Large language models' intelligence almost certainly does not work the same way as human intelligence.

Judged as of 2023. The claim (hedged 'almost certainly') is that large language models do not work the same way as human intelligence. True: the scientific consensus or predominant expert view in 2023 holds that LLM mechanisms (for example next-token prediction trained on text, transformer architecture) differ substantially from human cognition. Mostly true: the predominant view is that they differ, though a notable minority of research finds meaningful similarities in some respects. Mixed: expert opinion and evidence are roughly evenly split. Mostly false: substantial evidence suggests important functional similarities, and the strong claim of difference overstates matters. False: the evidence establishes that LLMs work essentially as human intelligence does.

Set before the research began.

1. E1 The debate over understanding in AI’s large language models PubMed Central (NCBI), article from PNAS · 21 March 2023

   “Indeed, these are precisely the abilities lacking in current AI systems, including state-of-the-art LLMs, although ever-larger LLMs have exhibited limited sparks of these general abilities.”

2. E2 The debate over understanding in AI’s large language models PubMed Central (NCBI), article from PNAS · 21 March 2023

   “finding that LLMs do, in some cases, exhibit humanlike responses on theory-of-mind tests (14, 75) and humanlike abilities and biases on reasoning assessments (76–78).”

3. E3 The debate over understanding in AI’s large language models PubMed Central (NCBI), article from PNAS · 21 March 2023

   “It could thus be argued that in recent years, the field of AI has created machines with new modes of understanding, most likely new species in a larger zoo of related concepts,”

4. E4 What Do NLP Researchers Believe? Results of the NLP Community Metasurvey arXiv (rendered via ar5iv) · 1 January 2021

   “respondents are split almost exactly in half on questions about the importance of artificial general intelligence, whether language models understand language, and the necessity of linguistic structure and inductive bias for solving NLP problems.”

A 2023 PNAS review (E1) says certain humanlike general abilities are lacking in current LLMs, apart from "limited sparks" in larger models. The same review (E3) describes AI as possibly a new mode of understanding rather than a copy of human intelligence, which supports Pinker's claim that the two differ. It also reports (E2) that some studies found humanlike responses on theory-of-mind tests and humanlike abilities and biases on reasoning assessments in some cases, so some findings point to similarities in some respects. A survey of NLP researchers (E4) found them split on whether language models understand language, but that question is about understanding, not about whether LLMs work the same way as humans.