The Prime Minister just called for a complete overhaul of how the British state operates, which puts pressure on the civil service to keep up with rapid tech changes, especially around generative AI. It’s clear that AI will change policy-making, but the real question is how officials can leverage it to help citizens.
Some aspects of policy-making won’t change much for now. Officials will still rely on their judgment to balance various interests within Whitehall. However, in other areas, the shift will be immediate and obvious. Tools like Redbox can significantly speed up how quickly a minister can get up to speed on new topics. Instead of just leaning on officials for information, they can now consult a large language model (LLM). This shifts how information flows to ministers.
LLMs will also reshape how policymakers create policies. They can now pull together existing evidence and suggest interventions to meet specific goals. LLMs, especially Redbox, are already making waves in Whitehall, with over 1,000 Cabinet and Department for Science users benefiting from its ability to summarize policy documents. The government claims another tool, Consult, can process public consultation responses a thousand times faster than human analysts. A demonstration at the 2024 civil service Policy Festival showed Redbox evaluating a National Grid report and presenting ideas from Ofgem on improving it.
But LLMs aren’t without limitations. While they’ve improved rapidly, they can still provide incorrect or biased information—a problem known as hallucination. Their summaries might overlook vital context that human expertise would catch. For instance, designing policy for hospital efficiency requires not only data but also an understanding of how the healthcare system works in practice. LLMs often stick to standard responses and may miss innovative ideas, particularly in swiftly evolving fields like AI.
Officials must be cautious about over-relying on LLMs since incorporating their outputs into policy advice can lead to bias, potentially overlooking a minister’s political stance.
Despite these challenges, the potential gains from LLMs are significant. In a future where AI assists policy-making, the key role of officials will be to infuse their unique knowledge into the process. They’ll likely focus on editing and refining LLM drafts—checking for errors and biases—something current policy makers already do. Another aspect will involve layering their insights over LLM outputs, perhaps even pushing the model to explore bolder ideas. Time saved by automating repetitive tasks could allow officials to gather real-time insights vital for crafting effective policies.
Integrating LLMs into the government landscape might also complicate skill development. As these tools become more prominent, gaining domain expertise could prove more challenging. LLMs excel at gathering and summarizing information—skills that newer policy makers generally need to develop. This creates a dilemma; the very tasks LLMs simplify are where officials usually learn the ropes.
Civil service has a couple of choices. They can keep some fundamental tasks for junior staff to help them build that necessary expertise. Alternatively, they can rethink how officials gain knowledge, perhaps by sending junior officials into frontline roles to experience state operations firsthand.
The civil service might find it beneficial to protect certain tasks while also trialing innovative methods for skill acquisition. Balancing between traditional solutions and new approaches could also help address issues like high turnover by allowing officials to stick to a policy area longer, fostering greater expertise.
Navigating the complexities of policy-making is challenging but essential. Blending human skill with LLM capabilities can lead to a process that’s not only more efficient but also more attuned to citizens’ needs. It’s crucial for the civil service to guide this technology adoption thoughtfully, maximizing benefits while managing risks.