Governments around the world are increasingly using artificial intelligence (AI) tools to improve policymaking. But when and how they should be used has become the subject of much debate.
Recently, the matter came to the fore when South Africa’s draft national artificial intelligence (AI) policy had to be withdrawn after it was found that the reference list included fictitious sources.
This incident served as a reminder that large language models (LLMs) such as ChatGPT and Gemini are capable but fallible, says Prof Willem Fourie, lead at Stellenbosch University’s Policy Innovation Lab. “The challenge, then, is no longer whether governments should use AI, but how they should use it responsibly.”
AI used wisely can strengthen government
Large language models can be of considerable value to government. They can rapidly collate thousands of pages of policy documents, identify emerging trends, assist officials with drafting reports, summarise public submissions and make complex information more accessible. Used well, they could help overstretched public servants work more efficiently and improve the quality of their work.
Yet the same systems that can save hours of work can also generate fabricated references, reinforce existing biases or present inaccurate information with remarkable confidence. Unlike a search engine, an LLM does not retrieve source material, although some models may make use of search engines. Instead, it predicts a likely sequence of words based on patterns in the data on which it was trained. The result is often fluent and persuasive, but not necessarily true.
This has significant implications for their use in government.
A framework for responsible use
To help address this challenge, the Policy Innovation Lab recently convened a national webinar on the responsible use of large language models in government, in partnership with the Policy and Research Services branch of The Presidency.
The discussion brought together experts from philosophy, mathematics, information science, public policy and industry to examine AI from ethical, technical, operational and governance perspectives.
Rather than ending with another broad discussion about AI’s risks and opportunities, the conversations produced practical outcomes.
Drawing on these discussions, Fourie and two other SU researchers, Dr Gray Manicom and Dr Tanya de Villiers-Botha, developed the CRAFT Principles, a simple, accessible framework designed to guide the responsible use of large language models in policymaking.
The framework is intentionally non-technical. It does not tell governments which AI system to use or which guardrails to put in place. Instead, it focuses on how people can use these tools responsibly.
C is for Controllability
The first principle recognises that people, not machines, must remain in control. Civil servants should retain the ability to complete essential policy work without relying entirely on an AI system. Maintaining this capability protects institutional knowledge, prevents overdependence on technology and ensures that human expertise remains central to government.
R is for Rigour
“Large language models often produce answers that sound convincing or appealing to the user, even when they are wrong,” says Dr Manicom. Every AI-generated claim should therefore be verified against reliable evidence, trusted sources and subject-matter expertise before it influences policy decisions. Rigour means treating AI as a useful assistant rather than an unquestioned authority.
A is for Accountability
One of the defining features of democratic government is that someone is accountable for decisions. That responsibility cannot be delegated to an algorithm. “Even where AI contributes to drafting or analysis, a clearly identified person must remain responsible for the final advice, recommendation or decision. Accountability remains human, not technological,” according to Prof Fourie.
F is for Fairness
AI systems reflect the data on which they are trained. Much of that data originates in wealthier, English-speaking parts of the world and under-represents many African contexts, languages, values and lived experiences, says Dr De Villiers-Botha. “If policymakers rely uncritically on AI outputs, these gaps and biases can unintentionally become embedded in public policy.” Fairness therefore requires officials to ask not only what an AI system says, but also what it may be leaving out.
T is for Transparency
Citizens deserve to know when and how AI has informed government decisions. Transparency does not require disclosing every interaction with an AI tool, but it does mean explaining its role in proportion to its influence on a decision. According to Dr De Villiers-Botha, the greater AI’s contribution, the greater the need for openness and disclosure of relevant details about its use. Transparency builds trust and enables meaningful public accountability.
Technology cannot replace judgement
The CRAFT Principles reflect a broader truth about artificial intelligence, notes Fourie. LLMs excel at processing and generating text, but they cannot exercise judgement, weigh competing public interests or accept responsibility for decisions.
Recent scholarship has argued that while AI may improve the efficiency of producing or analysing evidence for decision-making, it cannot replace the human practices through which sound judgement, expertise and accountability are developed. Those qualities emerge through experience, critical thinking, discussion and responsibility.
That insight is especially important in government, where policymaking involves balancing evidence with values, representing the public interest, navigating uncertainty and making decisions that affect people’s lives, says De Villiers-Botha.
From principles to practice
Artificial intelligence will almost certainly become a routine part of public administration over the coming years. By proposing the CRAFT Principles, SU’s Policy Innovation Lab hopes to shift the focus towards practical governance. “As technology evolves, the values that underpin democratic government must remain firmly in human hands,” concludes Fourie.
For the full document outlining the CRAFT principles, visit this page: https://arxiv.org/abs/2607.15704.