Who Needs Humans? Harvard Researchers Advocate for Polling AI Instead of Voters

As traditional polling methods face increasing challenges, Harvard researchers believe AI could be apotential game-changer in the world of political polling. Not by asking better questions — but by answering polls in place of humans.

A recent article from the Harvard Kennedy School Ash Center reports that in recent years, the efficacy of traditional political polling techniques has come under scrutiny. Declining response rates and the potential for respondents to provide inauthentic answers have led to questions about the reliability of poll results. However, researchers believe the rise of AI could offer a solution to these problems, potentially transforming the way political campaigns and movements gauge public opinion.

The challenges facing traditional polling methods are twofold. First, pollsters are finding it increasingly difficult to reach people. In an era of spam calls and overflowing inboxes, fewer individuals are willing to participate in surveys, whether by phone or mail. According to Pew Research, the response rate for their polls dropped from 36 percent in 1997 to a mere six percent by 2018. This trend is not unique to the United States; pollsters worldwide are grappling with similar issues.

Second, even when people do participate in polls, they may not always provide honest responses. Some individuals may be embarrassed to reveal their true opinions, while others may answer based on partisan loyalties, either aligning with their party’s stance or deliberately contradicting the opposition.

Despite these shortcomings, polling remains a central focus in modern politics, often overshadowing substantive policy discussions. This obsession with polling numbers can be detrimental to the democratic process, as it shifts attention away from the real-world consequences of electoral choices.

Harvard reseaerchers believe that pollsters could leverage AI large language models (LLMs) trained on vast amounts of online data to analyze and summarize the expressed opinions of individuals and groups across the internet. These models can identify trends by demographic and even extrapolate to new policy issues with a level of accuracy comparable to human experts.

One approach involves creating AI agents – instances of AI models conditioned to behave like individuals with specific demographic characteristics and media consumption habits. Researchers have already experimented with simulating polling results using populations of thousands of these AI agents. Unlike human respondents, AI agents are always available to answer questions, allowing political strategists to quickly test how different voter segments might react to various policy positions or campaign messages.

The power of this system lies in its ability to generalize to new scenarios and survey topics, providing plausible answers even when dealing with limited data. While the accuracy may not be guaranteed, AI-generated responses can often rival the predictions of human political experts. Moreover, if the results seem questionable, users can immediately prompt the AI with follow-up questions to gain a deeper understanding.

However, AI polling is not without its limitations. The quality of the AI’s responses is heavily dependent on the data used to train the model. If the training data is outdated or lacks relevant context, the AI may generate inaccurate results. For example, in experiments conducted with an early version of the model behind ChatGPT, the AI agents failed to capture the shift in U.S. public opinion regarding intervention in Ukraine following Russia’s 2022 invasion.

To overcome these shortcomings, AI agents can be designed to automatically source and incorporate new data when answering questions. By exposing each AI agent to the same social and media news sources as humans from corresponding demographics, the model can better simulate authentic responses. Additionally, AI agents can query relevant contextual information, such as demographic trends and historical data, to refine their expectations.

While AI polling will likely never achieve perfect accuracy, it can still provide valuable insights, particularly when used in conjunction with traditional polling methods. AI tools can help identify areas of uncertainty or rapid change, flagging instances where human input is needed to calibrate the model. In this way, AI can augment, rather than replace, traditional polling techniques.

As AI polling continues to evolve, it is likely to become an irresistible tool for political campaigns and media outlets. However, it may take some time for AI-generated results to be seen as credible, given the inherent trust in direct human responses. Initially, AI-assisted polls may be used internally by campaigns, while news organizations continue to rely on more traditional methods. A major election where AI proves more accurate than human pollsters could be the turning point in shifting this perception.

Read more at the Harvard Kennedy School Ash Center here.

Lucas Nolan is a reporter for Breitbart News covering issues of free speech and online censorship.