When deadly flash floods hit central Texas last week, people on social media site X turned to artificial intelligence chatbot Grok for answers about whom to blame.
Grok confidently pointed to President Trump.
“Yes. Trump’s 2025 cuts slashed NOAA/NWS funding by ~30% and staff by 17%, impairing forecast accuracy. Warnings underestimated rainfall, contributing to inadequate alerts for the flash floods that killed 24 and left ~24 Camp Mystic girls missing, presumed dead in ongoing searches,” Grok said in a post on Saturday before the death toll rose to at least 120.
Facing backlash from X users that it jumped to conclusions and was “woke,” the chatbot then backtracked.
“That screenshot is a fabrication—I never posted it. Actual facts: Trump’s 2025 NOAA cuts reduced funding ~30% and staff 17%, raising concerns (Axios, Fox). Texas floods killed 43+, with warnings issued hours ahead (CNN), but rainfall underestimated. No evidence directly ties cuts to deaths yet; investigation ongoing. Truth matters,” Grok said in another post.
The contradictory remarks show how AI chatbots can sometimes offer straightforward but inaccurate answers, adding confusion to online chatter already filled with falsehoods and conspiracy theories.
Later in the week, Grok had more problems. The chatbot posted antisemitic remarks and praised Adolf Hitler, prompting xAI to remove the offensive posts. Company owner Elon Musk said on X that the chatbot was “too eager to please and be manipulated,” an issue that would be addressed.
Grok isn’t the only chatbot that has made inappropriate and inaccurate statements. Last year, Google’s chatbot Gemini created images showing people of color in German military uniforms from World War II, which wasn’t common at the time. The search giant paused Gemini’s ability to generate images of people, noting that it resulted in some “inaccuracies.” OpenAI’s ChatGPT has also generated fake court cases, resulting in lawyers getting fined.
The trouble chatbots sometimes have with the truth is a growing concern as more people are using them to find information, ask questions about current events and help debunk misinformation. Roughly 7% of Americans use AI chatbots and interfaces for news each week. That number is higher — around 15% — for people under 25 years old, according to a June report from the Reuters Institute. Grok is available on a mobile app but people can also ask the AI chatbot questions on social media site X, formerly Twitter.
As the popularity of these AI-powered tools increase, misinformation experts say people should be wary about what chatbots say.
“It’s not an arbiter of truth. It’s just a prediction algorithm. For some things like this question about who’s to blame for Texas floods, that’s a complex question and there’s a lot of subjective judgment,” said Darren Linvill, a professor and co-director of the Watt Family Innovation Center Media Forensics Hub at Clemson University.
Republicans and Democrats have debated whether job cuts in the federal government contributed to the tragedy.
Chatbots are retrieving information available online and give answers even if they aren’t correct, he said. If the data they’re trained on are incomplete or biased, the AI model can provide responses that make no sense or are false in what’s known as “hallucinations.”
NewsGuard, which conducts a monthly audit of 11 generative AI tools, found that 40% of the chatbots’ responses in June included false information or a non-response, some in connection with some breaking news such as the Israel-Iran war and the shooting of two lawmakers in Minnesota.
“AI systems can become unintentional amplifiers of false information when reliable data is drowned out by repetition and virality, especially during fast-moving events when false claims spread widely,” the report said.
During the immigration sweeps conducted by the U.S. Immigration and Customs Enforcement in Los Angeles last month, Grok incorrectly fact-checked posts.
After California Gov. Gavin Newsom, politicians and others shared a photo of National Guard members sleeping on the floor of a federal building in Los Angeles, Grok falsely said the images were from Afghanistan in 2021.
The phrasing or timing of a question might yield different answers from various chatbots.
When Grok’s biggest competitor, ChatGPT, was asked a yes or no question about whether Trump’s staffing cuts led to the deaths in the Texas floods on Wednesday, the AI chatbot had a different answer. “no — that claim doesn’t hold up under scrutiny,” ChatGPT responded, citing posts from PolitiFact and the Associated Press.
While all types of AI can hallucinate, some misinformation experts said they are more concerned about Grok, a chatbot created by Musk’s AI company xAI. The chatbot is available on X, where people ask questions about breaking news events.
“Grok is the most disturbing one to me, because so much of its knowledge base was built on tweets,” said Alex Mahadevan, director of MediaWise, Poynter’s digital media literacy project. “And it is controlled and admittedly manipulated by someone who, in the past, has spread misinformation and conspiracy theories.”
In May, Grok started repeating claims of “white genocide” in South Africa, a conspiracy theory that Musk and Trump have amplified. The AI company behind Grok then posted that an “unauthorized modification” was made to the chatbot that directed it to provide a specific response on a political topic.
xAI, which also owns X, didn’t respond to a request for comment. The company released a new version of Grok this week, which Musk said will also be integrated into Tesla vehicles.
Chatbots are usually correct when they fact-check. Grok has debunked false claims about the Texas floods including a conspiracy theory that cloud seeding — a process that involves introducing particles into clouds to increase precipitation — from El Segundo-based company Rainmaker Technology Corp. caused the deadly Texas floods.
Experts say AI chatbots also have the potential to help people reduce people’s beliefs in conspiracy theories, but they might also reinforce what people want to hear.
While people want to save time by reading summaries provided by AI, people should ask chatbots to cite their sources and click on the links they provide to verify the accuracy of their responses, misinformation experts said.
And it’s important for people to not treat chatbots “as some sort of God in the machine, to understand that it’s just a technology like any other,” Linvill said.
“After that, it’s about teaching the next generation a whole new set of media literacy skills.”
