Position: We Can’t Understand AI Using our Existing Vocabulary

John Hewitt, Robert Geirhos, Been Kim

TL;DR: To understand AI, we must develop neologisms: new words for human and machine concepts.

Abstract:

This position paper argues that, in order to understand AI, we cannot rely on our existing vocabulary of human words. Instead, we should
strive to develop neologisms: new words that
represent precise human concepts that we want
to teach machines, or machine concepts that we
need to learn. We start from the premise that
humans and machines have differing concepts.
This means interpretability can be framed as a
communication problem: humans must be able to
reference and control machine concepts, and communicate human concepts to machines. Creating
a shared human-machine language through developing neologisms, we believe, could solve this
communication problem. Successful neologisms
achieve a useful amount of abstraction: not too
detailed, so they’re reusable in many contexts, and
not too high-level, so they convey precise information. As a proof of concept, we demonstrate how
a “length neologism” enables controlling LLM
response length, while a “diversity neologism” allows sampling more variable responses. Taken
together, we argue that we cannot understand AI
using our existing vocabulary, and expanding it
through neologisms creates opportunities for both
controlling and understanding machines better.

Lay Summary: Understanding AI systems is a critical problem as they become increasingly deployed in the world. We frame the problem of understanding and controlling AI systems as a communication problem, like how we try to communicate complex concepts between humans. In this opinion piece, we argue that we need to develop new words for ways in which AI systems see the world, and new words that teach AI systems how we see the world, in order to achieve this communication. Akin to how humans invent new words to discuss new or complex things, we need to do this with AI systems. We provide arguments for this position and early experiments showcasing how we might accomplish this in the future.

Primary Area: Model Understanding, Explainability, Interpretability, and Trust

Keywords: Interpretability, understanding, control

Submission Number: 346