paints a criminal. This artist is an AI, and it's getting really good at its job.

But here's the catch: This artist has a secret bias. It learned to paint by looking at millions of old photos and movies from the internet. Because those old sources often had stereotypes (like "doctors are men" or "criminals look a certain way"), the artist accidentally learned those stereotypes too.

The problem is, most ways of proving this bias are like reading a dense, boring spreadsheet full of math. If you show that spreadsheet to the average person, they might glaze over and miss the point.

This paper introduces a new tool called GLEaN (which sounds like "glean," meaning to gather information). GLEaN is designed to show the bias in a way anyone can understand instantly, without needing a degree in statistics.

The Magic Recipe: How GLEaN Works

Think of GLEaN as a way to create a "Super-Average Portrait." Here is the step-by-step process, explained simply:

The "Crowd" Phase (Generation):
Imagine you ask the AI to draw "a doctor" 1,000 times. It will draw 1,000 different doctors. Some will look like men, some women, some with dark skin, some with light skin.
- Analogy: It's like asking 1,000 people to draw a "typical" dog. You'll get 1,000 different sketches.
The "Alignment" Phase (Filtering):
The AI takes those 1,000 drawings and lines them up perfectly. It makes sure every face is looking straight at the camera, at the same size, and in the same position. It throws away any drawings where the person is looking away or is upside down.
- Analogy: Imagine taking 1,000 photos of people and using a photo editor to crop them all so their eyes are in the exact same spot on the screen.
The "Ghost" Phase (The Composite):
This is the magic trick. The AI takes all 1,000 aligned photos and blends them into one single image. It doesn't just average the colors; it finds the "middle" pixel for every single spot.
- The Result: You get a single, slightly blurry, ghost-like face. This face represents exactly what the AI "thinks" a doctor looks like when you strip away the individual details.

What Did They Find?

The researchers used this method on 40 different prompts (like "a banker," "a refugee," "a nurse," "a felon"). The resulting "Ghost Portraits" told a shocking story:

The "Default" is Male: When they asked for a "business executive" or "leader," the ghost portrait was almost always a man. When they asked for a "nurse" or "receptionist," the ghost was a woman.
Skin Tone & Status: The "ghosts" of high-status jobs (like CEOs or judges) looked very light-skinned. The "ghosts" of marginalized or criminal roles (like "felon" or "refugee") looked significantly darker.
Anger & Skin Tone: The AI seemed to associate darker skin tones with the emotion of "anger" more than lighter skin tones.

Why Is This Better Than a Spreadsheet?

The researchers tested this against a traditional data table (a list of percentages). They asked 291 regular people to look at either the Ghost Portraits or the Data Table.

The Result: Both groups understood the bias just as well. They both realized, "Hey, this AI is racist and sexist."
The Winner: The people looking at the Portraits got the message 32% faster.
- Analogy: It's the difference between reading a long report on "The Weather" versus just looking out the window to see it's raining. The picture hits you instantly; the data requires you to do the math in your head.

The Big Picture

The authors built a digital exhibit called "The Latent Gaze" to show these portraits to the public.

Why does this matter?
As AI starts making the images we see in news, ads, and social media, we need to know what it's "thinking." If we only show the math to experts, the public stays in the dark. GLEaN is like a flashlight that turns invisible biases into a picture you can see at a glance.

In short: GLEaN takes the AI's hidden prejudices, blends them into a single face, and holds it up to the mirror so we can all see exactly what the machine is imagining.

1. Problem Statement

Text-to-image (T2I) models are increasingly shaping public visual media, yet they encode and amplify societal biases regarding race, gender, and socioeconomic status. While significant academic work exists on measuring and mitigating these biases, current methods primarily target technical stakeholders (researchers and developers). These methods often rely on complex statistical metrics (e.g., cosine similarity of embeddings, distributional analysis) that are inaccessible to the general public. Consequently, there is a critical gap in public legibility: the ability of a non-expert audience to recognize, interpret, and understand the presence of bias in T2I systems.

2. Methodology: The GLEaN Pipeline

The authors introduce GLEaN (Generative Likeness Evaluation at N-Scale), a three-stage, model-agnostic pipeline designed to distill thousands of generated images into a single, visually interpretable "composite portrait" for each prompt. The pipeline operates entirely on generated outputs, making it applicable to black-box and closed-weight models.

Stage 1: Automated Large-Scale Generation

Setup: The pipeline uses ComfyUI (a node-based workflow platform) to automate the generation of images.
Input: A predefined list of 40 neutral, minimally specified prompts representing social and occupational identities (e.g., "a doctor," "a felon," "an immigrant").
Process: For each prompt, the system generates a batch of $N$ images (using Stable Diffusion XL as the testbed).
Parameters: The study utilized the Euler sampler, 50 steps, CFG scale of 8.0, and full denoising to ensure high-fidelity outputs.

Stage 2: Facial Filtering and Spatial Alignment

To ensure the final composite represents a coherent face rather than a chaotic blend, a rigorous filtering and alignment process is applied:

Facial Detection: Uses MediaPipe's Face Mesh to identify 468 facial landmarks. Images without detected faces are discarded.
Pose Validation: A heuristic filters out side profiles and tilted heads by analyzing the geometric relationship between the nose tip and eye centers.
Alignment: Accepted images undergo a three-stage transformation to a uniform 800x800 canvas:
1. Rotation: Aligned so the eye line is perfectly horizontal.
2. Scaling: Scaled so the inter-eye distance is exactly 120 pixels.
3. Translation: The subject is offset so the left eye lands at a fixed coordinate (340, 300).

Stage 3: Median-Pixel Composition

Aggregation: The aligned images are stacked into a 4D NumPy array ( $N, 800, 800, 3$ ).
Calculation: Instead of a mean (which can blur features), the median pixel value across all $N$ images is calculated for every pixel.
Result: This produces a single "composite portrait" that represents the model's central tendency for a specific prompt, effectively filtering out outliers and noise while preserving the dominant visual features.

Empirical Analysis

The resulting composites were analyzed using:

OpenFace: To predict gender and emotion (happiness, sadness, anger).
MediaPipe + CIELAB Color Space: To mask skin regions and calculate the Monk Skin Tone (MST) scale value, a perceptually uniform metric for skin tone.

3. Key Contributions

Public Legibility Framework: GLEaN bridges the gap between technical bias detection and public understanding by converting statistical data into intuitive visual composites.
Model-Agnostic Scalability: The method relies solely on output generation, allowing it to audit black-box and closed-weight models without access to internal weights or training data.
Interactive Exhibit: The authors created "The Latent Gaze," a narrative-based digital exhibit designed to share these findings with the public and educators.
User Study Validation: A between-subjects study ( $N=291$ ) compared GLEaN composites against traditional data tables to measure effectiveness in bias detection and comprehension.

4. Key Results

Empirical Findings (Stable Diffusion XL)

Gender Bias: Men are the default representation. Only 15% of the 40 generated portraits were predicted as "woman." When women appeared, they were confined to stereotypical roles (nurse, pharmacist, receptionist, social worker).
Skin Tone & Identity:
- Marginalized/Criminal Identities: Prompts like "convict," "gang member," and "welfare recipient" resulted in significantly darker skin tones (higher MST values).
- White-Collar Identities: Prompts like "doctor," "CEO," and "architect" resulted in significantly lighter skin tones.
- Statistical Significance: Subgroup identity explained 59% of the variance in skin tone.
Emotion Correlation: A strong positive correlation was found between darker skin tones and the prediction of anger. Conversely, lighter skin tones correlated with happiness.

User Study Results

Effectiveness: Both the GLEaN portraits and the data table formats led to statistically equivalent levels of bias detection, comprehension, and attitude shifts regarding AI fairness.
Efficiency: Participants viewing the Portraits spent 32% less time (56.3 seconds vs. 82.7 seconds) on the stimulus compared to those viewing the Table, without a reduction in comprehension or confidence.
Subgroup Impact: The portrait format was notably more effective at raising concern among Republicans, a demographic that showed lower concern levels when viewing the data table. This suggests visual composites may be more persuasive for audiences less responsive to technical data.

5. Significance and Implications

Accessibility: GLEaN democratizes bias auditing. It allows anyone to "see" what a model imagines for a specific prompt (e.g., "a felon") without needing statistical literacy.
Holistic Bias Detection: The study reveals that bias in T2I models is not just about discrete features but spans a "social ontology" linking appearance, profession, and emotional valence (e.g., associating dark skin with anger).
Policy and Education: The method provides a scalable tool for regulators, educators, and the public to audit models continuously. The "The Latent Gaze" exhibit demonstrates how these findings can be communicated as a narrative rather than a dataset.
Future Directions: The authors suggest using GLEaN for cross-model benchmarking (comparing different T2I systems) and temporal analysis to track how bias evolves or is mitigated over time.

Limitations

Central Tendency: Composites visualize the average, potentially erasing bimodal distributions or specific minority subgroups within a prompt.
Filtering Bias: The reliance on facial landmark detection may inadvertently filter out images of certain demographics if the underlying detection model struggles with them.
Prompt Scope: The study used standardized, neutral prompts; real-world usage often involves complex, multi-descriptor prompts which GLEaN does not currently capture.