Testing AI on language comprehension tasks reveals insensitivity to underlying meaning.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Cognitive Science

Background:

• Large Language Models (LLMs) are increasingly used in diverse applications, leading to claims of human-like linguistic capabilities.
• Moravec's Paradox highlights the difficulty of replicating human-like skills in AI, even for seemingly simple tasks.

Purpose of the Study:

• To systematically evaluate the language comprehension and reasoning abilities of state-of-the-art LLMs.
• To compare LLM performance against human baselines on a novel benchmark assessing linguistic understanding.

Main Methods:

• Seven state-of-the-art LLMs were tested on a benchmark of comprehension questions.
• Models were prompted in two settings (one-word and open-length replies) across a dataset of 26,680 data points.
• A human baseline was established by testing 400 individuals on the same prompts.

Main Results:

• LLMs performed at chance accuracy and exhibited considerable variability in their responses.
• Human participants significantly outperformed all tested LLMs in quantitative accuracy.
• Qualitative analysis revealed non-human errors in LLM language understanding.

Conclusions:

• Current LLMs, despite their utility, do not possess human-like language understanding.
• The lack of a compositional operator for grammatical and semantic information may explain LLMs' limitations.
• Further research is needed to bridge the gap between AI and human linguistic capabilities.