Before It Goes Public: Using AI to Test Government Messages

Before a government message goes public, the public officers working behind the scenes often have to answer a difficult question: Will citizens understand the message the way it was intended?

In January 2021, the UK government published a poster urging people to stay home and save lives. The illustrations showed women home-schooling, cleaning, and caring for children, while the one man pictured relaxed on the sofa. This was COVID-times, so the intended message was about public health. But that wasn’t what the public took away. What the public took away was a message about gender roles and who does the housework.

Photo: Covid lockdown poster published by UK Government in 2021

The poster was withdrawn ↗ (opens in new tab) within days. A misfire like that can affect a government’s credibility and public trust. The government officers who made the poster were not necessarily careless. They simply could not see their own work the way the public would. This is a familiar communications risk for any public officer: a sentence, image, or framing that seems clear internally may land very differently once it reaches citizens.

A team at GovTech has been exploring whether AI personas can help officers spot that gap earlier. The idea is not to predict public opinion, but to give officers an earlier dry run before a draft reaches real citizens.

Here’s what the team built, what early testing revealed, and why this approach is still no substitute for real public feedback.

Why testing a message is harder than it looks

Government agencies typically have established ways of checking how a message will land. Before a major announcement, they run focus group discussions across different demographics to understand ground sentiments. Those sessions remain the most direct way to hear what people actually think. A good focus group can surface perspectives that nobody in the drafting room anticipated, in the words of the person raising it.

The limitation is time. Recruiting participants, coordinating sessions, analysing responses, and translating findings into revised messages can take months. That limits how many times an agency can test and refine a message draft before going live.

Findings may also stay within the team that commissioned the focus group. When another agency later faces a similar communications challenge, it may not be easy to reuse what has already been learned.

For public officers, this creates a practical gap. You may know that a message needs testing but not have the time or resources to test every version as thoroughly as you would like. Communicating in a multi-racial and multi-religious society raises the stakes further. The same sentence can be taken as reassuring to one community but dismissive to another.

Red teaming, borrowed from cybersecurity

In cybersecurity, a red team attacks a system on purpose, looking for weaknesses before a real attacker uncovers them. The value lies in the adversarial posture. A team trying to break something notices what the team that built it cannot.

GovTech’s team applied a similar mindset to policy communications: test the message before the public does.

The team built an app called 'Echo' ↗ (opens in new tab) that probes a draft announcement for blind spots before it goes public. An officer pastes in a draft, and a set of AI personas representing different Singaporean profiles respond with how they might react.

Photo: Example of AI personas in Echo

Unlike a generic AI chatbot response, the personas are grounded in transcripts from real focus group discussions. Their responses trace back to opinions actual Singaporeans have expressed rather than stereotypes the model invented.

The app reviews accompanying visuals too, since an image can misfire as easily as a sentence, as the UK poster showed.

For public officers, the potential benefits are speed and focus. Instead of waiting until a formal focus group to discover where a message may be misunderstood, they can surface possible questions earlier and refine what needs to be tested with real people during focus groups.

The important caveat: these personas do not represent public opinion. They are a way to generate hypotheses, not conclusions.

Two personas, two different reactions

The team tested the app on a sample Budget announcement: a $2 billion Progressive Wage Credit Scheme top-up over three years, co-funding wage increases for lower-wage workers earning up to $3,000 a month.

The SME owner persona pushed back on cost, regulatory burden, and manpower, reflecting concerns employers had raised in earlier focus group sessions.

The critical intellectual persona went somewhere else entirely, asking what changes structurally once the three years of co-funding end.

Both are fair reactions, but they point to different kinds of questions a communications team may need to address. A cost objection might be answered by clarifying the type of support employers receive and when. The structural question is harder, because it asks about the policy rather than the wording.

That distinction matters for public officers. Some issues can be fixed in the wording. Others may need to be explained more clearly, escalated internally, or tested further through focus groups. Running the draft past both personas took just a single pass, well before anything reached a public channel.

Three principles for using AI personas

Three principles have emerged from the work so far:

1. First, a message may not land the way it was intended, but that risk can be spotted earlier. No tool removes the risk that a message lands differently from how it was intended. A dry run changes the cost of detecting that risk. Catching a potential misfire during drafting is much better than realising it after publication.

2. Second, AI personas generate hypotheses, not conclusions. Their output is a set of questions worth asking, not a verdict on how the public will respond. Treating a simulated reaction as evidence of public opinion misuses the tool.

3. Third, the goal is sharper focus groups, not fewer conversations with the public. An officer who runs a dry run first can walk into a focus group with a clearer set of questions to probe, making the time with real participants more productive. The AI-assisted dry run sets up the conversation, but does not replace it.

What this approach cannot do

The team is deliberate about the limits. Simulated personas are not a representative sample, and they cannot substitute for real engagement and feedback.

A persona built from past transcripts reflects what people have said before, making it a useful starting point but a poor guide to how opinions may have shifted since.

The app is currently in early testing stages. GovTech is working with interested agencies to assess how closely persona reactions match responses collected in live focus group discussions, testing predictions against fresh announcements and transcripts the system has never drawn on.

That evaluation matters because public officers need to know how much weight to give the output. A useful tool should not only produce plausible reactions, but should also make clear where human judgement, fresh evidence, and real public feedback are still needed.

Until that work matures, the output is best read as a prompt for further checking.

The team that built the app cautions clearly: the app is not a replacement for talking to real people, and it never will be.

Why this matters for public officers

For any government, the cost of a misunderstood message can be high. A message that misfires can chip away at public trust.

The app was built together with the policy and communications officers who use it, with their feedback shaping it throughout. It also shows what careful AI adoption in government can look like. The value is not about letting AI decide what the public thinks. It's about using AI to help officers ask better questions earlier, sharpen their drafts, and make better use of time with real citizens. That is what responsible AI in government communications looks like.

Public officers can try out the app via WOG AD login at: https://go.gov.sg/echo-tn ↗ (opens in new tab)

For the engineering that went into developing the app, read the full technical write-up on GovTech’s Medium publication ↗ (opens in new tab).

Connect with us 

Subscribe to the TechNews email newsletter ↗ (opens in new tab) 

Join the Tech Kaki user-testing community ↗ (opens in new tab)