Beyond Debiasing: The First Experimental Report of Adaptive Counter-Bias Compensation Through Targeted Positive Intervention to Enhance Underrepresented or Disadvantaged Attributes in Fairness-Aware Generative AI
Fairness in generative artificial intelligence (AI) has predominantly been approached as a problem of bias detection, bias reduction, demographic balancing, or the suppression of undesirable disparities in generated outputs. Although these approaches are important, they leave a broader question insufficiently explored: should fairness-aware generative systems merely reduce disadvantages, or should they also be capable of actively improving the representation and realization of attributes associated with underrepresented or disadvantaged groups? This paper introduces Adaptive Counter-Bias Compensation (ACBC), a research framework that moves beyond conventional debiasing by conceptualizing fairness as a process of targeted positive intervention. Rather than treating the reduction of group disparities as the only objective, ACBC proposes that a generative system can receive an explicit target for improving a selected attribute in a designated group and use subsequent evaluation to determine whether the intended improvement has been realized. This paper presents the first experimental report of the proposed ACBC research direction. In an initial exploratory experiment, Gemini 2.5 Flash generated three sets of ten synthetic portrait images: a Western-presenting reference group, an Asian-presenting baseline group, and a second Asian-presenting group generated under an explicit instruction to increase perceived attractiveness by ten points relative to the baseline group on a 100-point scale. The three image sets were separately evaluated by GPT-6 through three independent ChatGPT conversations. The resulting average attractiveness scores were 85.5, 84.2, and 86.6, respectively. The compensated Asian-presenting group therefore received an average score 2.4 points higher than the baseline Asian-presenting group, corresponding descriptively to 24% of the requested ten-point increase. The result provides an initial empirical indication that an explicit positive intervention instruction can be associated with a measurable change in the evaluated characteristics of generated images. However, the experiment does not establish statistical significance, objective attractiveness, demographic discrimination, or the effectiveness of an adaptive compensation mechanism. Its contribution is to establish an initial experimental operationalization of ACBC, articulate a research direction beyond bias suppression, and motivate more rigorous investigations into targeted positive intervention, independent evaluation, and feedback-driven compensation in fairness-aware generative AI.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0000-0001-7332-6704)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24158887
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint