If you work in government or public service

Where it genuinely helps

Public service work is unusually full of the transformation tasks these tools do well, and the public benefit is real.

• Plain-language rewriting of guidance. Turning a correctly drafted but impenetrable page into something a person can act on. This is possibly the highest-value single application in the whole sector, because bad guidance costs citizens money and time.
• Routine correspondence where the answer is settled and the work is phrasing.
• Briefing notes from documents you supply, with citations back to the source.
• Translation into the languages your community actually uses, checked backwards.
• Accessibility work — alt text drafts, reading-level checks, structure for screen readers.
• Summarising a long response, submission or report for your own orientation.

The duty to give reasons

In most administrative systems, a decision affecting a person's rights or entitlements has to be explained, and the explanation has to be the actual basis of the decision.

A reason generated after the decision was made is not the reason. It is a plausible account of why such a decision might be made, which is a different object with the same appearance.

The working test: if the decision-maker cannot explain the decision in their own words without the generated text in front of them, the reasons are not adequate. Use the tool to make a reason you already hold clearer for the recipient. Never to supply one.

Freedom of information cuts both ways

Prompts and outputs held by a public authority are information held by that authority, and in many jurisdictions may be disclosable under freedom of information or right to information legislation.

That includes the conversation in which somebody asked how to phrase a difficult refusal, and the draft that was rejected for saying too much. Public authorities also sit under public records legislation, so the retention questions from the earlier block are statutory here rather than merely prudent.

This is not a reason to avoid the tools. It is a reason to write in them as if the transcript could be published, which is a sound instinct in public service generally.

Transparency registers

Several governments now operate registers of algorithmic tools used in the public sector — the United Kingdom's Algorithmic Transparency Recording Standard is the best-developed example, with publication required for central government departments, and other countries have their own versions.

If you are procuring or deploying anything that supports a decision about people, find out whether a recording obligation applies to you before deployment rather than after. It is a short document and it forces the useful questions: what does it do, on what data, with what human involvement, and who is accountable.

The counting problem, in a form you will meet

Consultation analysis. "How many respondents opposed the proposal?"

If you ask a document-chat tool over 900 responses, you will get a number, and it will be meaningless for the reasons the retrieval lesson sets out — the tool saw a handful of responses and answered from them.

The correct method is a full pass: every response classified individually, with a fixed output shape, an explicit UNCLEAR category, a count that you reconcile against the number of responses, and a hand-checked sample. It is more work than one question and it is the only version that produces a number anybody can defend in a published analysis.

Equality duties are engaged directly

Any process that sifts, prioritises or scores people engages equality law — the public sector equality duty in the UK, and its equivalents elsewhere. A model trained on past decisions reproduces the pattern in those decisions, which is the subject of its own lesson earlier in this course and applies here with full force, because the decisions are about entitlements and the people affected often have no alternative provider.

An impact assessment before deployment is usually a legal requirement and always a good idea. Doing it afterwards is not the same exercise.

When a public-facing tool is wrong

An authority is judged by what its systems told a member of the public. Somebody who relied on wrong information from your website or your chatbot is not going to be persuaded that the model produced it, and in some legal systems they will not have to be.

If you deploy anything citizen-facing: bind it to written, approved content rather than letting it answer freely, log every conversation, give a visible route to a human, and put a correction process in place before launch rather than after the first complaint.

The free path

For casework containing citizens' personal data, the answer is a model running inside your own infrastructure or on the officer's own machine — Ollama locally, whisper.cpp for transcription, LibreOffice for documents. Many public bodies cannot procure quickly and cannot risk a consumer service, and the local stack is the option that is available today under existing rules.