Conformal Alignment is a framework for ensuring that foundation model outputs align with human values, providing finite-sample, distribution-free guarantees. The method leverages a set of reference data with known alignment status to train an alignment predictor, which then selects new units whose predicted alignment scores exceed a data-dependent threshold. This ensures that the selected outputs are trustworthy. The framework is based on conformal prediction principles and is applicable to any model and alignment criterion. It controls the false discovery rate (FDR) while maximizing the power of selected units. The method is demonstrated in question answering and radiology report generation tasks, where it effectively identifies trustworthy outputs with minimal training on reference data. The approach preserves the informativeness of original outputs and is lightweight, avoiding the need to retrain large models. The framework provides rigorous guarantees for downstream tasks, ensuring that selected outputs are reliable and align with human values.Conformal Alignment is a framework for ensuring that foundation model outputs align with human values, providing finite-sample, distribution-free guarantees. The method leverages a set of reference data with known alignment status to train an alignment predictor, which then selects new units whose predicted alignment scores exceed a data-dependent threshold. This ensures that the selected outputs are trustworthy. The framework is based on conformal prediction principles and is applicable to any model and alignment criterion. It controls the false discovery rate (FDR) while maximizing the power of selected units. The method is demonstrated in question answering and radiology report generation tasks, where it effectively identifies trustworthy outputs with minimal training on reference data. The approach preserves the informativeness of original outputs and is lightweight, avoiding the need to retrain large models. The framework provides rigorous guarantees for downstream tasks, ensuring that selected outputs are reliable and align with human values.