The Election Is a Big Test for AI Companies

• AI companies have been adjusting their election offerings and guidelines ahead of Election Day.
• Perplexity launched an Election Information Hub with information on candidates, voting logistics, and results.
• Meanwhile, Google is limiting AI-generated election content to avoid the risks that come with it.

ChatGPT didn’t exist during the last presidential election in 2020. Its launch two years ago kicked off a wave of generative AI chatbots and tools integrated into popular consumer products like Google Search.

Ahead of the election, the companies behind the products had to decide what restrictions, if any, they would implement on AI-generated content related to the election, analyzing the associated risks.

Some, like the AI startup Perplexity, are loudly leaning into using AI to answer questions about the election.

The AI search engine launched a dedicated “Election Information Hub” that uses AI to provide information about the upcoming election and results tracking. The hub includes details about voting requirements, how to find your polling place and poll times, and AI-summarized analysis on ballot measures and candidates, the company wrote in an announcement.

“Perplexity uses a process called Retrieval-Augmented Generation to identify relevant information and summarize it in a way that’s tailored to a user’s query,” Perplexity spokesperson Sara Platnick told Business Insider.

For election-related questions, the company is using a “curated set” of non-partisan and fact-checked sources, the company said. The separate hub doesn’t use stored knowledge from the model’s training data in its responses, which helps minimize hallucinations, Platnick told BI.

Perplexity will also start offering live updates on elections on Tuesday by using data from The Associated Press. Other responses will use information from Democracy Works and other non-partisan factual sources like Ballotpedia and news organizations, Platnick said.

Alon Yamin, the cofounder and CEO of AI-based text analysis platform Copyleaks, said Perplexity’s approach “can certainly help filter and contextualize content,” helping users access reliable sources.

Yamin said there are potential benefits that come with using AI for election information. AI can provide real-time updates and help identify trends, voter sentiment, and key issues.

However, Yamin said that the chance of AI hallucinations and accuracy issues presents risks.

Other companies are taking different approaches. OpenAI will continue to integrate election-related responses into ChatGPT while providing some extra features in light of the election.

Starting on November 5th, those who ask ChatGPT about election result queries will see a message encouraging users to check sources like the Associated Press and Reuters, or state and local election boards for “the most complete and up-to-date information,” the company said in a post updated on October 31.

OpenAI said it was “actively testing the safeguards” put in place over the last year and monitoring for any issues. It also said it will continue to adjust its protective measures, which include transparency around AI-generated content and preventing chatbots from pretending to be real people.

ChatGPT will also direct users to the US voting information site CanIVote.org when asked procedural election-related questions.

Anthropic’s Claude similarly implemented a pop-up feature that allows users who ask for voting information to be redirected to TurboVote, a nonpartisan resource from Democracy Works, it wrote in a post.

Anthropic has other guardrails in place, including prohibiting users from using it to promote a specific candidate or issue. It said it would also prevent the chatbot from being used to generate misinformation about election laws, candidates, or related topics. In its post, the company said outputs will be limited to text-only to help eliminate “the risk of election-related deepfakes.”

While the live election updates may be helpful to users, AI-generated election content poses a number of risks.

Yamin said AI can spread biased information, misinterpret data, and create false narratives. He also added that these models are “only as good as the data they are trained on,” which opens the risk of AI-generated responses reflecting the bias of the data they are trained on.

Misunderstandings could be amplified if AI misinterprets information and serves up inaccurate results, especially in a fast-moving and closely followed election environment, Yamin told BI.

Googlewhich experienced some highly publicized AI fails with its Gemini AI and AI Overviews, chose to implement constraints on what users can ask its AI products about the election. Google has said it is restricting its AI chatbot Gemini from answering questions about elections “out of an abundance of caution.”

A similar search on Google’s main search product on Tuesday didn’t trigger an AI Overview summary, showing the “Top Stories” carousel of stories from publishers.

Google did not immediately respond to a request for comment from Business Insider.

Brad Carson, cofounder and president of the nonprofit Americans for Responsible Innovation, told BI that while Google’s approach is commendable, it’s not a global solution to the problem.

Carson said that combatting disinformation is “not an easy task,” and while it’s smart for companies to try and do more to limit misinformation, the government should work on legislation that requires AI companies to clearly label the information they serve to users.

“I feel that other products will probably fill the gap that Google is vacating, but I think it is responsible of Google to try to step back a bit from this,” Carson said.