What AI can, and can’t, do for government

Welcome. This course is about practical decisions, the kind you might face this month. We’ll use examples from the Republic of Nyota, a fictional African country with a data protection law, a national digital ID and a government keen to modernise.

What AI is good at

For government purposes, AI is a tool that finds patterns in data and uses them to produce outputs: predictions, classifications, summaries, translations or drafts. It does well when:

• the task is repetitive and high-volume, such as sorting thousands of letters by topic;
• there is good, relevant data about the situation;
• a mistake can be caught and corrected easily;
• a person stays responsible for the decision.

Where AI goes wrong in government

• Decisions about people’s rights and entitlements. Benefits, licences, permits, policing. Errors here harm real people, and citizens usually can’t go elsewhere.
• Thin or biased data. If the data mostly describes urban, connected people, the system will serve rural and offline people worse.
• Problems that aren’t really about prediction. If a service is slow because of staffing or paperwork, AI won’t fix it.
• Tools built elsewhere. A system trained in another country may not understand local languages, names or conditions.

A simple way to think about it

Good fit	Use with care	Not a job for AI alone
Sorting and routing letters or complaints	Flagging tax returns for a human to review	Deciding who gets social support
Translating public information into more languages	Chatbots answering citizens’ questions	Deciding who is arrested or searched
Summarising public consultation responses	Predicting where services are needed	Rejecting a permit without a person looking

Key points

• AI is strongest at high-volume, correctable tasks with good data, with a person in charge.
• Be most careful where decisions affect people’s rights, and where your data leaves people out.
• Start by asking what problem you’re solving, not which tool to buy.

Try this: name one task in your department that’s a good fit, and one that isn’t. You’ll use them in the next lesson.

Practice

Optional. Nothing here is marked, and your results are private to you.

Sort and classifyPractice · not marked

Is this a job for AI?

Sort each government task by how suitable it is for AI.

Good fit · Use with care · Not a job for AI alone

1. Sorting incoming letters by topic so the right team answers them
2. Translating a public health notice into more languages, checked by a fluent speaker
3. Summarising thousands of public consultation responses
4. A chatbot answering questions about licence fees
5. Flagging tax returns for an officer to review
6. Deciding which households receive social support
7. Choosing who is stopped and searched at a checkpoint
8. Rejecting permit applications automatically

Show answers

1. Good fit: High volume, and mistakes are easy to spot and fix.
2. Good fit: Widens access, with a person checking before publishing.
3. Good fit: Saves time, as long as staff read a sample to check the summary.
4. Use with care: Helpful, but wrong answers can cost people money; offer a way to reach a person.
5. Use with care: Acceptable if officers decide and the flags are checked for unfair patterns.
6. Not a job for AI alone: Decides an entitlement; a person must decide, with reasons and appeal.
7. Not a job for AI alone: Affects liberty and risks discrimination.
8. Not a job for AI alone: A significant decision that people have the right to have reviewed by a person.

Your results aren't saved. Sign in to keep a private record of your plays.