This blog post is a reformatted and shortened post based off of Chris Bowdon's personal blog post about AI polling simulations and ultra-processed foods. Chris Bowdon is Polecat's Chief Technical Officer, with interests in AI and Big Data as well as maths and code. You can find the full length blog post from Chris on AI polling and ultra-processed foods here.
Public concern about ultra-processed food (UPF) has grown rapidly in recent years, driven by mounting health research and increasingly prominent media coverage. But measuring how these concerns evolve in real time—particularly in response to media narratives—remains a complex challenge for policymakers, analysts, and communicators alike.
In a recent exploratory project, Chris Bowdon, Polecat’s Chief Technology Officer, tested a new approach. Bowdon combined traditional opinion polling with AI-generated sentiment simulation to understand not just what the public thinks—but how media exposure might shape those perceptions over time.
A Fresh Perspective on Simulated Polling
The use of large language models (LLMs) in simulated polling is not a new concept. In fact, researchers and technologists have been experimenting with LLMs as stand-ins for human respondents across various domains. The innovation in this project, however, lies in how the simulation was applied—not simply to replicate public sentiment, but to model the effects of media exposure on demographic opinion.
The experiment focused on monthly polling data published by the UK Food Standards Agency (FSA) through YouGov, which tracks public awareness and concern about UPFs. In parallel, a suite of AI-generated personas—designed to mirror real demographic profiles—were “exposed” to different quantities and styles of media coverage related to UPFs before being asked polling questions. This allowed for comparison between actual survey results and simulated sentiment conditioned by curated media inputs.
Linking News Coverage to Public Perception
At the core of this approach is a novel hypothesis: that LLMs, when exposed to specific types of media narratives, may exhibit shifts in simulated opinion like those observed in human respondents. In other words, if a particular month sees a spike in alarming UPF headlines, the AI personas conditioned with that content might show increased concern—mirroring a pattern seen in real-world survey data.
Preliminary results suggested a compelling correlation between the intensity of media coverage and the direction of sentiment shift among the AI models. This creates the potential for a kind of explainable proxy, helping analysts explore whether public concern is likely being driven by media narratives, underlying scientific developments, or other factors.
Implications for Enterprise and Policy Insight
For organizations invested in real-time sentiment monitoring—whether around public health, ESG issues, or regulatory risk—this approach offers an exciting possibility: simulating how different population segments might react to emerging narratives before those responses are captured in traditional polls.
While simulated polling is not a replacement for established methodologies, it can serve as a rapid signal detection tool. The ability to inject a defined media diet into an AI model and observe likely opinion shifts offers new avenues for understanding narrative impact, testing message strategies, or anticipating stakeholder response to crises.
Crucially, this methodology also helps deconstruct the “why” behind public opinion, instead of simply measuring what people think, but how they may have arrived there.
Where This Could Lead
This proof-of-concept represents an early step in exploring how generative AI can be used to model narrative-driven perception at scale. With further refinement, such as more robust demographic grounding, more sophisticated media sampling, and longitudinal validation, these techniques could become powerful complements to conventional insight generation.
As enterprises seek to understand and navigate reputation dynamics in a media-saturated world, combining AI with empirical media modeling opens a new frontier: one that’s fast, flexible, and increasingly close to real-world complexity.