The more advanced artificial intelligence (AI) becomes, the more it tends to "hallucinate" and provide false or inaccurate information.

According to research by OpenAI, its most recent and powerful reasoning models—o3 and o4-mini—exhibited hallucination rates of 33% and 48%, respectively, when tested using the PersonQA benchmark. This is more than double the rate seen in the older o1 model. Although o3 offers more accurate responses compared to earlier versions, it also shows a higher tendency toward incorrect hallucinations.

This trend raises concerns about the accuracy and reliability of large language models (LLMs), such as AI chatbots, said Eleanor Watson, an IEEE member and AI ethics engineer at Singularity University.

"When a system produces fabricated information—like made-up facts, citations, or events—with the same fluency and coherence it uses for factual content, it can mislead users in subtle but significant ways," Watson told Live Science.

Related: Cutting-edge AI models from OpenAI and DeepSeek experience 'complete collapse' when faced with overly complex problems, study finds

Experts emphasize that this hallucination issue highlights the importance of carefully evaluating and monitoring the outputs generated by LLMs and reasoning models.

Do AIs dream of electric sheep?

The key feature of a reasoning model is its ability to tackle complex tasks by breaking them into smaller parts and developing strategies to solve each one. Unlike models that rely solely on statistical probability to generate answers, reasoning models create problem-solving strategies similar to human thinking.

To sign up for the Live Science daily newsletterIn order for AI to generate creative and potentially novel solutions, it must engage in some level of hallucination—otherwise, it would be limited to regurgitating data it has already learned.

"It's crucial to understand that hallucination is not a flaw but a feature of AI," said Sohrob Kazerounian, an AI researcher at Vectra AI, in an interview with Live Science. "As a colleague once put it, 'Everything an LLM generates is a hallucination. It's just that some of those hallucinations happen to be true.' If an AI only produced exact copies of what it had seen during training, AI would become nothing more than a massive search engine."

"This would mean writing only code that had been written before, discovering only molecules whose properties were already known, and answering only homework questions that had previously been asked. You wouldn't be able to ask an LLM to write lyrics for a concept album centered around AI singularity, combining the styles of Snoop Dogg and Bob Dylan."

Effectively, LLMs and the AI systems they power need to hallucinate in order to produce original content rather than simply repeating existing knowledge. Conceptually, this is similar to how humans dream or imagine scenarios to spark new ideas.

Thinking too much outside the box

However, AI hallucinations become problematic when the goal is to deliver precise and correct information, especially if users accept the output without verification.

"This is particularly concerning in fields where factual accuracy is critical, like medicine, law, or finance," Watson explained. "Although more advanced models may reduce obvious factual errors, subtler forms of hallucination persist. Over time, these fabrications undermine trust in AI systems and can lead to real-world harm when users act on unverified information."

Moreover, this challenge seems to grow as AI technology progresses. "As models improve, errors often become less obvious yet harder to detect," Watson noted. "Fabricated content is now embedded within convincing narratives and logical reasoning chains. This creates a unique danger: users may not realize there are errors and could treat the output as definitive. The issue shifts from filtering out clear mistakes to identifying nuanced distortions that only surface under close inspection."

Kazerounian supports this view. "Despite widespread optimism that AI hallucinations will decrease over time, evidence suggests that newer reasoning models may actually hallucinate more frequently than simpler ones—and there’s no consensus on why this happens," he said.

The situation is further complicated by the difficulty in understanding how LLMs arrive at their conclusions, drawing a parallel to how we still don’t fully comprehend how the human brain functions.

In a recent essay, Dario Amodei, CEO of AI company Anthropic, pointed out the lack of transparency regarding how AIs generate responses. "When a generative AI summarizes a financial document, we have no concrete understanding of why it makes specific word choices or why it occasionally errs despite usually being correct," he wrote.

Kazerounian stressed that the consequences of AI generating false information are already very real. "There is no reliable method to ensure an LLM correctly answers questions about any given dataset it has access to," he said. "Instances of non-existent references, customer service chatbots inventing company policies, and other inaccuracies are now alarmingly common."

Crushing dreams

Both Kazerounian and Watson told Live Science that eliminating AI hallucinations entirely may be difficult. However, there might be ways to reduce their impact.

Watson proposed that "retrieval-augmented generation," which ties a model's output to verified external knowledge sources, could help ground AI-generated content in factual data.

"Another strategy involves structuring the model's reasoning process. By prompting it to verify its own outputs, compare different viewpoints, or follow logical steps, scaffolded reasoning frameworks minimize unchecked speculation and enhance consistency," Watson explained. He added that this could be supported by training methods designed to encourage models to prioritize accuracy, along with reinforcement learning from human or AI evaluators to promote more disciplined and fact-based responses.

The above is the detailed content of Why is AI halllucinating more frequently, and how can we stop it?. For more information, please follow other related articles on the PHP Chinese website!