#1
#1 out of 122.86%
technology4h ago
ChatGPT Took a Science Exam Designed by College Professors and Scored a Low D, and That Should Worry You
- Latest finding: ChatGPT's effective accuracy drops to about 60% after adjusting for random guessing.
- Study notes a bias toward agreement, reducing reliability when claims are unsupported.
- Ten identical prompts yielded only 72.9% consistency across responses in 2025.
- The study warns that polished output is not the same as dependable reasoning.
- ChatGPT showed strongest performance on simple cause-and-effect questions.
- Inconsistency appeared as responses varied across identical prompts.
- Researchers stressed the need for human experts to verify logic.
- The Rutgers Business Review study examined hypotheses from open-access research.
- The study quantified false-claim detection, noting only 16.4% correct identification in 2025.
Vote 0