ChatGPT’s Confidence Masks Concerning Inconsistencies, New Study Finds

A new study reveals a troubling pattern in ChatGPT’s reliability, suggesting that the AI system’s confident tone may obscure significant flaws in its reasoning abilities. Researchers at Washington State University (WSU) discovered that when asked identical questions repeatedly, ChatGPT frequently delivers contradicting answers without any changes to the input prompt.

The research team, led by Mesut Cicek, conducted extensive testing by posing hundreds of hypotheses from published scientific research papers to ChatGPT, asking the system to determine whether each claim was true or false. By running identical prompts ten times, they uncovered a concerning inconsistency in the AI’s responses.

“We’re not just talking about accuracy, we’re talking about inconsistency because if you ask the same question again and again, you come up with different answers,” explained Cicek, highlighting a fundamental issue with the system’s reliability.

This inconsistency manifested starkly in some cases, with the AI oscillating between labeling the same claim as “true” in one instance and “false” in another, despite no changes to the input. Such reversals expose critical limitations in how large language models evaluate factual information.

At first glance, ChatGPT’s overall accuracy appeared reasonably strong, improving from 76.5 percent in 2024 to 80 percent in 2025. However, when researchers factored out the element of random chance inherent in true-or-false questions, the effective accuracy plummeted to around 60 percent—equivalent to a low D grade in academic terms.

The study identified specific weaknesses in ChatGPT’s reasoning capabilities. The AI performed particularly poorly when evaluating unsupported hypotheses, correctly identifying false claims just 16.4 percent of the time in 2025. This suggests a persistent bias toward agreement, with the system defaulting to “yes” responses because matching familiar language patterns is easier than detecting flawed ideas.

Furthermore, the research revealed that across ten repeated runs, only 72.9 percent of responses in 2025 remained consistently correct. This instability presents a significant challenge for users relying on these systems for decision-making, as a single response might appear reliable while repeated checks expose its fundamental fragility.

The AI demonstrated stronger performance on straightforward cause-and-effect relationships but struggled with context-dependent claims—precisely the kind of nuanced judgments central to everyday business decisions, from pricing strategies to policy tradeoffs.

These findings highlight the inherent limitations of large language models, which are trained on vast text datasets and function by predicting likely word sequences rather than verifying facts against reality. This design helps generate fluent, confident-sounding responses even when the system lacks a grounded way to assess their truthfulness.

OpenAI, the creator of ChatGPT, acknowledges that the system can produce “hallucinations”—responses that appear certain but contain factual inaccuracies. This combination of confidence and unreliability makes the system particularly challenging to use effectively, as incorrect answers can seem convincing enough to trust.

For businesses, scientific teams, and other professional users, this weakness transforms what might be a useful productivity tool into a potential source of risk. While AI-generated summaries can accelerate planning processes, a single flawed judgment could misdirect product development, budgeting decisions, or marketing campaigns.

“They just memorize, and they can give you some insight, but they don’t understand what they’re talking about,” Cicek noted, emphasizing the gap between the system’s linguistic fluency and its actual comprehension.

The research suggests several practical approaches for more responsible AI use. Users should treat AI outputs as preliminary drafts rather than final decisions, run identical prompts multiple times to identify inconsistencies, verify information against established sources, and consider potential missing context.

Though the study focused specifically on business hypotheses from open-access research evaluated through ChatGPT, its findings carry broader implications for AI reliability. Despite appearing more polished in its 2025 version compared to 2024, the system has not fundamentally evolved into a dependable reasoning tool.

As organizations increasingly integrate AI systems into their workflows, WSU’s research serves as a timely reminder: human expertise remains essential for evaluating AI-generated content, particularly when the stakes are high and the answers seem suspiciously straightforward.

The complete study has been published in Rutgers Business Review.