AI hallucinations worsen as advanced models invent facts and citations, risking public health
By isabelle // 2025-11-19
Copy
 
  • AI fabricates more than half of its academic citations in mental health research.
  • Its reliability plummets for topics outside the mainstream establishment narrative.
  • AI invents fake citations with real-looking links, making detection difficult.
  • The system is trained to guess and invent rather than admit uncertainty.
  • Newer, more powerful AI models are hallucinating at increasingly higher rates.
A groundbreaking study has exposed a dangerous flaw lurking within the artificial intelligence systems that millions now trust for information. Researchers at Australia’s Deakin University discovered that ChatGPT, specifically the GPT-4o model, is fabricating academic citations at an alarming rate when tasked with writing mental health literature reviews. This isn't a minor glitch; it is a fundamental failure that reveals how these "authoritative" AI models are built on a foundation of inaccuracies, potentially steering people toward harmful decisions about their health and well-being. The investigation found that more than half of all citations generated by the AI, a staggering 56%, were either completely fabricated or contained significant errors. Only 44% of the references were both real and accurate. This means that anyone using the AI for research, including students, writers, or even health practitioners, is being fed a diet of false information disguised as academic fact. The problem of AI "hallucination" is not random. The study revealed that the AI's reliability varies dramatically based on the topic. For well-known subjects like major depressive disorder, only 6% of citations were fake. However, when venturing into less mainstream areas like binge eating disorder and body dysmorphic disorder, the fabrication rates skyrocketed to 28% and 29% respectively. This pattern is a red flag for anyone seeking information outside the establishment narrative. It suggests that mainstream AI models, trained on so-called "authoritative" sources like Wikipedia and major media, are least reliable on topics that challenge conventional wisdom. For those exploring natural health, clean diets, or traditional healing methods, these models are likely to provide fabricated or erroneous support. The fabrications are particularly deceptive. When the AI invented a citation, 64% of the time it provided a real-looking Digital Object Identifier (DOI) that linked to a genuine, but completely unrelated, academic paper. This makes the falsehoods incredibly difficult to spot without rigorous, old-fashioned fact-checking. The root of this problem lies in how these AI models are trained and evaluated. According to research highlighted by IBM, AI developers create benchmarks that reward models for providing an answer and penalize them for admitting uncertainty. This incentivizes the AI to guess, bluff, and invent information rather than honestly stating "I don't know." Testing of OpenAI's o4-mini model confirmed this flawed incentive structure. The model guessed frequently but was wrong 75% of the time. It is being taught that a confident lie is better than a cautious truth. This is not a bug; it is a feature of a system designed to prioritize the appearance of intelligence over genuine accuracy. The consequences of this are not merely academic. When an AI hallucinates information about mental health treatments, nutrition, or medical advice, it can drive individuals to make poor choices that directly impact their physical and mental well-being. People who trust these systems are being misled by a sophisticated confidence game.

New models hallucinate more, not less

This crisis of accuracy is getting worse, not better. Despite promises from tech giants, newer and more powerful "reasoning" models are actually hallucinating more frequently. OpenAI's own tests found its new o4-mini model hallucinated 48% of the time on one benchmark, a significant increase from previous versions. The situation creates a perfect storm for the dissemination of inaccurate information. As one expert noted, these errors are becoming more subtle and embedded within plausible narratives, making them harder for the average user to detect. The very fluency of these models makes their falsehoods more convincing. This research serves as a critical warning. It underscores the vital importance of never trusting AI output at face value. Every claim, and every citation, must be subjected to rigorous human verification and due diligence. In an age where technology promises convenience, this study is a reminder that true knowledge requires skepticism and effort. Relying on AI for truth is a dangerous gamble, and it's one that could cost us not just our intellectual integrity, but our health. Sources for this article include: StudyFinds.org LiveScience.com NYTimes.com IBM.com
Copy