New CRAFT principles guide responsible LLM use in policymaking

Policymakers worldwide are increasingly exploring large language models (LLMs) to improve how they collect, interpret, and draft policy-relevant information. However, the technology carries significant risks: outputs can be fluent but factually wrong, training data may embed systemic biases, sensitive information can leak, and over-reliance can lead to deskilling and dependency. To navigate these trade-offs, researchers Willem Fourie, Gray Manicom, and Tanya de Villiers-Botha introduce the CRAFT principles in a new arXiv paper (2607.15704).

The CRAFT acronym stands for Control, Rigour, Accountability, Fairness, and Transparency. Control means humans remain the final decision-makers, not the model. Rigour demands that LLM outputs are validated against multiple sources and that uncertainty is clearly communicated. Accountability ensures responsibility for policy outcomes stays with people, not algorithms. Fairness requires training data and model outputs to be scrutinized for representational bias. Transparency calls for disclosing when and how LLMs are used in the policy process. Together, these principles offer a structured way for governments to adopt LLMs without eroding public trust.