8legs: Rutgers, usa

#1 out of 1

College professors gave ChatGPT a science exam, and its grade was a 'low D'

A Washington State University study found ChatGPT gave different answers to the same question across ten repeated runs.
The system showed a bias toward agreement, performing poorly on false claims in 2025.
Overall accuracy was around 60% when checks accounted for random guessing, close to a low grade.
ChatGPT can sound confident yet may not reflect grounded, verifiable knowledge.
Researchers warn against treating AI as a final decision-maker in high-stakes tasks.
The study tested hundreds of hypotheses from published scientific research papers.
Repeated questioning revealed instability in answers even with unchanged prompts.
The study notes better performance on simple cause-and-effect tasks than on context-dependent judgments.
Experts advise using AI as a drafting tool with human review for decisions.
The Rutgers Business Review paper calls for broader comparisons and longer prompt runs.

Vote 0