Publication
Unveiling biases in LLMS: a cross-model study on fairnessand utility
| Summary: | This study investigates biases in large language models, focusing on text-to-text and text-to image systems, to assess fairness, reliability, and utility across model generations. Through replication and extension of prior research, findings highlight persistent biases in knowledge based prompts stemming from skewed datasets, while reasoning-based prompts reveal systemic inequities linked to context. Larger architectures improve accuracy but risk amplifying biases. Industry interventions in text-to-image models show modest progress in fairness but fall short of demographic equity. A trade-off analysis underscores the critical balance between domain-specific applications and ethical AI deployment, providing actionable insights for advancing equitable AI systems. |
|---|---|
| Main Authors: | Rebuzzini, Davide |
| Subject: | Investments Bias LLMS Machine learning Perceptions bias Ageism Asset allocation Portfolio management Data analysis |
| Year: | 2025 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade Nova de Lisboa |
| Language: | English |
| Origin: | Repositório Institucional da UNL |
| Summary: | This study investigates biases in large language models, focusing on text-to-text and text-to image systems, to assess fairness, reliability, and utility across model generations. Through replication and extension of prior research, findings highlight persistent biases in knowledge based prompts stemming from skewed datasets, while reasoning-based prompts reveal systemic inequities linked to context. Larger architectures improve accuracy but risk amplifying biases. Industry interventions in text-to-image models show modest progress in fairness but fall short of demographic equity. A trade-off analysis underscores the critical balance between domain-specific applications and ethical AI deployment, providing actionable insights for advancing equitable AI systems. |
|---|
Funded activities
No funded projects found.