Artificial intelligence (AI) is increasingly being used by polling institutes to improve surveys and predict election outcomes. Traditional polling organizations are adopting AI to enhance their data, while newer companies are developing models that can predict election results without directly collecting responses from voters. This approach echoes a 1955 science fiction story by Isaac Asimov, A Voted, where a supercomputer determines election results based on a single voter's answers. Asimov's story was meant as a critique of the influence of polls, but today, they play a central role in shaping public discourse, even as concerns about their accuracy and reliability grow. In France, the polling institute Ifop has pioneered the use of "boosted samples," where AI generates synthetic data to supplement small survey samples. In 2024, Ifop used AI to create over 400 responses from teachers for a European elections survey, based on just 116 actual responses. Later that year, the same institute produced a survey on the Paralympic Games, where only 151 of the 1,151 responses came from real people with disabilities. The rest were artificially generated. Statistician Thomas Delclite explains that AI compares small survey groups with larger ones, using factors like location, income, and education, to generate plausible responses. However, this method risks amplifying existing biases and creating data that may not reflect reality. Polling companies are turning to AI partly due to rising costs and challenges in collecting data. Thomas Duhard, a director at Ifop, acknowledges that while AI can reduce costs, it still requires significant time and resources. He notes that Ifop uses AI-generated samples mainly for private marketing research, not for political polling. Meanwhile, companies are also creating "synthetic profiles" to model voter behavior throughout election campaigns. For example, Ipsos generated five fictional characters for the British elections, each with specific traits and political preferences. These profiles, while useful for modeling, often reflect the stereotypes of the people who create them. Some startups are even offering entirely AI-generated populations, known as "silicon samples." A US-based startup, Aaru, raised $50 million in 2025 for its AI-driven research, including a maternal health survey that initially misled media outlets. Other companies, such as Simile and Electric Twin, have also raised significant funding for similar services. In France, Eleya is developing this technology and has partnered with Ifop for marketing research. However, researchers warn that generative AI may not be reliable, as it tends to stereotype demographic groups and oversimplify human reasoning. The lack of transparency in how these synthetic surveys are created and interpreted raises concerns about their impact on public understanding. Polling institutes, which are private organizations with varying levels of transparency, are already under scrutiny for potential biases. Valérie Charolles, a member of the Polling Commission, highlights the risks of AI-generated data in elections, particularly given the lack of regulatory oversight. She notes that the commission has no authority to prevent the release of questionable surveys. In the case of Ifop, the AI data used was provided by Fairgen, an Israeli startup, but neither Ifop nor Fairgen disclosed this relationship. This raises fears of manipulation, similar to past scandals involving data misuse. Without clear legal guidelines, the responsibility for interpreting these polls falls on voters, who must critically assess the information presented.