A German-led research team conducted a systematic study on how anonymizing prompts by replacing personal information with placeholders affects five language models across eleven benchmarks and 6,210 samples, finding task-dependent impacts on output quality.

Stronger models tended to suffer larger average losses on retrieval tasks, implying that higher capability models rely more on specific entity knowledge for those tasks.

contextual details

Warning prompts that tell the model prompts are anonymized did not meaningfully improve performance, indicating current models do not significantly adapt reasoning strategies to such meta-information.

Implications for practice: the privacy-utility trade-off in prompt anonymization is non-uniform and task-dependent; organizations should consider task requirements when applying anonymization, and there is a need for models that are inherently robust to missing personal data or task-aware anonymization.

Anonymization significantly degrades retrieval-augmented tasks: GPT-4o mini’s RGB score dropped from 0.80 to 0.32, and similar collapses occurred for other models, indicating that replacing entities hinders cross-source fact integration.

Reasonable gains were observed on a truthfulness benchmark (TruthfulQA) for several models when personal names were hidden, suggesting removing triggers for memorized associations can reduce falsehoods in some cases.

Five anonymization methods were tested; reversible methods (pseudonymization, randomization, masking) generally preserved more utility than irreversible methods (generalization, redaction), with pseudonymization performing best in several benchmarks.

Regulatory and industry context: the study discusses GDPR-related interpretations of pseudonymization/anonymization, noting evolving guidelines (EDPB, EDPS) and regional policy developments affecting how organizations treat anonymized data in AI workflows.

Key figures and entities: study led by Tobias Deuß er and Max Hahnbück with colleagues from the University of Bonn, Fraunhofer IAIS, the Lamarr Institute, and Microsoft Germany; arXiv version 2 posted September 15, 2026; models evaluated include GPT-4o mini, Teuken-7B, Llama-3.1-8B, Gemma-2-27B, and Qwen2.5-72B; benchmarks include ARC, BIG-Bench Hard, HellaSwag, IFEval, MedQA, MMLU-PRO, MUSR, RGB, TruthfulQA, WMT 2014, and others; RGB specifically highlighted as most affected by anonymization.