Can GPT-4 formulate and test a novel hypothesis? Yes and no
In this paper, the author challenges the latest version of ChatGPT Plus, powered by the GPT-4 algorithm, to generate a novel mathematical or scientific hypothesis, and then generate a Matlab script capable of testing its hypothesis through one or more numerical experiments. Five different responses of GPT-4 are examined, with varying degrees of success. In one case, GPT-4 presented a well-studied mathematical theorem involving prime number twins as its own “novel” hypothesis. In another case, GPT-4 presented an overly general hypothesis related to graph theory with a Matlab program with an undefined function. In three other attempts, GPT-4 seemed to successfully generate simple (and false) hypotheses with Matlab scripts that correctly disproved these hypotheses. It is concluded that the generation and testing of novel hypotheses may indeed be within reach as this technology continues to advance in the future.
Email Address of Submitting Authormike.email@example.com
Submitting Author's InstitutionVanderbilt University
Submitting Author's Country
- United States of America