💻 computer science

Evaluating AI-Generated Images of Cultural Artifacts with Community-Informed Rubrics

This paper proposes a framework for integrating community perspectives into AI evaluation by involving marginalized groups in the "systematization" phase to define culturally appropriate representations of artifacts, which are then operationalized into automated measurement instruments using multimodal LLMs.

Original authors: Nari Johnson, Deepthi Sudharsan, Hamna, Samantha Dalal, Theo Holroyd, Anja Thieme, Hoda Heidari, Daniela Massiceti, Jennifer Wortman Vaughan, Cecily Morrison

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nari Johnson, Deepthi Sudharsan, Hamna, Samantha Dalal, Theo Holroyd, Anja Thieme, Hoda Heidari, Daniela Massiceti, Jennifer Wortman Vaughan, Cecily Morrison

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 are hiring a very fast, very smart robot artist to draw pictures of things from around the world. You ask it to draw a "white cane" for a blind person, or a "Kasavu saree" from India, or a "snake boat" from a festival.

The robot is fast, but it's also a bit of a dreamer. Sometimes, it draws a cane that looks like a candy cane, or a boat that looks like a toy, or a drum that has the wrong shape. To the people who actually use these items or wear these clothes, these mistakes aren't just "artistic choices"—they feel like the robot doesn't understand them at all. It's like if someone drew your grandmother's favorite recipe using ingredients she's never heard of; it looks like food, but it's not her food.

This paper is about how a team of researchers fixed this problem by asking the right people for help. Here is the story of how they did it, broken down simply.

1. The Problem: The Robot is Guessing

The researchers noticed that AI image generators (like DALL-E or Stable Diffusion) often get cultural details wrong. They might mix up a traditional Indian game with a Western board game, or draw a blind person's tool in a way that makes it useless.

The usual way to fix this is to have a computer program check the pictures. But the researchers realized: A computer doesn't know what "cultural respect" feels like. It only knows what it was trained on, and if it was trained mostly on Western images, it will miss the nuances of other cultures.

2. The Solution: The "Community Recipe Book"

Instead of letting the computer guess, the researchers went straight to the source. They held workshops with three specific groups:

Blind and Low Vision people in the UK.
Residents of Kerala, a state in South India.
Residents of Tamil Nadu, a neighboring state in South India.

They didn't just ask, "Is this picture good?" Instead, they treated the community members like expert taste-testers. They showed the group AI-generated pictures and asked:

"Would you show this to your family?"
"What is wrong with this picture?"
"What must be in the picture for it to be real?"

3. The Process: From "Vibe" to "Rulebook"

This is the most important part of the paper. The researchers turned the community's feelings into a Rulebook (or Rubric).

Think of it like this:

The "Vibe": A community member says, "This boat looks wrong. It needs to be long and narrow, and the rowers need to face backward."
The "Rulebook": The researchers turn that into a strict checklist:
1. Is the boat long and narrow? (Yes/No)
2. Are the rowers facing the back? (Yes/No)
3. Is the front of the boat a sharp point? (Yes/No)

If the AI picture passes every single check, it gets a "Pass." If it fails even one, it gets a "Fail."

They did this for six different items: a white cane, a Braille notetaker, a snake boat, a traditional drum, a board game, and a saree.

4. The Twist: Can a Robot Read the Rulebook?

Once they had these community-made rulebooks, the researchers asked a big question: Can we use a different AI (a "Judge AI") to automatically grade the pictures using these rules?

They tried it out. They fed the AI-generated pictures and the community rulebooks into a powerful AI judge.

The Good News: The AI judge was surprisingly good at spotting obvious mistakes. It could tell if a cane was the wrong color or if a boat looked like a house.
The Bad News: The AI judge sometimes struggled with the tricky stuff. For example, it had trouble checking if the "dots" on a Braille device were arranged in the correct pattern, or if the fabric of a saree looked like the right type of cotton.

It's like hiring a robot to grade a math test. It's great at checking if the numbers add up, but it might miss if the student used the wrong method to get there.

5. The Big Lesson: Humans First, Robots Second

The paper concludes with a very important message: You can't automate the "soul" of a culture.

The "Systematization" Phase (The Human Part): This is where you need real people. You need the blind community to tell you that a cane must have a white section, or the Kerala residents to tell you that a boat must have a specific pointed tail. This is the "Recipe."
The "Operationalization" Phase (The Robot Part): This is where you can use AI to check the recipe. The robot can quickly scan 1,000 pictures and say, "Hey, 900 of these have the wrong boat tail."

The Takeaway

This paper is a blueprint for how to build AI that respects everyone. It says:

Don't let the AI decide what is "culturally correct." It doesn't know the difference between a candy cane and a white cane.
Ask the community first. Turn their lived experiences into a clear checklist.
Use AI to do the heavy lifting. Once you have the checklist, let the robot check thousands of images to find the mistakes.

By doing this, we move from an AI that accidentally erases cultures, to an AI that is guided by the people who actually live those cultures. It's like giving the robot a map drawn by the locals, rather than letting it guess the way.

1. Problem Statement

Generative AI systems (specifically text-to-image models) frequently fail to accurately represent marginalized cultures, often replicating historical biases, reinforcing stereotypes, or causing cultural erasure. Current evaluation practices rely on "off-the-shelf" benchmarks and automated metrics that lack validity in low-data or culturally specific contexts. These standard metrics often fail to incorporate the "lived-experience expertise" of the communities being depicted. Furthermore, while there is a growing trend toward using Multimodal Large Language Models (MLLMs) as automated judges ("LLM-as-a-judge"), these models are trained on historical datasets that may lack diversity, leading them to misrepresent non-Western or marginalized cultural artifacts.

The core challenge is how to bridge the gap between community expertise (qualitative, lived experience) and automated measurement (quantitative, scalable evaluation) without losing the nuance required for valid cultural assessment.

2. Methodology

The authors propose a framework grounded in social science measurement theory, specifically adapting the Adcock and Collier framework (Systematization $\rightarrow$ Operationalization $\rightarrow$ Application). They conducted a study involving three distinct communities to develop and test this framework:

A. Community Engagement (Systematization Phase)

The researchers engaged with three communities to define "cultural appropriateness" for specific artifacts:

Blind and Low Vision (BLV) individuals in the UK: Focused on assistive technologies (Guide Cane, Braille Notetaker).
Residents of Kerala, India: Focused on cultural artifacts (Kasavu Saree, Chundan Vallam snake boat).
Residents of Tamil Nadu, India: Focused on cultural artifacts (Pallanguzhi board game, Mridangam drum).

Process:

Workshops: Synchronous online workshops were held where participants reviewed AI-generated images and real photographs.
Criteria Elicitation: Participants identified visual features that made an image "culturally appropriate" or "inappropriate." They distinguished between functionality (can the object be used as intended?) and recognizability (does it look like the specific cultural object rather than a generic substitute?).
Rubric Creation: The team synthesized these qualitative insights into binary evaluation rubrics (Yes/No criteria). These rubrics were organized under thematic headers (e.g., "The object must be functional," "The object must not resemble hegemonic objects").
Comparison: The team also generated rubrics automatically using GPT-4o to compare community-informed criteria against standard LLM-generated criteria.

B. Operationalization (MLLM-as-a-Judge)

The authors tested whether these community-informed rubrics could be applied automatically:

Dataset: Generated 300 images (50 per artifact) using five state-of-the-art models (DALL·E 3, Stable Diffusion 3/3.5, GPT Image-1, Flux.1).
Automated Evaluation: They used GPT-4o as an MLLM judge. The model was prompted with the specific rubric criteria and asked to score each image (1 = meets criterion, 0 = fails).
Human Validation: Research team members manually annotated the same images using the rubrics to establish a ground truth for comparison.

3. Key Contributions

Community-Centered Systematization: A demonstrated methodology for translating complex, subjective community knowledge into concrete, binary evaluation rubrics. This moves beyond simple "crowdsourcing" to deep engagement in the definition of metrics.
Empirical Evidence of LLM Limitations: The study shows that off-the-shelf LLMs (when used to generate rubrics) often miss critical cultural nuances. For example, LLMs suggested Braille notetakers should have visual screens (inaccessible to blind users) or failed to identify the specific black membrane on a Mridangam drum.
Feasibility of MLLM-as-a-Judge: The paper provides a rigorous evaluation of using MLLMs to apply community rubrics. It identifies that while MLLMs can achieve high agreement (78–88%) with human annotators on simple visual features (e.g., color), they struggle with complex spatial arrangements, material inference, and subtle cultural distinctions (e.g., the specific shape of a boat's stern or the validity of Braille cell arrangements).
Diagnostic Framework: The rubrics allow for the identification of specific failure modes in different AI models. For instance, the study revealed that while GPT Image-1 performed best overall, it still made specific errors (e.g., adding decorative patterns to drums) that other models did not.

4. Key Results

Model Performance: State-of-the-art models perform poorly on culturally specific artifacts. Only 40% of Guide Cane images were deemed culturally appropriate by human annotators; for other artifacts (like the Chundan Vallam boat), the rate was 0%.
Rubric Validity: Community-informed rubrics aligned with the majority judgment of workshop participants in >80% of cases, validating the systematization process.
LLM vs. Human Agreement:
- Overall Agreement: MLLM judges agreed with human annotators on the final "appropriate/inappropriate" label for 78–88% of images.
- Criterion-Level Variance: Agreement varied significantly by criterion type.
  - High Agreement (>90%): Simple visual features (e.g., "Is the cane white?").
  - Low Agreement (<50%): Complex spatial or material features (e.g., "Is the Braille arranged in valid 2x3 cells?", "Is the boat's stern pointed?").
Community vs. LLM Rubrics: LLM-generated rubrics frequently contained factual errors (e.g., suggesting Braille notetakers have screens) or omitted essential functional details (e.g., the leather straps on a Mridangam). Community rubrics were essential for capturing these nuances.

5. Significance and Future Directions

Shift from Data to Measurement: The paper argues that the bottleneck in evaluating generative AI for marginalized groups is not just data scarcity, but the lack of valid measurement instruments.
Pragmatic Participation: The study proposes a "pragmatic" approach where communities participate in the systematization phase (defining the rubric) rather than the entire lifecycle, making it feasible for industry to scale while retaining community input.
Limitations & Open Questions:
- Scalability: Engaging communities for every cultural artifact is resource-intensive. The authors suggest using LLMs to generate draft rubrics that are then validated/edited by communities.
- Subjectivity: Cultural representation is inherently contested. The binary nature of the rubrics simplifies this complexity, potentially masking intra-community disagreements.
- MLLM Reliability: The study highlights that MLLMs are not yet reliable enough to fully replace human judgment for complex cultural criteria, particularly regarding material properties and spatial logic.

Conclusion:
This paper provides a critical pathway for making AI evaluation more inclusive. It demonstrates that while automated MLLM judges are useful for scaling, they must be grounded in community-informed systematization to be valid. Without this grounding, automated evaluation risks perpetuating the very biases it aims to measure.