Errors Tend to Occur with AI-Generated Content
Paul Taylor/Getty Images
AI chatbots from tech giants like OpenAI and Google have seen several inference upgrades in recent months. Ideally, these upgrades would lead to more reliable answers, but recent tests indicate that performance may be worse than that of previous models. Errors called “hallucinations,” particularly in the “hagatsuki” category, have been persistent issues that developers have struggled to eliminate.
Hallucination is the broad term used to describe specific errors generated by large-scale language models (LLMs) from organizations like OpenAI’s ChatGPT and Google’s Gemini. It primarily refers to instances where these models present false information as fact, but it can also describe instances where a generated answer is accurate yet irrelevant to the question posed.
A technical report from OpenAI evaluating the latest LLMs revealed that the O3 and O4-MINI models, released in April, exhibit significantly higher hallucination rates compared to earlier O1 models introduced in late 2024. For instance, if O4-MINI had a summary accuracy of 33%, the hallucination rate for O3 was similarly at 33%, whereas the O1 model maintained a rate of only 16%.
This issue is not exclusive to OpenAI. The popular leaderboard showcases various inference models from different companies assessing their hallucination rates, including the DeepSeek-R1 model. This model has shown increased hallucination rates compared to previous versions, undergoing several reasoning steps before reaching a conclusion.
An OpenAI spokesperson stated, “We are actively working to reduce hallucination rates in O3 and O4-MINI. Hallucinations are inherently more common in inference models. We will continue our research across all models to enhance accuracy and reliability.”
Some potential applications of LLMs can be significantly impeded by hallucinations. Models that frequently produce misinformation are unsuitable as research assistants, and a bot stating fictitious legal cases could endanger lawyers. Customer service agents falsely citing obsolete policies can also create significant challenges for businesses.
Initially, AI companies believed they would resolve these issues over time. Historically, models had shown reduced hallucinations with each update, yet the recent spikes in hallucination rates complicate this narrative.
Vectara’s leaderboard ranks models based on their consistency in summarizing documents. This indicates that for systems from OpenAI and Google, “hallucination rates are roughly comparable for inference and irrational models,” as noted by Forest Shen Bao from Vectara. Google has not provided further comments. For leaderboard assessments, the specific rates of hallucinations are less significant than each model’s overall ranking, according to Bao.
However, these rankings may not effectively compare AI models. For one, different types of hallucinations are often conflated. The Vectara team pointed out that the DeepSeek-R1 model demonstrated a 14.3% hallucination rate, but many of these hallucinations were “benign,” being logically deduced yet not appearing in the original text.
Another issue with these rankings is that tests based on text summaries “reveal nothing about the percentage of incorrect output” for tasks where LLMs are applied, as stated by Emily Bender at Washington University. She suggests that leaderboard results don’t provide a comprehensive evaluation of this technology, particularly since LLMs are not solely designed for text summarization.
These models generate answers by repeatedly answering the question, “What is the next word?” to formulate responses, thus not processing information in a traditional sense. However, many technology companies continue to use the term “hallucination” to describe output errors.
“The term ‘hallucination’ is doubly problematic,” says Bender. “On one hand, it implies that false output is abnormal and could potentially be mitigated, while on the other hand, it inaccurately anthropomorphizes the machine since large language models lack awareness.”
Arvind Narayanan from Princeton University argues that the issue extends beyond hallucinations. Models can also produce errors by utilizing unreliable sources or outdated information. Merely increasing training data and computational power may not rectify the problems.
We may have to accept the reality of error-prone AI, as Narayanan mentioned in a recent social media post. In some circumstances, it may be prudent to use such models solely for tasks requiring fact-checking. The best approach might be to avoid relying on AI chatbots for factual information altogether.
Source: www.newscientist.com
