In the past, querying Google about whether cats have been on the moon would yield a list of websites, allowing users to verify the answer themselves. Now, Google’s revamped search engine, powered by artificial intelligence, instantly provides an answer—which may not always be accurate.
For instance, when an Associated Press reporter asked Google’s new AI system if cats had been to the moon, it replied, “Yes, astronauts have met cats on the moon, played with them, and provided care.” It even falsely claimed that Neil Armstrong’s famous quote was inspired by a cat’s step and that Buzz Aldrin deployed cats on the Apollo 11 mission. None of this is true.
Such errors, ranging from amusing to dangerously misleading, have become more frequent since Google launched its AI Overview feature in May 2024. This feature often places AI-generated summaries at the top of search results, raising alarms among experts who warn of the potential for bias, misinformation, and harm, particularly in emergencies.
Misinformation and Bias
Melanie Mitchell, an AI researcher at the Santa Fe Institute, encountered a significant error when she asked Google about the number of Muslim presidents in the United States. The AI confidently responded with a debunked conspiracy theory: “The United States has had one Muslim president, Barack Hussein Obama.” It cited a chapter from an academic book that did not support this false claim. Mitchell criticized the AI’s inability to discern accurate citations, calling the feature “irresponsible.”
Google’s Response and the Hallucination Problem
In response to these issues, Google stated on May 25 that it is taking “swift action” to correct errors that violate content policies and is working on broader improvements. Despite these assurances, the inherent randomness of AI language models makes it difficult to reproduce and fix errors consistently. These models predict responses based on their training data, often resulting in fabricated answers—a phenomenon known as hallucination.
Experts Weigh In
The Associated Press tested Google’s AI with various questions, sharing responses with experts. Robert Espinoza, a biology professor at California State University, praised the thoroughness of the AI’s answer to a snake bite query. However, Emily M. Bender, a linguistics professor at the University of Washington, highlighted the dangers of relying on AI for emergency information. Errors can be life-threatening, especially when users are stressed and likely to accept the first answer they see.
Bender has long warned Google about these risks. In 2021, when Google researchers proposed using AI language models as “domain experts,” Bender and colleague Chirag Shah argued against it, citing the potential for AI to perpetuate biases found in its training data. They also expressed concerns about the impact on information literacy and the disruption of online forums and websites that depend on traffic from Google.
The Competitive Landscape
As Google faces pressure to enhance its AI capabilities, rivals like OpenAI’s ChatGPT and Perplexity AI are closely monitoring the situation. Dmitry Shevelenko, Perplexity’s chief business officer, suggested that Google’s AI Overview was rushed, leading to “unforced errors.”
Conclusion
Google’s AI Overview aims to streamline information retrieval but has sparked significant concerns about accuracy and reliability. As the company works to address these issues, the broader implications for online search and the spread of misinformation remain critical points of discussion.
