Text Understanding in GPT-4 vs Humans
This work addresses the problem of evaluating AI text understanding for researchers and educators, though it is incremental as it builds on existing testing methods.
The study compared GPT-4's text understanding to humans using standardized and more difficult comprehension tests, finding that GPT-4 performed slightly better on the standardized test and considerably better on the difficult passages, with evidence of generalization and inference.
We examine whether a leading AI system GPT4 understands text as well as humans do, first using a well-established standardized test of discourse comprehension. On this test, GPT4 performs slightly, but not statistically significantly, better than humans given the very high level of human performance. Both GPT4 and humans make correct inferences about information that is not explicitly stated in the text, a critical test of understanding. Next, we use more difficult passages to determine whether that could allow larger differences between GPT4 and humans. GPT4 does considerably better on this more difficult text than do the high school and university students for whom these the text passages are designed, as admission tests of student reading comprehension. Deeper exploration of GPT4 performance on material from one of these admission tests reveals generally accepted signatures of genuine understanding, namely generalization and inference.