An architect is reviewing a Claude-based candidate-screening tool prior to deployment. A stakeholder asserts that because the model was not trained…

Even if an LLM is not explicitly trained on a company's proprietary data, it inherits demographic and societal biases from its massive pretraining corpora. In high-stakes domains like employment screening, these latent biases can manifest as discriminatory outcomes against protected classes. Therefore, it is imperative to rigorously evaluate the integrated system using outcome-disparity metrics across protected groups before deployment, ensuring the application complies with fairness standards and ethical AI guidelines.

A. Transparency (disclosing AI use) is a separate ethical requirement and does not mitigate or measure the actual bias present in the system's decisions.

B. Bias can emerge from proxy variables and semantic associations in the pretraining data, even if explicit protected-class labels are absent from the prompt.

E. A vendor's responsible-use policy is a guideline, not a technical guarantee; the deploying organization remains responsible for evaluating fairness in their specific application context.

1. Anthropic Research (2023)

"Evaluating and Mitigating Discrimination in Language Model Decisions".

2. NIST Artificial Intelligence Risk Management Framework (AI RMF 1.0)

Section 4.1 (Fairness and Bias)

pp. 22-24. https://doi.org/10.6028/NIST.AI.100-1