The irony is that the plan was either to - somehow, incomprehensibly, vet all 3 million responses - which would at minimum cost $10m - feed it as training data into some LLM, which would almost certainly make tens of thousands of mistakes at minimum.

That would be let alone the many other costs - including any fees to fight the uphill battle to prove the legality of this.

"Move fast and break things" is cute for a prototype when mistakes cost only time and pay dividends in experience. On a scale of government it's like taking a bulldozer to thousands of Chesterton's fences a day. Which is efficient, from a certain perspective...