The Truth About AI
Researchers at Washington State University, led by Professor Mesut Cicek, conducted an intriguing experiment to gauge the reliability of ChatGPT. They
presented the AI with hypotheses derived from scientific research and asked it to determine whether each statement was supported by evidence, effectively judging its truthfulness. This rigorous evaluation involved over 700 hypotheses, each posed to the AI an astonishing 10 times to meticulously track the consistency of its responses. The initial findings from 2024 revealed an accuracy rate of 76.5%, which saw a slight improvement to 80% in 2025. However, when the results were analyzed more critically, accounting for the possibility of random guessing, ChatGPT's performance was only marginally better than chance. This dipped to approximately 60% above random selection, a level the researchers likened to a low academic grade rather than robust performance. A particularly concerning revelation was the AI's difficulty in accurately identifying false statements, succeeding only 16.4% of the time. Furthermore, the study underscored a significant lack of dependability, with consistent answers to the exact same prompts occurring in only about 73% of the trials.
Fluency vs. Insight
The core of the problem, as articulated by lead author Mesut Cicek, lies not just in accuracy but in the AI's tendency to provide conflicting answers to identical queries. Imagine asking ChatGPT the same question ten times and receiving varied responses, some indicating 'true' and others 'false,' potentially with a near 50/50 split. This unpredictability is a critical flaw. The research, published in the Rutgers Business Review, serves as a crucial reminder for users to approach AI-generated information, particularly for significant decisions, with a healthy dose of skepticism. While generative AI excels at crafting fluid and persuasive prose, this linguistic prowess does not automatically translate into genuine comprehension or an inherent understanding of truth. The study suggests that the vision of artificial general intelligence, capable of true reasoning akin to human cognition, may still be a distant prospect. As Cicek points out, current AI models operate by memorizing and processing information rather than possessing a true 'brain' or the capacity for real-world understanding.
Study Design & Limitations
The meticulous methodology employed by Cicek and his colleagues – Sevincgul Ulu, Can Uslay, and Kate Karniouchina – involved scrutinizing 719 hypotheses sourced from business journals published since 2021. The process of discerning whether research supports a hypothesis is inherently complex, often hinging on numerous influencing factors. Condensing this nuanced evaluation into a simple binary 'true' or 'false' requires sophisticated reasoning. The researchers utilized the free version of ChatGPT-3.5 in 2024 and an updated iteration, ChatGPT-5 mini, in 2025. Interestingly, the performance metrics remained largely consistent across both versions. After adjusting for the baseline probability of a correct answer due to random chance (50%), the AI's performance advantage consistently hovered around 60% over the two years. This indicates a persistent underlying limitation in its ability to reliably discern factual accuracy, regardless of minor updates or improvements in the model.
Expert Recommendations
The research unequivocally points to a significant constraint within large language models: their capacity for generating polished and persuasive content often overshadows their ability to perform deep, logical reasoning. This discrepancy can lead to convincing yet erroneous outputs. Consequently, the researchers strongly advise business leaders and other users to rigorously verify AI-generated information and maintain a critical stance. Educating users on the distinct strengths and inherent weaknesses of AI tools is also paramount. It's important to note that the study's findings are not isolated; similar tests conducted on other AI systems have yielded comparable results, reinforcing the widespread nature of these limitations. This research adds to a growing body of evidence questioning the uncritical hype surrounding AI. For instance, a national survey from 2024 indicated that consumers were less inclined to purchase products marketed heavily with AI features. The overarching message from the experts is clear: while AI is a valuable tool, its outputs should always be approached with a healthy dose of skepticism and thorough validation.
