Letting AI Decide Your Vote

AI chatbots have become a major source of political information. According to Anthropic, the company behind the prominent chatbot “Claude,” over 5% of conversations with their chatbot concern politics. Both candidates and countries have already taken note. Campaigns are specifically trying to sculpt the information about candidates provided by chatbots, following the emerging areas of Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO). Already back in March Brazil’s Superior Electoral Court forbade chatbots to specifically rank or recommend candidates, something it now struggles to enforce. In Israel, high chatbot use turned the October 27th election into a test case for the challenges AI brings to electoral integrity.

How should we feel about this emerging layer of information access? What ethical challenges does it represent?

One concern is the acceleration of the same philosophical challenges represented by misinformation (incorrect information) and disinformation (deliberately deceptive information). To name a few: First, chatbots homogenize across sources into a single authoritative voice so the user may not know what information is higher and lower quality. This concern is exacerbated by the fact that many credible sources, which rely on subscriptions or clicks for financial support, bar chatbots from directly accessing them. Second, chatbots are not free from manipulation and can be “poisoned” by large-scale actors seeding the internet with disinformation. Third, chatbots can hallucinate, making up plausible-sounding information. Fourth, chatbots are notoriously “sycophantic,” following and amplifying the user’s biases; unless used carefully, they can quickly turn into a personalized echo chamber.

Beyond merely being frustrating, misinformation and disinformation undermine the health and legitimacy of democracy. “Instrumental” defenses of democracy contend democracy is good partly because it delivers good outcomes. It has often been noted that through harnessing the wisdom of crowds, democracies may be quite good at making decisions. However, this relies on participants in democracy having access to good information. A corrupted informational environment undermines this advantage.

Misinformation and disinformation also undermine consent and political legitimacy. Imagine a candidate who, after winning an election, dramatically tore off a latex mask to reveal they were a different person with different political ideas. We would likely feel we had been cheated, perhaps that the election was altogether invalid. Lies and distortions around a candidate represent a similar threat of misrepresentation and lack of consent. While a voter can always choose not to do the research, with disinformation we vote based on a deliberately false and misleading account of what we are voting for.

However, acknowledging the challenges presented by large language model chatbots is not the same as claiming they are uniquely misinformative. The Brennan Center for Justice found that chatbots can also push back against misinformation.

We may also worry about the threat they hold to decision making and the ownership of our decisions. Consider two ways to research candidates. In one, we look at each candidate and read their platforms as well as some associated coverage. In the other, we simply allow a chatbot to give us a recommendation. (To be clear, there are more active and critical ways to use chatbots.) Passively trusting a chatbot does not advance us as agents, as people able to delve into difficult questions. It instead offloads the thought and decision-making. The concern is especially acute for chatbots because of their verbal fluency.

Philosopher C. Thi Nguyen writes about the “seductions of clarity.” The clear, polished, authoritative voice of a chatbot can lead potential voters into an “aha” feeling, an undue sense of understanding about a candidate. However, at the same time, it is often true that we must rely on the insight and knowledge of others to make informed choices. The key question becomes who to trust, and with a new technology with known weaknesses – from manipulation to sycophancy – trust may be premature. But this does not mean AI cannot in principle become a trusted technology. It depends on who controls it, how it is regulated, and how it is used.

And then there are the values behind the models. During training, models are endowed with “values.” In other words, model developers try to give the model a set of behavioral tendencies that align with ethical considerations. This can include considerations about how models identify, integrate, and present information. These values can be a cause for concern. Qwen, made by the Chinese firm Alibaba, had apparently been designed to give positive answers about China. Researchers have also raised concerns with Chinese models regarding censorship of politically sensitive information.

Government censorship is less of an explicit concern in the US, but a small number of private corporations are now in the position to make enormously influential decisions about knowledge gathering and potentially have their own biases. As it stands, companies currently value strong fact-finding capabilities in their models. And AI doesn’t enjoy the same protections against liability for content provided like Facebook and Instagram do. But both of these trends could easily shift, and with them any aspiration for chatbot accuracy.

However, the fact LLMs currently represent a widely used access point to information also represents an opportunity. And through regulation or legislation there is space for democratic input about what the guiding values of models should be, including how models should assess, summarize, and present information.

A challenge for governments with strong free speech protections is balancing speaker rights against inaccurate or misleading information. The US Supreme Court has been especially leery of appointing the government as the arbiter of true and false political speech. But a large language model is not a speaker and its “speech” is not protected in the same way. Chatbots instead gather, summarize, and integrate information that has been provided by others. This does not mean anything goes. Free speech can also concern the rights of the listener to hear a variety of information. It may be a tricky line to draw, but there is room to shape the presentation of information that falls well short of censorship. It is plausibly in line with a model that, for instance, reasons carefully, flags information experts have questioned, mentions dubious provenance, and does not provide fringe claims unless asked.

Undoubtedly, AI chatbots represent distinct informational and democratic challenges. Not the least of which is ensuring they can effectively co-exist with the human journalists and researchers actually providing the information that LLMs then represent. However, we are not stuck with one way of using chatbots, either personally or collectively. Perhaps they can provide solutions to such challenges and not just further complications.