A recent study reveals that while ChatGPT may deliver confident answers, its inconsistency raises concerns about trustworthiness. Researchers from Washington State University (WSU) discovered that identical prompts could yield divergent responses, highlighting a fundamental limitation in the AI’s reliability. Specifically, ChatGPT struggled with unsupported claims, identifying false hypotheses only 16.4% of the time, despite an overall accuracy improvement. This discrepancy suggests the model often defaults to familiar language rather than complex reasoning. With only 72.9% of responses consistent across multiple trials in 2025, users must exercise caution, especially for critical decisions. ChatGPT’s design, rooted in text prediction rather than factual validation, can lead to misleading “hallucinations.” To mitigate risk, it’s essential to treat AI outputs as drafts, verify information through secondary sources, and maintain a critical approach. The study emphasizes that human oversight remains crucial in leveraging AI effectively, especially in high-stakes environments.
Source link
