What AI Can and Cannot Do in Public Service

AI assistants are tools that write text. Some are general, such as ChatGPT, Gemini and Claude. Others are built into office software, such as Microsoft 365 Copilot. This level gives you a working picture of what these tools do well and what they do badly. It also shows how UK public bodies are actually using them, with dates and sources. Later levels cover approved tools, data protection and fact-checking in detail.

How an AI Assistant Produces Text

An AI assistant is not a database or a search engine. Here is what it does, step by step:

1. It was trained on a very large amount of text, from which it learned patterns of language.
2. When you type a prompt, it predicts the most likely next words, one piece at a time.
3. It keeps predicting until it has a complete answer.

The result reads well because the model is very good at language patterns. But a fluent answer is not the same as a checked answer. The model has no built-in step that confirms a fact is true. If the most likely-sounding words are wrong, it can still write them confidently, for example a statute section, a date or a quotation that does not exist. These are often called "hallucinations".

Analogy: think of a very well-read new colleague who answers instantly and never says "I don't know". Their drafts can save you time, but you would not sign off their work unchecked.

What AI Assistants Do Well and Badly

Usually helpful:

• First drafts of letters, emails and briefing outlines
• Summarising a long document you supply
• Rewriting text in plainer English or a different tone
• Suggesting structures, headings and questions you may have missed
• Grouping large amounts of text into themes

Weak or risky:

• Facts, figures, dates and references you have not supplied yourself
• Recent events and your organisation's local context
• Nuance, conflicting views and politically sensitive judgement
• Anything where a decision affects a person's rights, money or safety

Task 1: Sort Your Own Work into Three Groups

Do this now. List five tasks from a typical week. Then put each one into a group:

1. AI can help: low-risk drafting or reformatting that you will review. Example: turning your bullet points into a first-draft team update.
2. AI with caution: useful, but the output needs careful checking against sources. Example: summarising a 60-page guidance document, where you check key points against the original.
3. Must stay human: decisions, judgement and accountability. Example: deciding whether a resident qualifies for support, or signing off advice to a minister or councillor.

Write one reason next to each task. A good test is: "If this output were wrong, who would be harmed, and would I notice?" If the answer is "a member of the public, and maybe not", the task goes in the cautious or human group.

Task 2: Read the Evidence, with Dates

Here are three sourced examples. For each one, note what it shows and what it does not show.

1. Microsoft 365 Copilot cross-government experiment (30 September to 31 December 2024; findings published June 2025). Around 20,000 civil servants in 12 organisations took part. Participants reported saving an average of 26 minutes a day, mostly on email, drafting and meeting preparation.

• What it shows: many staff found the tool useful for routine tasks.
• What it does not show: that every role benefits equally. Savings varied widely: some participants reported no time saved, and the report notes that policy-focused teams struggled with nuance and conflicting sources. Users also raised concerns about incorrect information. Time savings were reported by participants, which is not the same as checking the quality of the work.

2. "Consult", a government-built tool (reported May 2025). Part of the Humphrey suite of tools, it was used on a Scottish Government consultation about non-surgical cosmetic procedures, covering about 2,000 responses. Officials checked and refined the themes the AI identified, and they also reviewed all the responses by hand. The AI and the human reviewers largely agreed.

• What it shows: AI can help sort large volumes of consultation text into themes.
• What it does not show: that it is ready to replace human analysis. The government said more evaluation was needed before wider rollout, and experts stayed involved throughout.

3. Local Government Association "State of the sector: AI" 2025 update (survey run December 2024 to February 2025). About a third of English councils responded. Most of them said they were using or exploring AI, and generative AI was the most common type. Uses included chatbots, caseworker assistants, easy-read documents and translation.

• What it shows: AI is widely being tried in local government.
• What it does not show: what happens in councils that did not respond. "Exploring" is not the same as using a tool safely at scale.

Habit to build: whenever you hear a claim about AI, ask who measured it, when, how, and what was left out?

Your Checklist

• I have sorted five of my own tasks into three groups, with a reason for each
• I can explain why fluent AI text can still be wrong
• I can name one dated UK example and say what it does and does not prove
• I have marked which of my tasks must always stay human

Key takeaways

• AI assistants predict likely words. They write fluently but do not check facts for you.
• They are strongest at drafting, rewording and summarising material you provide, and weakest at facts, nuance and judgement.
• Decisions that affect people, and accountability for them, stay with you.
• UK evidence (the Copilot experiment of 2024, Consult in 2025, the LGA survey of 2025) shows real promise, alongside limits and continued human checking.
• Always ask who measured a claim, when, and what it leaves out.