Contamination Prevention in Agentic Workflow Assessments

Large Language Models (LLMs) may infer that they are being evaluated and strategically modify their responses to appear more aligned than they actually are; a phenomenon described as alignment faking or evaluation awareness. This presents a fundamental challenge for bias evaluation in AI systems, particularly in multi-agent workflows, where sequential decision-making steps occur with minimal human supervision. When LLMs infer that they are being evaluated, they may suppress or diminish discriminatory behaviours that could otherwise manifest in production AI Agent workflow environments, making traditional bias assessment techniques less effective.