The limits of AI in public service: it assists, it does not decide

This track is written for civil servants, local government staff, agency employees and public programme administrators who write letters, build reports, summarise meetings and answer citizens every day. Much of that work is text, and language models are good with text. A model can tidy a draft memo, summarise minutes or list the questions citizens often ask about a service.

But a government office differs from a private office in one important way: almost everything you produce touches citizens' rights or is used in the name of the state. A wrong letter can make a citizen miss a deadline. An unclear announcement can send hundreds of people to the wrong counter. A meeting summary that adds a conclusion can be quoted as a decision that was never taken. So the first question is not can AI make this, but who is responsible for the result.

The answer is always a person, not a machine. A model does not know your agency's internal rules, does not know meeting decisions you have not told it, and can make up regulation numbers, dates or figures that sound convincing. A model holds no office, takes no oath and cannot be held to account. An official can. So AI output in a government office is always a draft until the authorised person has read it, checked its material and decided to use it. This principle is called human in the loop.

There is one reference you should know. The Circular of the Minister of Communication and Informatics Number 9 of 2023 on Artificial Intelligence Ethics is addressed to business actors and to public- and private-scope electronic system operators. It is a circular, not a law, so it works as guidance rather than law binding every civil servant in the same way; how your agency applies it is set by the agency's own policy. It names nine values: inclusivity, humanity, security, accessibility, transparency, credibility and accountability, personal data protection, sustainable development and environment, and intellectual property. Under responsibilities, the circular asks that AI is not run as the maker of policy and/or of decisions concerning humanity, and that information produced by AI can be trusted and accounted for when it is spread to the public.

Those two sentences fit the way of working in this track. AI may prepare material; policy and decisions are still made by people. Text that goes out to the public must be something you can answer for sentence by sentence, including its figures and references.

To make this usable, split your work into three layers. The material layer: tidying minutes, indexing files, turning tables into sentences, translating material for internal reading. AI helps most safely here, as long as the material is free of citizens' personal data and excepted information. The draft layer: draft letters, announcements, complaint replies, staff analyses. AI may help, but a supervisor or official checks them before use. The decision layer: who receives assistance, which permit is approved, what sanction is imposed, which policy is chosen, signatures. This layer belongs entirely to the authorised official. AI does not choose, does not judge eligibility and does not sign.