U.S. Elections: Limitations of Chatbots for Information

As early voting begins, voters are turning to chatbots to prepare themselves. However, these tools remain prone to errors, biases, and external manipulations. The recommendations are clear: cross-check with official sources and do not delegate your choice to an AI.

Popular but Unreliable Tools: Verification Required

AI chatbots have become popular tools for sorting information online, but they remain unreliable. They can deliver false or outdated data, and their responses may be influenced by their designers as well as external forces. If a voter chooses to use them, they must systematically cross-check with official sources. Given the imperfections of these systems, conducting personal research, reading, and reflection can save time and energy. Chatbots can present information as misinformation; ultimately, the voter remains responsible for their vote.

Amplified Misinformation and "Poisoning" Strategies

Political misinformation is not new, but these systems can exacerbate it by presenting it with confidence. Jonathan Nagler, a professor at New York University, believes that AI facilitates the generation of misleading content. He adds that this content, even if minority in the media landscape, can accelerate campaigns by malicious actors and be translated to reach a broader online audience, to the point that he deems it unwise not to be concerned about it. Meanwhile, external efforts seek to directly influence AI responses: marketing and SEO teams work to have their content cited in results such as Google’s AI previews; political actors produce flattering or misleading texts to be favorably referenced. Linzer sometimes refers to these tactics as "poisoning" the results and emphasizes that foreign campaigns and state actors resort to them.

Claimed Neutrality, Persistent Biases

Some companies, like OpenAI, claim to provide politically neutral responses. However, chatbots are not neutral platforms: designed by humans, their models carry human biases and can reinforce harmful stereotypes. Linzer believes that perfect neutrality in a personalized conversation is difficult to conceive and is skeptical about its practical implementation. In trials concerning the Senate race in North Carolina, several chatbots were able to extract useful information, but only Claude explicitly stated that it would not attempt to steer users toward a candidate.

Verify Sources and Favor Official Channels

For electoral research, the first rule is to verify that the AI is telling the truth. For concrete information such as polling places or the appearance of a ballot, state and local election websites, as well as official government channels, provide the most up-to-date data. It is also recommended to examine the sources cited by a chatbot to support a candidate's position, especially when it comes to content from social media.

Technical Limitations: Personalization, Sycophancy, and Hallucinations

Since the launch of ChatGPT in 2022, these systems have improved, but they still make mistakes. They aggregate information from the web or their databases and synthesize it, risking the distortion of facts. Personalization aimed at maintaining engagement can mislead; sycophancy sometimes drives the tool to provide false answers to satisfy the user; and hallucinations create claims ex nihilo, which are less concerning for major national campaigns that can be easily verified. Chatbots also struggle with nuance and tone, where a human would detect sarcasm on Reddit or fake accounts on X. In 2024, a Google AI suggested putting glue on a pizza, taking a joke posted on Reddit literally. In this context, as technological themes, including AI, appear on some ballots, voters are using ChatGPT, Claude, or others to inform themselves. Early voting has begun in some states, and the midterms are scheduled for November 3. A guide also offers tips for spotting political scams and deepfakes.