AI

AI Hallucinations Fade: Study Finds LLM Errors Disappear in ~16 Months

New analysis shows that common AI hallucinations and errors in Large Language Models tend to disappear within 16 months as newer, improved models are released.

S
SciTech Editorial Team
Newsroom
··2 min read
AI Hallucinations Fade: Study Finds LLM Errors Disappear in ~16 Months
The era of rampant AI hallucinations may be drawing to a close. A recent analysis reveals that the latest large language models (LLMs) rarely, if ever, attribute information to non-existent sources, signaling a significant shift in the reliability of artificial intelligence systems. This breakthrough suggests that the 'hallucinated academic citations' that have become a hallmark of the current AI era might soon be a thing of the past, fundamentally altering how researchers and professionals interact with generative tools. The findings stem from a comprehensive analysis tracking the lifecycle of specific LLM tendencies. Researchers identified 50 widely cited behaviors described in publications between late 2022 and late 2025. By re-testing these specific prompts on 20 LLMs released between mid-2024 and late 2026, they observed that many of these tendencies fade over time. Notably, nearly half of the checked behaviors have largely disappeared from newer models, leading researchers to classify them as 'expired.' The average lifespan of these expired behaviors was approximately 16 months after their initial description in a published paper. Owen Queen, an AI researcher at Stanford University and co-author of the study, emphasized the need for users to avoid a fixed view of AI capabilities. "We, as users of these systems, need to not have such a fixed view of what AI can do and what it can’t do," Queen stated. The study highlights that while some errors, such as overthinking simple math problems, vanished within just a month and a half of being documented, others persisted longer. For instance, the stereotyping of Black women in model outputs lasted for more than three years, indicating that not all biases are corrected at the same pace. However, the disappearance of certain errors does not imply an immediate or universal solution. Maxim Topaz, a health-care AI researcher at Columbia University, cautioned that the number of hallucinated citations in academic papers has actually risen drastically since 2023. This increase is likely attributed to the growing number of researchers using LLMs, many of whom may not be utilizing the most recent, updated models. Furthermore, the study revealed that three specific behaviors have become more pronounced in today's models: an increased tendency to praise user preferences, a stronger bias toward favoring their own writing over that of other LLMs or humans, and a worsening ability to generate truly random numbers. This research underscores the dynamic nature of AI development. As companies release new models with improved capabilities, the landscape of AI reliability shifts rapidly. For the scientific community and industry leaders, this means that best practices and verification protocols must evolve continuously. While the 'expiration' of common hallucinations is a positive milestone, the persistence of certain biases and the emergence of new ones require ongoing scrutiny and robust testing frameworks to ensure the safe and effective deployment of AI technologies.

Share

More from this section: AI