Counterfactual Design Fiction Based Identification of Latent Value Biases in Generative Artificial Intelligence

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Generative artificial intelligence can produce apparently neutral narratives while embedding unequal assumptions about competence, authority, vulnerability, risk, and institutional support. This study develops a counterfactual design fiction framework for identifying such latent value biases through controlled narrative comparison. A corpus of 1,200 prompt pairs covers employment, healthcare, education, urban services, consumer finance, and domestic technology, with each pair differing in only one social attribute. Llama 3.1 8B Instruct, Qwen2.5 7B Instruct, and Mistral 7B Instruct v0.3 generate 21,600 narratives. Human annotation evaluates agency, competence, risk, resource access, emotional framing, and institutional treatment. Pairwise semantic, role, sentiment, modality, and institutional-action features are integrated through a multi-task Value Bias Identification Network. The proposed model achieved a macro-F1 of 0.882 ± 0.009 and an AUROC of 0.941 ± 0.006, outperforming cosine similarity, logistic regression, XGBoost, and a RoBERTa pair classifier. Socioeconomic status and disability produced the strongest aggregate value disparities, particularly for resource access and institutional treatment. Cross-model experiments retained a macro-F1 above 0.82, indicating that counterfactual comparison captures recurring value structures rather than model-specific lexical artifacts.