by the CAQDAS Networking Project
The CAQDAS Networking Project aims to raise awareness and build capacity concerning the appropriate use of digital tools for qualitative data analysis. Part of that effort is accomplished via our free webinar series, where we invite developers and users of qualitative tools to present work-in-progress and discuss the process and implications of their use.
We were delighted to host Eileen Irvin and Marzia Azarbadegan from Ipsos UK to present the first CAQDAS webinar of 2025, titled Marginal Moments: Can Gen-AI help better understand the general election? See below for more information about Eileen and Marzia.
This post summarises their webinar. Watch the full recording and check out the Ipsos UK Marginal Moments project website for more information.
What “Marginal Moments” is all about
The “Marginal Moments” project aimed to analyse how undecided voters engaged with election information during the July 2024 UK general election campaign across five constituencies. The project aimed to explore how AI Personas could provide insights into voter behaviour and engagement with election information over time as the election campaign evolved, particularly in marginal constituencies.
Five AI Personas were created based on demographic data and local voting trends to simulate individual voter perspectives, and to investigate their reactions to materials – such as party manifestos, campaign launch speeches, and local candidate descriptions – presented to them during the election campaign.
The AI Personas were not expected to replicate real voter behaviour exactly; rather, they were designed to provide insights into broader trends and individual perspectives based on demographic data. The use of AI allowed for engagement with extensive materials that real voters might find overwhelming, and it aimed to uncover hidden or unconscious perspectives that humans might hesitate to share, thereby leveraging AI’s lack of social perception concerns.
The process of creating the AI Personas
The AI Personas were created using Ipsos’s in-house AI assistant, which utilized large language models. The Personas were based on demographic data from five constituencies, incorporating social, political, and economic backgrounds, as well as personal details, to create diverse and relatable characters. This approach aimed to ensure a variety of perspectives while maintaining consistency in their responses throughout the election period.
The AI Persona’s are shown below.
Insights about the general election from the use of AI Personas
The AI Personas showed varied responses to party manifestos, reflecting strategic voting considerations, with some leaning towards Labour and others towards Conservative. They exhibited differing levels of engagement and satisfaction with the political parties based on the materials presented, with some showing a shift in their voting intentions as they processed the information.
The project revealed that local candidate descriptions significantly influenced the Personas’ decisions, contrary to typical polling insights.
The AI Personas were asked to creatively interpret the manifestos, for example, to describe them as films. This demonstrated their ability to capture the tone and messaging of the political parties in a unique way, providing insights that might not emerge from customary human analysis.
Overall, the project showcased the potential of generative AI in understanding communication strategies in political contexts.
Challenges with using AI Personas
The project highlighted challenges in generating realistic images for the Personas and the need for consistent prompts to maintain their character throughout the research. There were also some issues with model bias, for example, assuming that the Chloe Persona, a 19-year-old, could not engage with political debates. There were some contradictory outputs which emphasises the importance of context and detailed descriptions when prompting, to generating relevant responses.
Discussion with webinar attendees
The session finished up with Sarah Bulloch facilitating a Q+A session with the webinar attendees. The following questions were posed and answered by Eileen and Marziah. Head over to watch the recording to hear their answers.
- To what extent does the sequential AI prompting used emulate a customary interview encounter, or do you think it’s a very different kind of interaction?
- The issues encountered with generating a suitable image for the Chloe Persona brings to mind the bias that we’ve all heard about within Large Language Models. Did you see any other examples of biases coming through? And if so, how did you handle it?
- What was said by the AI Personas about the Conservative Party when they were asked to describe each manifesto as a film?
- What is the usefulness of using AI Personas for the future of political and electoral polling? And does this highlight any particular dangers or benefits?
- For those of us who are new to this field, what is the difference between synthetic respondents and Personas?
- Do you think that these types of Personas would ever help us to predict unexpected or surprising marginal events which real life political focus groups may be unlikely to bring out?
- When humans engage with election campaigns, as well as the official documentation that you showed the AI Personas, are all kinds of informal and unofficial background influences, such as the media and so on. What sort of thinking have you done around how the Personas might be able or not to take that stuff into account, and how much do you take that into account generally when investigating election behaviour?
- How do you think these Personas manage misinformation?
- Do you think the “stickiness” of large language models might get better as they are trained on more data, or do you think that’s kind of a characteristic of them?
- What is Ipsos UK doing with AI more generally at the moment? Are there ever situations where clients ask for things that you don’t feel are appropriate to do and if so, in what kinds of circumstances? So just a bit of context about how this is fitting to you in your work more broadly.
- What sort of vision do you have in the future for the Personas, in terms of how you’d like to be able to use AI going forward? What future does this have for Ipsos and this kind of work more broadly?
For a full list of previous CAQDAS webinars, including links to the recordings, and to register for upcoming webinars, visit our website. All webinars are Free to attend, and registered participants get the recording links and any associated materials afterwards.
If you have an idea for a CAQDAS webinar, or would like to deliver one yourself, get in touch with us at caqdas@surrey.ac.uk
We are also always actively seeking guest authors for our blog.
Webinar Presenters:
Eileen Irvin is Co-Head of the Ipsos UK Survey Research Methods Centre, where she specialises in high quality, large scale random probability surveys. Marzieh Azarbadegan is a Senior Research Executive at Ipsos UK, specialising in research focussed on education, children, and families. Both Eileen and Marzieh have particular experience in research involving innovative methods and the use of Artificial Intelligence (AI) in the research process.
AI use: This post was created collaboratively. The AI-generated webinar transcript was summarised using CoLoop, producing a thematic and a conversational summary. These were then reviewed and edited by Christina Silver, and cross-checked with the webinar recording to ensure accuracy. Fairly extensive edits were made in this process. The draft blogpost was then checked and adjusted by Eileen and Marzia until they felt it reflected the webinar. The result is what you read above.
