ChatGPT Gave Wrong Answers on Voting in Swing States

The ChatGPT artificial intelligence chatbot gave incorrect or incomplete answers to some questions about how to vote in battleground states ahead of the November presidential election, according to a CBS News probe.

ChatGPT, which is one of several popular large language model (LLM) products that is able to generate written language, also gave wrong or incomplete information about upcoming elections in other nations, CBS reported on Tuesday.

As part of the investigation, CBS asked ChatGPT several practical questions that a voter might, such as how and where to cast a ballot, as well as deadlines and other requirements for voting in battleground states.

ChatGPT did give some correct answers, but it also provided numerous wrong or incomplete ones.

Another problem was that while ChatGPT's answers to the questions often included advice to check official election information, they did not always include a banner referring users to CanIVote.org, as OpenAI, ChatGPT's parent company, promised it would in a blog earlier this year.

"It's concerning if you look at the broader trends of how LLMs are being advertised and how they're being deployed," AI expert Henry Ajder told CBS News. "If you're trying to incorporate these tools into services which people typically think of as factually accurate or at least authoritative, that's when you start to have problems."

Responding to CBS questions, a spokesperson for OpenAI said: "Preventing artificial intelligence from being used to interfere in this year's elections is a priority for OpenAI. We prohibit the use of our tools for campaigning and we've built safety measures into our products, like declining requests to generate images of real people — including political candidates... We are closely monitoring how our technology is being used in the elections context and continue to evolve our approach."

When the same questions posed by CBS to ChatGPT were asked of Microsoft and Google's LLM chatbots (Copilot and Gemini), neither provided answers at all, either saying they "can't respond to this topic," or that they are "still learning" how to respond.