TL;DR
A working paper from Harvard Kennedy School, written by Ed Balls, the former shadow chancellor, with Dylan Evans and Nyasha Weinberg, who both previously worked in the civil service, tests how well public generative AI tools handle real policymaking tasks. Chris Smyth, who edits public policy coverage at the FT, highlighted it in the paper’s Inside Politics newsletter. On his reading, AI is strong at summarising evidence and linking datasets, but produces bland, spin-led proposals that miss political trade-offs.
What the paper tested
The authors ran “policy simulations” using publicly available tools, first across typical tasks in three broad policy areas, then in depth in regional growth, where they have expertise. They also tested how expert checking of outputs and the choice of source material changed the results. The paper is dated 19 October and describes its findings as indicative.
Their summary says generative AI offers “significant value” for some functions, provided its use is “specific, deliberate, and targeted”. It does not remove the need for deep policy expertise, and they warn of significant risks without meaningful human engagement with what the tools produce.
The FT’s reading
Smyth singled out the weaknesses. When the tools were set difficult problems, NHS productivity and the supply of housing among them, he wrote, the AI absorbed government spin and turned out “policy slop”, technocratic ideas that are “beige” variations on existing approaches. It struggled with political trade-offs, failed to come up with anything genuinely new and was poor at judging which policies work.
He set this against a Whitehall keen on AI’s savings, noting that some permanent secretaries talk of slimming the civil service sharply and that AI is due to be a central theme when Antonia Romeo, the cabinet secretary, reviews how Whitehall works. His irony: a transformational technology could deepen the “warm bath of managed decline” that Keir Starmer once criticised.
Looking Forward
The authors call for more direct experimentation with government, alongside workforce, procurement and governance changes, and stress keeping sovereignty over internal data. In our view, the finding most useful to UK organisations outside government is the split itself: AI tools speed up the evidence-gathering stage, while judgement about trade-offs and what is politically or commercially workable remains a human job. The welfare reforms Smyth was writing about, informed by Alan Milburn’s review of young people not in work or education, will test that.