Marginal Moments: Can Generative AI help us better understand the general election?

Overview

We created five AI personas of undecided voters in key marginal constituencies across England: Penny, Darren, Priya, Rupert and Chloe.

Our ‘undecided’ personas carefully considered and reviewed a variety of stimulus over the course of the campaign. Having changed their minds along the way, all did choose to vote at the election. 

Overall, 3 of our 5 personas voted for the winning candidate in their constituency. Final voting decisions were driven strongly by local candidates and manifesto pledges, contextualised by polling projections in their local area.

The true value of our AI personas was in observing how they engaged with the language, tone and specific pledges of each party.

Key findings: final stages of the campaign 

In the final stages of the campaign, we asked our AI Personas to review local and national polls and make their final decisions. Post-election, we then provided them with the local and national results:

• Our personas were generally surprised by the polling, both locally and nationally, suggesting some lag in the understanding of current politics from the Large Language Models used by our GenAI tool.
• Polling did not impact our candidates’ perceptions of the parties, but did mean they considered tactical voting to ‘keep out’ parties they were particularly concerned about, and to ensure their vote wasn’t “wasted”
• Having been unsure who to vote for at the start of the campaign, all of our personas decided to vote 
• Overall, 3 of 5 of our AI personas voted for the winning candidate in their constituency
• Even where they hadn’t voted for the winner, our personas were generally positive about their new MP, but had mixed in reactions to a Labour government nationally
• Overall, our personas were cautiously hopeful for the future, particularly for the NHS and rise in minimum wage, but more concerned about implications of tax and nationalisation

When thinking about their expectations for the new Labour government as films, our personas suggested films that focussed on people collaborating and working creatively and resiliently in the face of large challenges, such as The Martian and Apollo 11.

I'm hoping for the best, but bracing myself for a bit of a bumpy ride

-Darren 

Key findings: methodological reflections 

As an R&D project, it’s been really useful to unpick what we have learnt about using AI personas for this type of research, and how to get the most from this approach.

• Using one sequential ‘conversation’ for each persona improved authenticity and encouraged consistency in their responses, but also made it harder to test prompts before they were used. Future studies would benefit from a ‘prompt sandbox’ to fully test prompts when ‘live’ in the field. 
• Our individual personas did show variation in their reaction to the same stimulus or prompt. This demonstrates the statistical variability that exists within GenAI. 
• There is a trade off in how prescriptive to be when building AI personas. Greater specificity may introduce false guardrails and inhibit minority views.

Personas were particularly useful for:

• Digesting and reacting to large quantities of information quickly and thoroughly
• Providing creative responses that captured and explained more nuanced perceptions
• Explaining why answers were given, and what aspect of the persona these related to
• Responding to questions with answers of information the persona would feel uncomfortable sharing

It is important to recognise that our personas cannot fully replicate the complexities of human experiences and decision making. They are likely to add most value in supporting innovation and early creative processes and in bringing segmentations to life.

Key findings from second half of the campaign

• Local candidate statements had a large impact on our AI personas, with 4 of 5 of our AI personas preferring a candidate for a party they had not previously preferred
• The AI personas prioritised candidates with connections to the local area and specific links to local interventions
• Reviewing relevant news summaries of the debates had limited impact on our persona’s voting preferences
• When considering the debates, personas tended to focus on the parties they had previously been considering voting for

Our persona’s have been heavily swayed by local candidate statements, while less driven by reporting of the debates. All have changed their mind at least once.

After reviewing campaign materials from her local constituency, Penny was torn between policy pledges and experience.

I think my favourite is a toss-up  … [the Liberal Democrat candidate] seems like he really cares about the community, but [the Conservative candidate] seems like she could actually get things done.

Darren preferred the Conservative candidate in his local constituency, who had a similar job history to his persona.

He's got experience in the army and in business, and he's talking about supporting local businesses and jobs, which is what we need around here. Plus, he's focused on the energy sector.

For Chloe, readings statements from the local candidates reinforce her previous preference for the Greens manifesto, and alleviated some of her concerns about practicalities, as she felt the Green candidate had the most direct, local experience.

The Green candidate seems like a normal, down-to-earth guy who really cares about the things that are important to me, like the environment and the cost of living. … Plus, he's local and understands the issues facing our area.

Key findings from first half of the campaign 

• Election ‘launch’ speeches failed to impress. Our AI personas were uninspired by the visions set out by party leaders early in the campaign. 
• Manifestos did help 4 of our 5 AI personas start to identify with a party. 
• Although they didn’t have all the answers our voters were looking for, manifestos did provide our AI personas with a sense affinity for one party over others, based on common priorities and, on occasion, social norms.
• Our AI personas have a low trust in politicians and the main parties – most shared doubt as to whether parties could deliver on their promises, or  whether they were fully costed.
• National service was spontaneously identified as the least popular policy – unpopular among young AI personas and those with children.
• Our AI personas are a creative and powerful tools for translating and distilling complex policy narratives into something more tangible. This includes bringing manifestos to life with reference to film and songs: from quirky indie comedy to gritty social reality drama.

As we reach polling day, our AI personas will be reviewing what their own local candidates have to say.  Who will they decide to vote for on 4th July!?

Emerging affinity to national campaigns

All our AI personas were undecided on who to vote for on the day the election was called. Despite being uninspired by the launch speeches, all but Chloe had leaned towards a party after reading five manifestos from back to front.

Bringing manifestos to life

One of the strengths of GenAI is the ability to be creative. Having digested all five manifestos, our AI personas shared their ideas on how each manifesto would be depicted as a film and song.

Technical note

Marginal Moments is an AI R&D project, using AI personas (synthetic voters) to explore how election promises are engaging the public. 
We have created five different personas of undecided voters. These personas are based on Ipsos polling on the demographics of undecided voters, and their key election priorities; they have been placed in five constituencies that are likely to be marginal seats across England . Our personas also have an expanded backstory to cover a range of demographics, interests and news sources.

Marginal Moments uses Ipsos Facto, a secure AI assistant that draws on best in class Large Language Models from Google and Chat GTP.
The personas have been set up as sequential AI chats, which allows researchers to continue conversations over time, as each personas learns from its own reactions to the campaign.

Personas are established by providing demographics, information about the persona’s family, work, and life, as well as the political and social context of their constituency, and their main political concerns. 

Once the personas are established, each persona is then provided with consistent prompts, providing information about the election, and asking their opinions, priorities, attitudes and concerns. Additional prompts may then be used to encourage personas to expand, where not enough details were provided.

This work is intended to tell us more about:

The election:

• Communicating broader trends and analysis through the lens of individual voters
• Understand reactions to extensive materials that would not be feasible for participants to read (e.g. multiple manifestos in full)
• Tap into hidden or unconscious perspectives that voters may be unwilling to share

Generative AI

• How similar are AI and genuine voters?
• What are generative AI personas useful for?
• What are the limitations? What cannot be understood without engaging with real people?

To help us do this, our final results will be triangulated against other findings – including polling data, video diaries of real undecided voters, and voting behaviour in the actual constituencies.