Public trust and transparency

Why trust is the real test

A government AI system can work well technically and still fail if citizens do not trust it. People who believe a machine is deciding their case in secret will suspect unfairness, complain more and lose confidence in the service as a whole. People who are told openly what is happening, why and how to challenge it are far more likely to accept the result. Transparency is not a public relations exercise. It is part of making AI use legitimate.

Tell people when AI is used

As a general principle, citizens should know when AI plays a meaningful part in a service they use or a decision about them. Good practice includes:

• Clear notice at the point of contact. On a form, a website, a letter or a helpline message, say plainly that AI is used and for what.
• Plain language in Urdu and English. And in regional languages where the service is delivered in them. Avoid technical terms.
• Honest description of the human role. Say whether a person reviews every case, some cases or none. Do not suggest more human involvement than there is.
• No disguised chatbots. If citizens are talking to an automated assistant, they should be told. They should also be able to reach a person.

The level of detail should match the risk. A tool that helps staff draft internal notes needs less public explanation than a system that helps decide who receives a service.

Publish what systems do

Beyond individual notices, a department can publish a short, plain description of each significant AI system it uses. Some governments keep an algorithmic register for this purpose: a public list of systems and what they do. Whether or not a central register exists, your department can publish its own information. A useful entry covers:

• the name and purpose of the system
• which service or decision it supports
• what data it uses
• what role humans play
• the main risks identified and how they are managed
• who is responsible and how to raise a concern

Check your department's instructions and any government guidance before publishing. Make sure nothing published creates a security risk. Most of this information is not sensitive and publishing it builds confidence.

Explain decisions

When AI contributes to a decision about a person, they are entitled to an understandable explanation. Explainability does not mean publishing technical details. It means being able to tell someone, in plain words, the main reasons for the outcome and what information it was based on. If your department cannot do this for a particular system, that system should not be used for decisions about individuals.

Handle complaints well

Complaints about AI-assisted decisions are valuable information. They are often the first sign that a system is failing. Make sure that:

• citizens can complain through normal channels, without needing to know technical terms
• complaints that mention AI or automated decisions are tagged so they can be tracked
• a human with authority reviews each complaint and can change the outcome
• patterns are reported to the system's owner, not just resolved case by case
• the citizen is told what was found and what was done

This is the practical side of redress, covered in module 3. A spike in complaints or complaints clustering in one group or district, should trigger a review of the system itself.

When something goes wrong

At some point an AI system will make a mistake that becomes public. How the department responds matters more than the mistake itself. Prepare in advance: know who will speak, what you will say and how affected citizens will be contacted. Acknowledge the problem, explain what happened in plain terms, correct the affected decisions and say what will change. Defending a failing system damages trust far more than admitting a fault.

Try it

Write a short public notice, in plain English, for an imaginary AI tool that helps sort citizen complaints so that urgent ones reach staff faster. Cover what the tool does, what it does not do, the human role and how to raise a concern. Keep it under 150 words. Then ask a colleague outside your field whether they understand it.

Check yourself

1. Give three elements of good practice in telling people that AI is used.
2. What should a published description of an AI system include?
3. Why are complaints about AI-assisted decisions especially valuable?

Answers: (1) Any three of: notice at the point of contact, plain language in the right languages, honest description of the human role, no disguised chatbots. (2) Purpose, the service it supports, data used, human role, main risks and their management, responsibility and how to raise a concern. (3) They are often the first sign that a system is failing; patterns can reveal problems that individual cases hide.