This study investigates two fundamental questions about the relationship between artificial intelligence and democratic governance. First, how and under which circumstances does AI-generated political information affect citizens' perceptions of elections and democracy? Second, what can be done to mitigate the negative impact of AI on democracy? Understanding whether particular institutional arrangements or contextual factors can buffer against democratic harm is essential for developing effective regulatory responses.
To address these questions, the experimental design manipulates three theoretically-motivated dimensions to assess their independent and interactive effects on democratic perceptions. The first dimension examines the investigating authority: who conducts the investigation into AI-generated content may significantly affect public trust in the findings. We compare three types of authorities representing different institutional positions—electoral management bodies (Electoral Commission), civil society organizations (independent fact-checking organizations), and private sector platforms (a social media platform). The second dimension manipulates the prevalence of AI-generated false information. We vary whether false information "dominated" online content (high prevalence), was found in "a moderate amount" (moderate prevalence), or was "only minimally found" (low prevalence). The third dimension examines voter identification ability: citizens' capacity to recognize AI-generated content may serve as a protective factor against democratic harm. We manipulate whether "almost no voters," "some voters," or "most voters" could identify AI-generated political content.
The experiment employs a 3×3×3 factorial between-subjects design, yielding 27 unique experimental conditions. This full factorial design allows us to estimate main effects for each dimension as well as all two-way and three-way interactions, providing insight into how these factors operate independently and in combination. Respondents are randomly assigned with equal probability to one of the 27 experimental conditions using the survey platform's randomization function, which occurs at the individual level after respondents complete baseline demographic and attitudinal questions.
The experimental stimulus consists of a hypothetical vignette describing a post-election investigation into AI-generated content. The vignette structure holds constant the basic scenario—a federal election in a hypothetical country followed by an investigation—while systematically varying the three experimental dimensions.
The vignette reads:
"We want to ask you about a hypothetical situation about the use of AI in election campaigns. Imagine that a federal election took place in a hypothetical country. After the election, [INVESTIGATING AUTHORITY] conducted an investigation into the use of AI to generate political content during the election campaign. The investigation found that AI-generated false information [PREVALENCE LEVEL]. Additionally, the investigation revealed that [VOTER IDENTIFICATION LEVEL]."
Quota allocations :
1) <Investigating authority>:
- the Electoral Commission – 33%
- independent fact-checking organizations – 33%
- a Social Media Platform – 33%
2) <Prevalence Level>:
- dominated online content related to the election. – 33%
- was found in a moderate amount of online content related to the election. – 33%
- was only minimally found in online content related to the election. – 33%
3) <Voter identification level>:
- Almost no voters could identify AI-generated political content – 33%
- Some voters could identify AI-generated political content. – 33%
- Most voters could identify AI-generated political content. – 33%
This hypothetical framing allows us to isolate causal effects while avoiding potential confounds from respondents' prior knowledge or partisan attachments to real elections.
The experiment is embedded within the December 2025 ANUpoll. The particular wave of the survey series will be conducted on the Online Research Unit’s (ORU’s) Australian Consumer Panel. The survey commenced with pilot data collection of around 70 respondents on the 9th of December, 2025. It is expected that data collection will be completed by mid-December, with an eventual sample size of around 3,500 respondents.
The target sample is 1,000 respondents aged 18 to 24 years, and the remaining 2,500 respondents aged 25 years and over.