The newest and most powerful technologies – so-called reasoning systems from companies including OpenAI, Google and the Chinese startup DeepSeek – are generating more errors, not fewer. As their math skills have notably improved, their handle on facts has gotten shakier. It is not entirely clear why.
Today’s AI bots are based on complex mathematical systems that learn their skills by analysing enormous amounts of digital data. They do not – and cannot – decide what is true and what is false. Sometimes, they just make stuff up, a phenomenon some AI researchers call hallucinations. On one test, the hallucination rates of newer AI systems were as high as 79%.
These systems use mathematical probabilities to guess the best response, not a strict set of rules defined by human engineers. So they make a certain number of mistakes. “Despite our best efforts, they will always hallucinate,” said Amr Awadallah, CEO of Vectara, a startup that builds AI tools for businesses, and a former Google executive. “That will never go away.”
For several years, this phenomenon has raised concerns about the reliability of these systems. Though they are useful in some situations – like writing term papers, summarising office documents and generating computer code – their mistakes can cause problems.
The AI bots tied to search engines like Google and Bing sometimes generate search results that are laughably wrong. If you ask them for a good marathon on the West Coast, they might suggest a race in Philadelphia. If they tell you the number of households in Illinois, they might cite a source that does not include that information.
Those hallucinations may not be a big problem for many people, but it is a serious issue for anyone using the technology with court documents, medical information or sensitive business data.
“You spend a lot of time trying to figure out which responses are factual and which aren’t,” said Pratik Verma, co-founder and CEO of Okahu, a company that helps businesses navigate the hallucination problem. “Not dealing with these errors properly basically eliminates the value of AI systems, which are supposed to automate tasks for you.”
Cursor and Truell did not respond to requests for comment.
For more than two years, companies such as OpenAI and Google steadily improved their AI systems and reduced the frequency of these errors. But with the use of new reasoning systems, errors are rising. The latest OpenAI systems hallucinate at a higher rate than the company’s previous system, according to the company’s own tests.
The company found that o3 – its most powerful system – hallucinated 33% of the time when running its PersonQA benchmark test, which involves answering questions about public figures. That is more than twice the hallucination rate of OpenAI’s previous reasoning system, called o1. The new o4-mini hallucinated at an even higher rate: 48%.
When running another test called SimpleQA, which asks more general questions, the hallucination rates for o3 and o4-mini were 51% and 79%. The previous system, o1, hallucinated 44% of the time.
In a paper detailing the tests, OpenAI said more research was needed to understand the cause of these results. Because AI systems learn from more data than people can wrap their heads around, technologists struggle to determine why they behave in the ways they do.
“Hallucinations are not inherently more prevalent in reasoning models, though we are actively working to reduce the higher rates of hallucination we saw in o3 and o4-mini,” a company spokesperson, Gaby Raila, said. “We’ll continue our research on hallucinations across all models to improve accuracy and reliability.”
Hannaneh Hajishirzi, a professor at the University of Washington and a researcher with the Allen Institute for Artificial Intelligence, is part of a team that recently devised a way of tracing a system’s behaviour back to the individual pieces of data it was trained on. But because systems learn from so much data – and because they can generate almost anything – this new tool can’t explain everything. “We still don’t know how these models work exactly,” she said.
Tests by independent companies and researchers indicate that hallucination rates are also rising for reasoning models from companies such as Google and DeepSeek.
Since late 2023, Awadallah’s company, Vectara, has tracked how often chatbots veer from the truth. The company asks these systems to perform a straightforward task that is readily verified: summarise specific news articles. Even then, chatbots persistently invent information.
Vectara’s original research estimated that in this situation chatbots made up information at least 3% of the time and sometimes as much as 27%.
In the year and a half since, companies such as OpenAI and Google pushed those numbers down into the 1% or 2% range. Others, such as San Francisco startup Anthropic, hovered around 4%. But hallucination rates on this test have risen with reasoning systems. DeepSeek’s reasoning system, R1, hallucinated 14.3% of the time. OpenAI’s o3 climbed to 6.8%.
(The New York Times has sued OpenAI and its partner, Microsoft, accusing them of copyright infringement regarding news content related to AI systems. OpenAI and Microsoft have denied those claims.)
For years, companies like OpenAI relied on a simple concept: the more internet data they fed into their AI systems, the better those systems would perform. But they used up just about all the English text on the internet, which meant they needed a new way of improving their chatbots.
So these companies are leaning more heavily on a technique that scientists call reinforcement learning. With this process, a system can learn behaviour through trial and error. It is working well in certain areas, such as maths and computer programming. But it is falling short in other areas.
“The way these systems are trained, they will start focusing on one task – and start forgetting about others,” said Laura Perez-Beltrachini, a researcher at the University of Edinburgh who is among a team closely examining the hallucination problem.
Another issue is that reasoning models are designed to spend time “thinking” through complex problems before settling on an answer. As they try to tackle a problem step by step, they run the risk of hallucinating at each step. The errors can compound as they spend more time thinking.
The latest bots reveal each step to users, which means the users may see each error, too. Researchers have also found that in many cases, the steps displayed by a bot are unrelated to the answer it eventually delivers.
“What the system says it is thinking is not necessarily what it is thinking,” said Aryo Pradipta Gema, an AI researcher at the University of Edinburgh and a fellow at Anthropic.
This article originally appeared in The New York Times.
Written by: Cade Metz and Karen Weise
Photographs by: Erik Carter, Cayce Clifford and Kelsey McClellan
©2025 THE NEW YORK TIMES
