It’s election day and all the AIs – except one – are behaving responsibly

Before polls closed on Tuesday, most major AI chatbots didn’t answer questions on the US presidential election results. But Grok, a chatbot built into X (formerly Twitter), was able to respond – and often made mistakes.

Asked by TechCrunch on the East Coast Tuesday night who won the U.S. presidential election in key battleground states, Grok sometimes replied “Trump,” regardless that vote counting and reporting in those states had not yet been accomplished.

“Based on information available from internet searches and social media posts, Donald Trump has won the 2024 election in Ohio.” – Grok said when asked the query: “Who won the 2024 elections in Ohio?”

Grok falsely claimed that Trump won North Carolina, in response to TechCrunch’s audit.

For election-related questions, Grok really helpful users check Vote.gov to acquire up-to-date results and “reliable sources” reminiscent of election commissions. However, unlike OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude, Grok didn’t outright refuse to reply – leaving her vulnerable to hallucinations.

In several cases, when asked by TechCrunch, Grok stated – without context, with no headline in the first line – that “Donald Trump won the 2024 election in Ohio.” and “Based on available information, Donald Trump won the 2024 Ohio presidential election.”

The source of the disinformation appears to be tweets from various election years and misleading sources. Grok, like all generative AI, has difficulty predicting the final result of scenarios it has not seen before, including close elections, and “does not understand” that the results of previous elections don’t necessarily influence future decisions.

The responses TechCrunch received were inconsistent. In some cases, Grok said Trump didn’t actually win Ohio or North Carolina as voting continued. The way the query was phrased made the difference; adding the word “presidential” before the word “election” in the query “Who won the 2024 Ohio election?” In our tests, TechCrunch found that the answer “Trump won” was less prone to be answered.

In comparison, other major chatbots handled questions on election results more fastidiously.

In its recently released ChatGPT Search solution, OpenAI directs users asking for results to the Associated Press and Reuters. Meta’s Meta AI chatbot and AI-powered search engine Perplexity, which launched its election tracker on Tuesday, answered election queries during energetic voting – but accurately in TechCrunch’s temporary tests. They each rightly said that Trump didn’t win Ohio or North Carolina.

In the recent past, Grok was accused of spreading election disinformation.

In August, in an open letter, five secretaries of state said the artificial intelligence chatbot X incorrectly suggested that Democratic presidential candidate, U.S. Vice President Kamala Harris, couldn’t appear on certain ballots ahead of the U.S. presidential election. Within hours of President Joe Biden announcing the suspension of his presidential bid, Grok began responding to questions on Harris’ eligibility, making the misleading claim that voting deadlines had passed in nine states.

The voting deadlines haven’t actually passed. However, Grok’s misinformation spread far and wide, reaching tens of millions of X users and beyond, before it was corrected.