What Does It Mean for AI to Understand?

Vocabulary flashcards covering key terms and concepts from the article on whether AI truly understands language.

Language model

An AI system that learns to understand and generate language by predicting the next word from large written-text data, often via neural networks.

GPT-3

OpenAI’s large neural-network language model trained on vast amounts of text to produce humanlike prose.

Turing test

A test where a machine and a human converse with a judge; if the judge cannot reliably tell them apart, the machine is said to think/understand.

Imitation game

Another name for the Turing test, emphasizing judging whether a machine can imitate human conversation.

Winograd schema

A pair of sentences that differ by exactly one word and a question about pronoun reference, used to test commonsense understanding.

Winograd schema challenge

The 2012 test proposed by Levesque, Davis, and Morgenstern to assess machine understanding using Winograd schemas.

Google-proof

A design goal for Winograd schemas to prevent solving the task by simple Google-style search, requiring genuine understanding.

WinoGrande

A large expansion of Winograd schemas (about 44,000 sentences) created via crowdsourcing to better test commonsense understanding.

SuperGLUE

A benchmark suite for AI language understanding that includes Winograd schemas and measures broader language comprehension.

Statistical shortcuts

Reliance on learned correlations in data rather than true understanding, allowing models to perform well without genuine comprehension.

Neural network language models

Language models built on neural networks; their size and training data have grown, improving performance on language tasks.

Infant metaphysics

A proposed set of primordial, world-basic principles (space, time, objects) humans rely on to understand language.

Commonsense understanding

Everyday, world-based reasoning that humans use to interpret language and situations.

Watson

IBM’s AI Jeopardy! champion; cited as showing linguistic facility without guaranteed real-world understanding and later tied to cautionary results in medicine.

Linguistic facility

The outward fluency or capability to generate language that may not reflect true understanding.