Rates Of Hallucination In AI Models From Google, OpenAI On The Rise
Experts warn that Google's AI Overviews are increasingly "hallucinating" false information, diverting users from accurate sources. Launched in May 2024, these summaries generated by Google's Gemini AI have been criticized for their inaccuracies, with recent studies showing hallucination rates of 33% to 48% in OpenAI models. Despite Google's claims of improvement, experts note that the frequency of errors is rising, leading to a significant drop in click-through rates to legitimate articles. This trend raises concerns about the efficiency of AI systems and the spread of misinformation.
Google's AI Overviews are "hallucinating" false information and drawing clicks away from accurate sources, experts warned The Times of London late last week.
Google introduced its AI Overviews, a feature that aims to provide quick answers to search queries, in May 2024. Summaries are written by Google's Gemini AI – a large language model similar to ChatGPT – scans through the results of the search to create the graphs and includes links to some of the sources.
Don't Miss:
- Maker of the $60,000 foldable home has 3 factory buildings, 600+ houses built, and big plans to solve housing — this is your last chance to become an investor for $0.80 per share.
- Peter Thiel turned $1,700 into $5 billion—now accredited investors are eyeing this software company with similar breakout potential. Learn how you can invest with $1,000 at just $0.30/share.
Google Vice President of Search Elizabeth Reid said In a blog post that the overviews were designed to be a "jumping off point" that provided higher-quality clicks to webpages. "People are more likely to stay on [those pages], because we've done a better job of finding the right info and helpful webpages for them."
However, experts told The Times of London that these answers can be "confidently wrong" and direct searchers away from legitimate information.
When generative AI imagines facts or otherwise makes mistakes, computer scientists refer to it as hallucinating. These hallucinations can include references to non-existent scientific papers, like those NOTUS found were cited in Health and Human Services Secretary Robert F. Kennedy Jr.'s "Make America Healthy Again" report, and a host of other errors in judgment.
Trending: Invest where it hurts — and help millions heal: Invest in Cytonics and help disrupt a $390B Big Pharma stronghold.
Shortly after AI Overviews were launched last year, users began to point out the frequency with which these summaries included inaccurate information, the Times of London reports. One of its most notorious hallucinations was the suggestion that a user add non-toxic glue to pizza sauce to help the cheese stick better.
Pre-IPO Offer: Get A Piece Of A Nearly $5T Global Opportunity By Joining BOXABL As An Early Shareholder At Just $0.80/Share
- Massive Demand & Disruptive Potential – Boxabl has received interest for over 190,000 homes, positioning itself as a major disruptor in the housing market.
- Revolutionary Manufacturing Approach – Inspired by Henry Ford’s assembly line, Boxabl’s foldable tiny homes are designed for high-efficiency production, making homeownership more accessible.
- Affordable Investment Opportunity – With homes priced at $60,000, Boxabl is raising $1 billion to scale production, offering investors a chance to own a stake in its growth.
Share Price: $0.80
Min. Investment: $1,000
Valuation: $3.5B
Google pushed back, claiming that many of the examples circulating were fake, but Reid acknowledged in her blog post that "some odd, inaccurate or unhelpful AI Overviews certainly did show up. And while these were generally for queries that people don't commonly do, it highlighted some specific areas that we needed to improve."
According to the experts who spoke to The Times of London, despite the technological advancements and improvements, hallucinations are getting worse rather than better. New reasoning systems are producing more incorrect responses than their predecessors, and designers aren't sure why.
In a recent study, it was found that two recent OpenAI models, o3 and 04-mini, hallucinated in 33% and 48% of answers, respectively, according to The Times of London. These percentages are more than double those of previous models.
Features like Google's AI Overviews or ChatGPT summaries are also drawing clicks away from more accurate resources. Laurence O'Toole, founder of tech firm Authoritas, tracked the impact of AI Overviews and told The Times of London that when they exist click click-throughs to articles go down by 40% to 60%.
The compounding issues of AI presenting inaccurate information and causing searchers to click through to more accurate articles have many worried about efficiency and the spread of fake news.
"You spend a lot of time trying to figure out which responses are factual and which aren't," the chief executive of Okahu, Pratik Verma, told The New York Times last month. Okahu works with AI engineers to make the technology better, and helps companies troubleshoot issues, including hallucinations. "Not dealing with these errors properly basically eliminates the value of A.I. systems, which are supposed to automate tasks for you," he said.
- Are you rich? Here’s what Americans think you need to be considered wealthy.
- Inspired by Uber and Airbnb – Deloitte's fastest-growing software company is transforming 7 billion smartphones into income-generating assets – with $1,000 you can invest at just $0.30/share!
Image: Shutterstock
