Generative AI Models Suffer from Degenerated and Biased Human Behavior

Strategic Insights -

Researchers from the Technical University of Darmstadt have found that generative AI models, despite their impressive results, are susceptible to degenerated and biased human behavior. These models, which rely on large datasets scraped from the internet, can perpetuate and even amplify existing biases. To address this issue, the researchers have developed a novel strategy called Fair Diffusion, which enables the attenuation of biases during the deployment of generative text-to-image models.

Key Takeaways:

  • Generative AI models have achieved impressive results in quality but suffer from degenerated and biased human behavior due to their reliance on large datasets.
  • The models can perpetuate and amplify existing biases, making them a concern for fairness and ethics.
  • The research has developed a novel strategy called Fair Diffusion to attenuate biases during the deployment of generative text-to-image models.
  • Fair Diffusion enables instructing generative image models on fairness without requiring data filtering or additional training.
  • The empirical evaluation of the research demonstrates the effectiveness of Fair Diffusion in controlling bias.
  • The research has been peer-reviewed and published in the journal AI and Ethics.

Statistics:

  • Generative AI models rely on billion-sized datasets randomly scraped from the internet.
  • The research used a novel strategy called Fair Diffusion to attenuate biases during the deployment of generative text-to-image models.
  • The empirical evaluation of the research demonstrated the effectiveness of Fair Diffusion in controlling bias, with results showing a significant reduction in bias.
  • The research was funded by the HORIZON EUROPE European Research Council, Deutsches Forschungszentrum fur Kunstliche Intelligenz, Hessisches Ministerium fur Wissenschaft und Kunst, and Technische Universitat Darmstadt.

Sources:

  • Auditing and instructing text-to-image generation models on fairness. AI and Ethics, 2024;5(3):2103-2123
  • NewsRx LLC. Technical University of Darmstadt Reports Findings in Artificial Intelligence and Ethics (Auditing and instructing text-to-image generation models on fairness). Robotics & Machine Learning. June 9, 2025; p 950.