文章以英国为主要案例，探讨 AI 是否能让公共服务更有效率。工党部长希望更广泛使用 AI 能带来更好、更快、也更便宜的服务，而 NHS 一份草案提案甚至表示，AI 可能把生产力提高到足以「完全取代」某些职位。同时，文章也提出一个更难的问题：要如何判断 AI 是否真的提升了公共部门的表现，尤其是在那些并非以市场价格出售、因此很难定义生产力的服务中。

目前的证据喜忧参半，而且往往只涵盖有限范围。National Audit Office 发现，到 2024 年初，几乎四分之三的政府部门和国家机构不是已经在部署 AI，就是正在试行 AI；但 Ada Lovelace Institute 表示，关于其有效性与影响的证据却意外地缺乏。其简报批评一些研究只聚焦于第一层级的时间或成本节省，忽略了技术的终身成本，也忽视成果如何取决于工作流程、资料系统、文化、员工技能与实施方式。例子包括 AI 转录帮助社工节省时间，其中一人表示，这把他的个案负荷能力从每月约 6 人提高到超过 25 人；但其他人因系统不相容而几乎没有看到好处。

文章的主要警告是，新增的需求可能会抵消这些收益，而这些需求本身就是由公众创造的。据报导，英国政府对 AI 相关生产力节省的自估，已从每年 £45bn 下修一半至 £23.6bn。在规划方面，政府的 Extract 工具能把一份提交文件的处理时间从超过 1 hour 缩短到 3 minutes，但 AI 辅助的反对工具可能让提交量成倍增加，迫使政府用自己的 AI 系统来应对。类似模式也出现在 Freedom of Information 申请与基于 AI 的 GP 预约系统中：更容易，或更令人沮丧的存取方式，都可能产生额外工作，或把人导向其他地方。更广泛的含意是，AI 能在特定用途上提供帮助，但它不会让公共服务整体上提升 20 percent 的效率，因此，对效益与伤害进行更好的衡量至关重要。

The article examines whether AI can make public services more efficient, using the UK as the main case. Labour ministers hope wider AI use will deliver better, faster, and cheaper services, and a draft NHS proposal even suggests AI could boost productivity enough to “completely substitute” for some roles. At the same time, the article asks a harder question: how to tell whether AI is actually improving public-sector performance, especially when productivity is difficult to define in services that are not sold at market prices.

Evidence so far is mixed and often narrow. The National Audit Office found that almost three-quarters of government departments and national bodies were either deploying or piloting AI by early 2024, but the Ada Lovelace Institute says there is a surprising lack of evidence on effectiveness and impact. Its briefing criticizes studies that focus on first-order time or cost savings, ignore lifetime technology costs, and overlook how outcomes depend on workflows, data systems, culture, staff skills, and implementation. Examples include AI transcription helping social workers save time, with one saying it raised caseload capacity from about 6 people to more than 25 a month, while others saw little benefit because their systems were incompatible.

The article’s main warning is that gains can be offset by new demand created by the public itself. The UK government’s own estimate of AI-related productivity savings was reportedly revised down by half, from £45bn per year to £23.6bn. In planning, the government’s Extract tool cuts one submission from more than 1 hour to 3 minutes, but AI-assisted objection tools may multiply submissions and force the state to counter them with its own AI systems. Similar patterns are seen in Freedom of Information requests and AI-based GP booking systems, where easier or more frustrating access can generate extra work or push people elsewhere. The broader implication is that AI can help in specific use cases, but it will not make public services 20 percent more efficient across the board, so better measurement of both benefits and harms is essential.