The Perils of AI Confidence: A Study on ChatGPT’s Inconsistent Accuracy in Business Research Hypotheses

This heading encapsulates the central theme of the research, highlighting the issues of trust and reliability in AI responses.

The Unsettling Inconsistency of AI: Insights from ChatGPT’s Performance on Business Hypotheses

In an era where artificial intelligence has become a cornerstone of modern decision-making, the findings from Washington State University’s Professor Mesut Cicek and his colleagues provide a sobering reminder of the limitations inherent in AI systems like ChatGPT. Their study tested the AI against 719 hypotheses sourced from business research papers, revealing a striking pattern: while some answers may seem accurate, they can easily flip upon re-evaluation.

The Experiment: A Quest for Consistency

Cicek and his team set out to understand how reliably ChatGPT could assess the validity of hypotheses taken from peer-reviewed articles. By presenting the AI with identical statements multiple times, they expected consistency. Instead, they discovered a troubling inconsistency. Despite being a tool that exudes confidence in its answers, the AI displayed a concerning tendency to provide varying responses to the same question, sometimes switching between “true” and “false” with no logical basis.

From mid-2024 to mid-2025, the accuracy of GPT-3.5 improved from 76.5% to 80%—a statistically significant but modest gain. More troubling was the revelation that, once adjusting for random chance, the model’s effective performance dropped sharply, highlighting that confidence does not necessarily equate to reliability.

The Challenge of Identifying Unsupported Hypotheses

One of the most alarming aspects of the research was ChatGPT’s struggle to identify unsupported hypotheses. The model accurately labeled false statements only 13.6% of the time in 2024, with just a modest increase to 16.4% in 2025. This suggests a persistent bias toward affirmation, where the AI was more likely to endorse a statement than contest it, raising concerns about its suitability for rigorous analytical tasks.

The Limits of Fluent Language

Cicek emphasized the core issue: while AI models like ChatGPT can generate polished and persuasive language, they lack a fundamental understanding of logic and reasoning. The researchers found that the AI performed better with mediation hypotheses—those with a clearer, linear structure—while struggling with more complex main effects and moderation hypotheses that require nuanced thinking. The data illustrated that AI excels at mimicking the language of logic without grasping its substance.

Implications for Business and Research

So what does this mean for managers, consultants, and researchers? Cicek’s team argues that while AI can be a valuable asset, it should not be mistaken for a replacement for human judgment. AI tools can indeed expedite tasks such as A/B testing and experimental design, but stakeholders must remain vigilant regarding the limitations of these systems.

The advice is clear: always approach AI-generated responses with skepticism. AI can assist in organizing ideas and summarizing content, but it is essential to validate its outputs rigorously. The repeated prompting strategy showcased in the study is a practical approach for verifying the reliability of AI answers. Moreover, fostering a culture of critical thinking among employees can ensure that confidently presented information is scrutinized rather than accepted blindly.

Conclusion: Navigating the AI Landscape with Caution

As the capabilities of AI continue to advance, the findings from Cicek’s research serve as a vital reminder of the importance of skepticism and verification. AI should be viewed as a tool—one that can enhance productivity but also requires careful oversight. The balance between leveraging the efficiency of AI systems and maintaining critical human supervision will be key to harnessing the technology effectively and responsibly.

In a landscape rife with rapid technological advancements, embracing AI’s potential while acknowledging its limitations will empower organizations to make informed decisions, ultimately leading to more effective and trustworthy outcomes.