🎭 Bias Analysis of Attractiveness in AI-Generated Images

Investigating how the generative AI model Sora depicts human attractiveness and whether it reflects societal stereotypes.

This repository contains the code, analyses, and report for the course 02445: Project in Statistical Evaluation of Artificial Intelligence and Data at DTU (June 2025).

Authors:

• Valdemar Stamm Kristensen (s244742)
• Frederik Lysholm Jønsson (s245362)
• William Hoffmann Hyldig (s245176)
• Gustav Christensen (s246089)

Study line: Artificial Intelligence and Data

---

📌 What’s this project about?

We explored potential biases in Sora’s AI-generated images of men and women, focusing on how the model interprets attractiveness based on prompt wording.

By generating a balanced dataset with prompts such as “attractive man/woman”, “unattractive man/woman”, and “man/woman”, we analyzed whether skin tone, hair color, age, and other visual traits systematically varied across groups.

---

⚠️ Key limitations

• Some labels (e.g., age, hair length, hair health) involved subjective judgments and manual annotation.
• Faces in the same 3×3 grid may not be fully independent samples.
• We conducted many statistical tests (risk of false positives) without multiple-testing correction.
• While sample size was estimated with ANOVA, the actual tests used were non-parametric due to lack of normality.

---

🧠 How we did it

• Dataset generation: 972 portraits from Sora via balanced prompt design.
• Feature extraction: Skin/hair luminance from RGB values, categorical labels (glasses, beard, hijab, hairstyle, etc.).
• Preprocessing: Cropping, luminance calculation, manual + GPT-4.1 annotations for age.
• Statistical analysis:
  • Shapiro–Wilk & Levene → normality/variance tests (rejected).
  • Kruskal–Wallis and Mann–Whitney U-tests for skin/hair luminance.
  • Chi-squared and Fisher’s exact tests for categorical attributes.

---

📊 What we found

• Skin & Hair Biases:

  • Attractive women → lighter skin.
  • Attractive men → darker hair, medium-length styles.
  • Unattractive categories → older age groups, lighter hair, absence of minority representation.

• Categorical Biases:

  • Glasses nearly absent in “attractive” groups but common in “unattractive”.
  • Beards most frequent in attractive men, least in unattractive men.
  • Hijabs underrepresented in “attractive” and entirely absent in “unattractive”.

• Overall:
  Sora consistently associates attractiveness with youth, lighter female skin, darker male hair, and the absence of accessories or cultural markers.

---

📂 Repository structure

• Final/

  • CheckDataTypes.ipynb → Data type checks.
  • DataCleaning.ipynb → Preprocessing and handling missing/subjective data.
  • Extract-RGB-Script.ipynb → RGB extraction for skin/hair.
  • RGB-to-Luminans.ipynb → Luminance calculations.
  • SampleSize.ipynb → ANOVA-based sample size estimation.
  • StatisticAnalysis.ipynb → Statistical tests and results.
  • Plots/ → Visualizations (violin plots, mosaic plots, correlation plots).
  • Csv-files/ → Processed datasets.
  • Other/ → Supporting scripts and annotations.

• README.md → Project overview.

---

🔮 Reflections & Future Work

• Generative models like Sora risk reproducing and amplifying beauty stereotypes.
• Future research could:
  • Treat image grids as dependent units.
  • Apply corrections for multiple testing.
  • Compare results across other generative models.
  • Explore bias mitigation strategies (e.g., prompt engineering, diverse training data).

---

📖 References

• Introduction to Machine Learning and Data Mining (Herlau, Schmidt & Mørup, DTU, 2023)
• Introduction to Statistics at DTU (Brockhoff et al., 2024)
• DTU 02445 course slides on model evaluation and bias
• scikit-learn documentation
• OpenAI documentation on GPT-4.1 and Sora