When South Africa's draft national AI policy had to be withdrawn after officials discovered its reference list included fictitious sources, the incident underscored an uncomfortable reality for governments worldwide: large language models are powerful but unreliable. That contradiction is the central challenge facing public administrations that want to harness AI's efficiencies without surrendering the accountability that defines democratic governance.
To help civil servants navigate this tension, researchers at Stellenbosch University's Policy Innovation Lab have developed the CRAFT Principles, a five-point framework designed for the responsible use of large language models in policymaking. The principles emerged from a national webinar that brought together experts from philosophy, mathematics, information science, public policy and industry.
"The challenge, then, is no longer whether governments should use AI, but how they should use it responsibly," said Prof Willem Fourie, lead at the Policy Innovation Lab.
What the CRAFT Principles require
The framework is intentionally non-technical. It doesn't prescribe which system to buy or which technical guardrails to install. Instead, it addresses how people working inside government should interact with AI tools day-to-day.
Controllability demands that civil servants retain the ability to complete essential policy work without relying on a model. If the system goes down, the work continues. That principle protects institutional knowledge and prevents overdependence.
Rigour means treating every AI-generated claim as suspect until verified against reliable sources. "Large language models often produce answers that sound convincing or appealing to the user, even when they are wrong," said Dr Gray Manicom, a co-author of the framework.
Accountability puts responsibility back on people. "Even where AI contributes to drafting or analysis, a clearly identified person must remain responsible for the final advice, recommendation," said Fourie. "Accountability remains human, not technological."
Fairness acknowledges that training data skews wealthy and English-speaking, which means many African contexts, languages and lived experiences are under-represented. "If policymakers rely uncritically on AI outputs, these gaps and biases can unintentionally become embedded in public policy," said Dr Tanya de Villiers-Botha.
Transparency demands that citizens know when and how AI has informed government decisions. The greater the AI contribution, the greater the need for openness. For civil servants looking for practical guidance on this topic, the AI for Government resource track offers more detailed operational examples of how accountability frameworks apply in practice.
Where the technology stops
LLMs excel at processing and generating text. They cannot exercise judgment, weigh competing public interests or accept responsibility for consequences said the researchers. Those human qualities emerge through experience, critical discussion and accountable decision-making.
That limitation matters most in government, where policy work requires balancing evidence with democratic values, representing the public interest and making decisions that directly affect lives. The CRAFT framework does not try to solve all those problems. Instead, it sets a baseline for trustworthy use while technology continues evolving.
Fourie also connected the rise note of AI in government to a broader strategic shift. To governments already integrating these tools into their operations, the AI for Policy Makers learning path outlines ways to structure oversight as routine use grows.
Why this matters for government professionals
The question is no longer whether your agency will start using large language models - in most governments, that decision has already been made or is being made right now. The CRAFT principles offer a handful of concrete questions you can raise in your next meeting: Who is accountable for outputs? Have we verified claims against independent sources? What is invisible blind spots in the model? Getting those answers right isn't bureaucratic overhead - it's the difference between a useful tool and a liability that could cost an election, a policy, or public trust.
Your membership also unlocks:
