Generative AI like ChatGPT could help boost democracy — if it overcomes key hurdles

The dawn of artificial intelligence systems that can be used by almost anyone, like ChatGPT, has revolutionized business and alarmed policymakers and the public.

Advanced technologies can feel like unstoppable forces shaping society. But a key insight from scholars of philosophy and of the history of technology is that people can, in fact, exert a lot of control over how and where we use these tools.

To us, as political scientists, this new technology offers some interesting opportunities to improve democratic processes, such as through increasing civic knowledge and facilitating communication with elected representatives — if key challenges are met. And we’ve begun to research how this might happen.

Increasing civic knowledge

Politics can feel impossibly complicated, with emotion-laden negative marketing campaigns and political winds that seem to change almost daily. Many cities, states and countries provide little to no information to the public about policy issues, political candidates or policy referendums. So even when citizens have the ability to exercise their democratic freedoms, they may not feel well enough informed to do so.

Generative AI could help. Building on platforms like isidewith.com, politicalcompass.org and theadvocates.org, AI could help people answer questions about their core beliefs or policy positions, and then help them determine which political candidates, parties or choices best match their views.

Existing websites like Ballotpedia, Vote Smart and Vote411 have made tremendous advances in providing critical information like sample ballots, polling place locations and candidate positions to voters. But these websites can be difficult to navigate. AI technologies may be able to provide enhanced services, and at local, state, regional, national and international levels. These systems may eventually be able to use automation to provide continuously updated information on candidates and policy issues.

AI chatbots could also interactively help people think through complex issues, learn new skills and determine their policy stances, while also providing relevant news stories and facts.

However, at the moment, generative AI systems aren’t really ready to answer democracy-related questions reliably or without bias. Large language models generate text based on statistical frequencies of words in their training data, with little regard for whether the statements are fact or fiction.

For example, AI systems could hallucinate by fabricating nonexistent politicians or by generating inaccurate candidate positions. These systems also appear to generate output with political biases. And the rules protecting the privacy of users and compensating individuals or organizations whose data are used by these systems are not yet clear, either.

Before generative AI is ready to enhance democracy, then, there’s a lot to understand and address.

Facilitating constituent communication

One area for inquiry: Could generative AI help constituents communicate with their elected representatives?

Contacting a politician can be intimidating, and many Americans may not even know where to start. Survey research reveals that fewer than half of Americans can name the three branches of government. Knowing the names of their own representatives, much less getting in contact with them, is even less common. For example, in 2018, only 23% of survey respondents in a poll by the Pew Research Center indicated that they had contacted an elected official in the past year, even at a time with significant developments in national politics.

To encourage greater outreach to representatives, generative AI could help citizens not only identify their elected officials, but even draft detailed letters or emails to them.

We examined this idea in a recent study we conducted as part of our work at the Governance and Responsible AI Lab at Purdue University. We ran a survey of American adults in June 2023 and found that 99% of respondents had at least heard of generative AI systems like ChatGPT, and 68% had tried them out personally. However, 50% also reported that they have never contacted any of their elected political representatives.

As part of the survey, we showed some survey respondents an example of a message written by ChatGPT to a state legislator about an education funding bill. Other respondents, the control group, saw the same example email, but with no indication it was written by AI.

Survey respondents who heard about this possible use of AI said they were significantly more likely than those in the control group to support the use of AI for communication with politicians, both by individuals and by advocacy groups. Because of their support for using this new technology, we had expected they would be inclined to reach out to politicians more frequently and see AI as making that process easier. But we found that not to be true.

Nevertheless, we have identified an opportunity. For example, public-interest groups could use AI to improve mass advocacy campaigns by helping citizens more easily personalize emails to politicians. If they can ensure the AI-generated messages are factual and valid reflections of citizens’ views, many more people who might not historically have contacted their politicians might consider doing so.

There are risks, though, including that politicians might be skeptical of communications they think are written by AI.

Maintaining authenticity and the human touch

One of the most significant drawbacks to using generative AI for political communication is that it might make the receivers of messages suspicious that they are not actually in conversation with a real human. To test this possibility, we included a warning for some of the people taking our surveys that the use of mass AI-generated messages could lead politicians to doubt whether the messages were authentically created by humans.

We found that these people, compared to those in the control group, felt that legislators would indeed be less likely to pay attention to emails, and that emails would be less effective at shaping policymakers’ opinions or decisions.

Strikingly, however, these people still supported the use of generative AI in political communication. One possible explanation for this finding is something called the “trust paradox” of AI: Sometimes, even when people find AI untrustworthy, they still support its use. They may do so out of a belief that future versions of the technology will be better, or because they lack effective alternatives.

So far, our early research into the implications of generative AI for political communication suggests a few key lessons.

First, even with ostensibly easy-to-use AI tools, politics is still out of reach for many of those who have historically lacked opportunities to share their thoughts with politicians. We even found that survey respondents with higher baseline trust in government or who had had prior contact with government were less likely to support AI use in this context, perhaps to preserve their heightened existing influence in government. Therefore, greater availability of AI tools might not mean more equal access to politicians, unless these tools are carefully designed.

Second, given the importance of human contact and authenticity, a critical challenge is making use of AI’s opportunities while also preserving the human touch in politics. While generative AI could enhance aspects of politics, we shouldn’t be too quick to automate away the relationships that underpin our social fabric.