AI Jargon Glossary: Demystifying the Core Terms of Generative AI

The Numbers Behind the Words

Before diving into the alphabetical glossary, it is helpful to understand how artificial intelligence models process human language. AI models do not read words the way humans do. Instead, they break text down into smaller units called tokens. Understanding these numbers is crucial for managing API costs and working within model limitations.

According to documentation from OpenAI and Google, here is how tokens translate to everyday English:

Tokenization is highly context-dependent. For example, OpenAI's tokenizer assigns different token IDs to the word "red" depending on capitalization and spacing: " red" (lowercase with a space) is token 2266, " Red" (capitalized with a space) is token 2297, and "Red" (capitalized at the start of a sentence) is token 7738 [1].

A to Z

Abstention Rate

The frequency with which an AI model declines to answer a question when it is uncertain, rather than guessing. In OpenAI's SimpleQA evaluation, the reasoning model gpt-5-thinking-mini demonstrated an abstention rate of 52% (resulting in a 26% error rate), whereas o4-mini had a 1% abstention rate (resulting in a 75% error rate) [3].

AI Accountability

The process of holding entities answerable for the risks and harms of the AI systems they develop or deploy. This relies heavily on transparency, governance functions, and "accountability inputs" like system evaluations [4].

AI Assurance

A set of informational and evaluative practices that provide justified confidence that an AI system operates in a trustworthy fashion and as claimed within its specific context [4].

AI Audit

An independent evaluation of an AI system's performance or processes against transparent, specified criteria to verify claims or assess its trustworthiness [4].

AI Model

A component of an AI system that implements computational, statistical, or machine learning techniques to produce outputs from a given set of inputs [4].

AI Red-Teaming

A structured, adversarial testing effort designed to find flaws, vulnerabilities, and security risks in an AI system, such as harmful outputs or unexpected behaviors, in a controlled environment [4].

AI System

An engineered or machine-based system that can, for a given set of human-defined objectives, generate outputs such as predictions, recommendations, or decisions that influence real or virtual environments [4].

Context Window (Length Limits)

The maximum amount of information an AI model can process and remember in a single conversation. On paid Claude plans, the newest models support context windows up to 1 million tokens, while others support 500K or 200K tokens [5]. When conversations approach this limit, some systems use automatic context management to summarize earlier messages [5].

Hallucination

An instance where an AI model confidently generates a plausible but false statement. OpenAI research indicates that standard training and evaluation procedures often reward guessing over acknowledging uncertainty, which contributes to these errors [3].

Memory

A feature in conversational assistants (such as Claude) that allows the AI to store details across different chats to reduce repetitive explanations. Users can view, edit, or delete saved topics. By default, sensitive topics like health, politics, or religious beliefs are not stored unless explicitly enabled by the user [6].

Tokens

The basic building blocks of text processed by language models. They can represent single characters, partial words, or full words, and include spaces and punctuation [1], [2].

Trustworthy AI

As defined by the NIST AI Risk Management Framework, AI that is responsive to multiple values, including being valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed [4].

Usage Limits

The "conversation budget" that restricts how many messages a user can send to an AI model over a specific time period, distinct from the length limits of a single conversation [5].

Quick answers

What is the difference between a token and a word?

A token is the raw unit of text processed by an AI, whereas a word is a unit of human language. In English, one token is roughly equal to 4 characters or 0.75 words. This means a 100-word paragraph will typically consume about 130 to 140 tokens [1], [2].

Why do AI models hallucinate?

AI models hallucinate primarily because standard training and evaluation metrics reward accuracy (getting a lucky guess right) over honesty about uncertainty. When models are penalized for leaving a blank answer but get partial or lucky points for guessing, they learn to generate confident but incorrect answers [3].

What is a context window?

A context window is an AI model's working memory for a single conversation. It dictates how much text (measured in tokens) the model can read and write at one time. If you exceed this limit, the model may forget earlier parts of the conversation or require automatic summarization to continue [5].

Continue the path

Now that you understand the core terminology of generative AI, you can explore how to apply these concepts in your daily workflow. Learn how to write your first prompt or discover how to fact-check an AI answer to protect yourself against model hallucinations.

• Read our guide: How to Write Your First Prompt
• Read our guide: How to Fact-Check an AI Answer