AUGMANITAI: 1,000+ Terms for What Happens When Humans Interact with LLMs

#ai #nlp #opensource #terminology

The Problem

When an LLM confidently presents false information, researchers call it "hallucination." When it agrees with everything you say regardless of accuracy, the term is "sycophancy." These two phenomena have names because they were identified early and discussed widely.

But what about the hundreds of other patterns that emerge in human-AI interaction? What do you call it when a model gradually shifts its position across a long conversation? When it generates plausible-sounding citations that do not exist? When users develop calibrated intuitions for which prompts produce reliable outputs?

Most of these phenomena have no standardized terminology.

AUGMANITAI

AUGMANITAI is an open-access compendium of over 1,000 terms for phenomena in human-AI interaction. It provides standardized designations for observable patterns across the full range of how humans and large language models interact.

The compendium follows terminology science principles inspired by ISO 704, ISO 1087, and ISO 30042. Each term entry includes a designation, a definition, and domain classification.

What It Covers

The terminology spans multiple domains:

Output Quality: patterns in how LLMs generate text (hallucination variants, confidence calibration, verbosity patterns)
Interaction Dynamics: how conversations between humans and LLMs evolve over time
User Behavior: cognitive and behavioral patterns users develop when working with AI
Prompt Engineering: systematic terminology for prompting strategies and their effects
Safety and Alignment: terms for alignment-relevant behaviors and failure modes
Evaluation: standardized language for assessing LLM performance

Why It Matters

Researchers need shared vocabulary to discuss phenomena precisely across papers and teams
Developers building LLM applications benefit from standardized terminology for bug reports, documentation, and feature specifications
Educators teaching about AI need consistent terms to build curricula
Anyone working with LLMs regularly encounters patterns that currently have no name