Study Reveals ChatGPT's Poor Accuracy on Scientific Questions—Key Findings Exposed.

How ChatGPT Performed in a Scientific Accuracy Test

According to TSN.ua: A study led by Mesut Çiçek at Washington State University found that ChatGPT gives inconsistent answers to scientific questions, with low accuracy—especially when dealing with unproven hypotheses. The research showed that the AI can produce different responses to the same question, even when it is repeated up to ten times.

Answer Reliability Under Scrutiny

In 2025, ChatGPT’s overall accuracy on scientific questions was around 80%. But after factoring in random guessing, that figure dropped to just 60%. When it came to identifying false statements about unconfirmed hypotheses, the system was correct only 16.4% of the time. Moreover, just 72.9% of answers stayed consistently correct across ten identical queries—highlighting serious issues with the system’s stability and reliability.

The researchers stress that AI should be treated as a supporting tool, given its limitations in both precision and consistency. These findings underscore the need for a critical mindset when using artificial intelligence in scientific research and other fields.

The study’s outcomes could significantly shape how scientists and professionals across various industries integrate AI into their work. Acknowledging the constraints of AI in terms of accuracy and stability encourages more cautious and informed use of these technologies—an increasingly important consideration as AI evolves rapidly. This may also drive further research aimed at refining algorithms and systems to boost their dependability in both academic and real-world applications.