Artificial intelligence systems often underperform in new contexts, limiting their reliability in business domains. This study introduces the reference aware delexicalization (RAD) framework, a theory-driven approach that improves artificial intelligence (AI) model performance across different domains without requiring massive computational resources. RAD addresses a fundamental problem: AI models often memorize surface patterns from training data rather than learning transferable reasoning skills. By systematically abstracting domain-specific terms while preserving semantic relationships, RAD enables models to focus on underlying logical structures that generalize across contexts. For practitioners, RAD offers measurable benefits. Healthcare organizations can deploy clinical decision support systems that maintain accuracy when processing records from different departments or institutions. Financial institutions can build fraud detection systems that adapt to emerging threats without extensive retraining. Social media platforms can improve content moderation consistency across languages and cultural contexts. For policymakers, RAD demonstrates that effective AI does not require ever-larger models with corresponding environmental and economic costs. Organizations can achieve robust, adaptable AI systems through principled data augmentation techniques. This finding supports policies encouraging efficient, interpretable AI development over resource-intensive scaling approaches, promoting both technological sustainability and broader access to reliable AI capabilities.
Suntwal et al. (Thu,) studied this question.