The landscape of opinion polling is undergoing a significant transformation, driven by the increasing integration of artificial intelligence. Companies like Naratis, a French firm founded in 2025, are pioneering the use of conversational AI to conduct qualitative research, a traditionally slow and expensive segment of the polling industry. Pierre Fontaine, the 28-year-old engineer behind Naratis, explains that their AI agents engage respondents in conversations, moving beyond simple tick-box surveys to explore the nuances of how people form their opinions.
Naratis aims to revolutionize qualitative studies, which typically involve small groups or one-on-one interviews that can take weeks to complete and analyze. By replacing human interviewers with AI, Naratis claims its process is "10 times faster, 10 times cheaper and 90% as accurate as human polling." This allows for rapid data collection, often within 24 hours, enabling clients to react to unfolding events in near real-time. The efficiency stems from what Fontaine terms "parallelisation," where AI agents can conduct numerous interviews concurrently, a stark contrast to the sequential nature of human interviews.
This shift towards AI-driven polling arrives at a critical juncture for the industry. Stéphane Le Brun, an AI consultant, highlights a sharp decline in survey response rates, dropping from over 30% in the 1990s to below 5% currently. This trend makes polling more costly and less representative, contributing to a decline in public trust. The industry faces challenges in accurately capturing public sentiment, with past high-profile polling failures, such as predictions for Brexit and the 2016 US presidential election, casting a shadow over traditional methods.
Fontaine distinguishes Naratis's approach from the quantitative polling that often struggles with predictive accuracy. He argues that qualitative research, which his company focuses on, is more about understanding the 'why' behind opinions rather than forecasting election outcomes. This distinction is crucial, as it positions AI as a tool for deeper insight into public sentiment, such as gauging reactions to campaign messages, rather than solely for predicting vote counts.
Established polling firms are also embracing AI, albeit with varying degrees of integration. Ipsos, a major player, extensively uses AI in market research. Instead of relying solely on self-reported data, researchers are experimenting with having respondents film themselves, allowing AI to analyze their behavior directly. This observational approach offers a potentially more objective understanding of consumer habits and preferences.
Furthermore, AI is being employed to analyze vast amounts of social media data and to develop sophisticated tools like "digital twins" and "synthetic people." Digital twins are virtual models designed to mimic the responses of real individuals, while synthetic data involves generating entirely new profiles based on observed patterns. These technologies are particularly useful for studying small or hard-to-reach demographic groups, a persistent challenge in polling. Researchers sometimes blend real and simulated respondents, using human participants to validate the AI-generated insights.
However, the application of AI in politically sensitive polling remains a subject of caution. Ipsos, for instance, refrains from using AI-generated respondents in political surveys. Bruno Jeanbart, CEO of OpinionWay, echoes this sentiment, stating that his firm would "never publish an opinion poll based on AI-generated data" due to concerns about maintaining public trust. This highlights a significant ethical and practical consideration: the potential for AI to introduce new forms of bias or to generate data that lacks genuine human grounding.
The advantages of AI in polling are undeniable: increased speed, reduced costs, and enhanced flexibility. AI enables the collection of richer, more detailed data and allows for swift responses to current events. It also holds the potential to mitigate certain human biases, as individuals might feel more comfortable expressing candid opinions to a machine, particularly on sensitive topics. This could, for example, help address the consistent underestimation of far-right support observed in French political polling.
Despite these benefits, the risks associated with AI in polling are substantial. AI systems are known to "hallucinate," generating plausible but inaccurate information. They can also produce "common sense" responses that reflect general societal views rather than the specific, potentially unconventional, opinions the poll aims to capture. The very nature of polling is to uncover what people *actually* think, not just what is commonly believed.
The use of synthetic data raises fundamental questions about what is being measured. If responses are generated rather than collected from real individuals, the validity and interpretation of such data become complex. Trust remains a paramount concern, especially as polling is already subject to intense political scrutiny and regulation. The introduction of AI, particularly in data generation, could exacerbate these concerns, potentially leading to stricter regulations or even prohibitions on certain AI-driven polling methods, as Jeanbart anticipates for countries like France.
Even proponents of AI acknowledge its current limitations. Le Brun emphasizes that while end-to-end automation is a long-term goal, completely removing human oversight would be "unsafe and socially unacceptable" at present. Human involvement is deemed essential for validating AI-generated results and for taking ultimate responsibility for the findings. The most probable future for opinion polling appears to be a hybrid model.
In this hybrid future, AI will likely expand the capabilities of polling by enabling large-scale conversational surveys, integrating diverse data sources like social media, and delivering insights more rapidly. Techniques such as digital twins and synthetic data may find specialized applications, particularly in market research where the stakes for political outcomes are lower. However, in the realm of political polling, the distinction between augmenting human data and simulating it is expected to remain a critical boundary.
Companies like Naratis are positioning themselves at the forefront of this evolution, betting that the true innovation lies not in replacing human respondents but in transforming the way their voices are captured and understood. By turning surveys into conversations and conversations into vast datasets, they aim to redefine public opinion measurement. Ultimately, whether this technological shift restores or further erodes public trust in polling will depend on its implementation, transparency, and regulatory frameworks, rather than solely on the technology itself. The persistent economic pressures on the polling industry will undoubtedly continue to drive the pursuit of greater automation and efficiency.
