| Model |
The computational engine that reads text and generates what follows. A product may use several models. |
| Token |
A basic piece used by a model to process text. Tokens do not correspond one to one with characters or words, and vary by language. |
| Reasoning token |
A token used internally by a reasoning model before answering. Invisible, but counted and billed as output. |
| Context window |
The total instructions, conversation, material and answer a model refers to at once. Its working memory. |
| Context rot |
Declining accuracy and recall as context grows longer. |
| Output limit |
The maximum number of tokens that can be generated in one response. Separate from the context window. |
| Probabilistic generation |
Generating text by choosing successive tokens according to probabilities. It has no fact-checking step. |
| Temperature |
An API setting controlling randomness. Lowering it does not guarantee accuracy, and even 0 may not produce identical results. |
| Knowledge cutoff |
The reference point for a model's training material. Later facts may be unknown or inaccurate and need outside evidence. |
| Reliable knowledge cutoff |
The point up to which knowledge is broadest and most reliable. The training data cutoff covers a wider range of material. |
| Grounding |
Anchoring answers in material retrieved through search or document lookup. Citations begin verification. |
| Hallucination |
Plausibly generated content that differs from facts or provided material. |
| Sycophancy |
Favoring answers that match the user's beliefs over truthful answers, such as changing an answer without evidence after an objection. |
| Reasoning model |
A model that internally works through several steps before answering. Strong at complex judgment, but slower and costlier. |
| Model tier |
A level within a model family differing in size, speed and cost. |
| Multimodal |
Ability to handle images, PDFs, audio, video and other inputs beyond text. |
| Completion criteria |
Checkable conditions defining when a result is finished. |
| Zero-shot / few-shot |
Instructions without examples / instructions accompanied by a few examples. |
| Prompt chaining |
Dividing work by passing one step's output into the next as input. |
| Self-correction |
A three-step chain: drafting, review against criteria, revision. |
| Temporary chat |
A conversation excluded from history and training. Short-term safety retention may still apply. |
| Custom instructions |
User-defined response guidance saved to an account. Scope depends on the product and conversation mode. |
| Memory |
A feature in which a product selects relevant information from conversations and other sources, saves it and uses it in later answers. |
| Project (knowledge space) |
A workspace collecting related conversations, files and instructions. |
| Project instructions |
Instructions applying only to conversations within that project. |
| Custom assistant |
A package of instructions, reference material and tools for consistent use by multiple people. |
| Connector |
A feature linking AI to outside services for retrieval and actions. Follows the connected user's permissions. |
| MCP |
Model Context Protocol, a public standard connecting AI applications to external data and tools. |
| Agent features |
A form in which AI plans multiple steps and takes actions across websites and apps. |
| Deep research |
A feature that plans research, reads multiple sources and produces a cited report. |
| Scheduled task |
Work executed repeatedly at specified times or in response to events. |
| API |
A route for software to call a model directly. Billed separately from chat subscriptions. |
| Batch processing |
Grouping high-volume requests without immediate-response needs for lower-cost processing. |
| Retrieval-augmented generation (RAG) |
Retrieving relevant material and supplying it as the basis for answers. |
| Fine-tuning |
Further training a model with additional material. Considered when behavioral consistency is the problem. |
| Usage policy |
Rules defining what an AI company's service must not be used for. |
| Human review |
A process in which a qualified person meaningfully examines AI output before deciding; human-in-the-loop. |
| High-impact AI |
Under Article 2(4) of Korea's AI Basic Act, an AI system that may significantly affect or pose risks to human life, physical safety or fundamental rights and is used in listed domains, such as hiring and loan screening. |
| Consumer / business plan |
A plan individuals obtain themselves / one an organization contracts for its members. Data-handling conditions differ. |
| Administrator controls |
Features through which organizational administrators configure and supervise access, retention, features, connectors and auditing. |
| Data minimization |
Excluding information unnecessary for a task. |
| Personal information |
Information identifying a living individual, including information easily combined with other information for identification and pseudonymized information. |
| Pseudonymization |
Removing or replacing portions so an individual cannot be identified without additional information. |
| Sensitive information |
Personal information whose processing is restricted in principle, including beliefs, political views and health (PIPA, Article 23). |
| Unique identifying information |
Identifiers assigned by law to distinguish individuals. Processing is restricted in principle (Article 24; resident registration numbers are covered by Article 24-2). |
| Prompt injection |
A vulnerability in which input text causes unintended changes in model behavior. Includes indirect injection through external material. |
| Bias |
Social imbalances in training material reflected in output. |
| Baseline |
A pre-adoption measurement used for comparison. |
| Total cost |
Actual cost combining fees, preparation, training, operations, review and rework. |
| Pilot |
A small-scale trial to assess impact before broader operation. |
| Vendor lock-in |
Deep dependence on a provider's product that makes switching difficult. |
| AI RMF |
NIST's AI Risk Management Framework, consisting of GOVERN, MAP, MEASURE and MANAGE. |
| AI fluency (4D) |
A framework describing AI use through Delegation, Description, Discernment and Diligence. |