A shared language for responsible AI.

87 terms

No matching terms. Try an acronym, full name, or a broader topic.

AI agent#

Agents & integration

An AI-enabled application that can select or perform actions using tools. Its authority comes from application controls and identities, not from a role described in its prompt.

Explore the related guide ↗

AI governance#

Governance & compliance

The arrangements for making decisions, assigning accountability, and checking how AI is used throughout its lifecycle. Governance should connect policy to actual work and evidence.

Explore the related guide ↗

AI inventory#

Governance & compliance

A maintained record of AI uses, owners, purposes, data, providers, permissions, and review status. Record the use case rather than relying only on a product list.

Explore the related guide ↗

AI supply chain#

Security

The models, datasets, dependencies, providers, build systems, and services an AI application relies on. Review integrity, access, change management, and operating responsibility across that chain.

AI system#

Foundations

The application and operating arrangement that uses AI to perform work. It includes more than a model: data, interfaces, tools, people, and surrounding processes also shape its behavior.

Explore the related guide ↗

API (application programming interface)#

Agents & integration

A defined interface through which software requests data or operations from another component. AI applications use APIs for models, retrieval services, and tools; access still needs authentication and authorization.

Artificial intelligence (AI)#

Foundations

Computer-based systems that perform tasks such as prediction, classification, content generation, or decision support. The capabilities and risks depend on the particular system and its use.

Explore the related guide ↗

Assurance#

Assurance & operations

Evidence-based confidence in a claim within a defined scope. State what the evidence supports, how it was obtained, and which important questions it leaves unanswered.

Attention#

Models & training

A mechanism that weights relationships between parts of the input or intermediate representation. It helps a model use context; it should not be interpreted as human awareness or a complete explanation of a decision.

Audit trail#

Assurance & operations

Records that help explain who did what, when, under which authority, and with what outcome. Collect enough to investigate while protecting sensitive information in the records.

Authentication#

Security

Establishing the identity of a user, service, or other actor. Proving identity does not decide which actions that identity is allowed to perform.

Authorization#

Security

Deciding whether an identity may perform a particular operation on a resource under the relevant conditions. Enforce it at the component that controls access or execution.

Autonomy#

Agents & integration

The extent to which a system can choose and perform actions without another human decision. Describe the actual actions, limits, and review conditions instead of relying on a broad autonomy label.

Benchmark#

Assurance & operations

A standardized task or dataset used to compare performance. A strong benchmark result may not predict performance on your own users, data, permissions, or operating conditions.

Bias#

Assurance & operations

Systematic patterns that can lead to distorted or unequal outcomes. In an assessment, specify which pattern, population, task, and measure you mean rather than using the word as a complete diagnosis.

ChatGPT#

Generative AI

OpenAI’s conversational AI application. A product such as ChatGPT includes interfaces and features around its models; the application name and the name of an underlying model are different things.

Chunking#

Retrieval & data

Splitting documents into smaller pieces for indexing and retrieval. Chunk size and boundaries affect which context is found, how permissions are attached, and whether an answer has enough supporting information.

Computer vision#

Foundations

Methods for extracting or generating useful information from images and video, such as recognizing objects or classifying scenes. Performance should be tested in the conditions where the system will operate.

Context window#

Generative AI

The amount of context a model can handle in a request, usually measured in tokens. The practical budget may include instructions, conversation history, retrieved material, and generated output.

Control#

Governance & compliance

A measure intended to prevent, detect, limit, or respond to an unwanted outcome. A documented control needs implementation and verification before its effectiveness can be assumed.

Data exfiltration#

Security

Unauthorized transfer of information out of a protected environment. In AI systems, examine retrieved context, generated links, tool actions, logs, and outbound connections.

Data minimization#

Retrieval & data

Limiting collected and processed information to what a defined purpose needs. For an AI application, examine uploads, context, generated output, logs, and retained copies.

Data poisoning#

Security

Deliberate manipulation of data to influence training, evaluation, or another data-dependent workflow. Identify the affected stage and intended outcome when describing a particular case.

Data provenance#

Retrieval & data

Information about where data came from, how it was produced or changed, and who handled it. Provenance supports investigation and assessment without guaranteeing data quality.

Data retention#

Retrieval & data

How long information is kept and what happens when its retention period ends. An AI workflow may need to address source files, indexes, logs, backups, and cached outputs.

Explore the related guide ↗

Deep learning#

Foundations

Machine learning using neural networks with multiple layers. These models learn representations of data and support many modern language, image, audio, and prediction systems.

DPIA (data protection impact assessment)#

Governance & compliance

A structured assessment of personal-data processing and risks to individuals’ rights and freedoms. Whether one is legally required depends on the relevant jurisdiction and processing context.

Embedding#

Retrieval & data

A numerical representation that captures useful relationships in data, often used for similarity search. An embedding is not encryption and should not automatically be treated as non-sensitive.

Evaluation#

Assurance & operations

A defined process for assessing behavior against criteria relevant to a decision. Keep the test material, configuration, expected results, and measurement limitations with the findings.

Explore the related guide ↗

Explainability#

Assurance & operations

The ability to provide useful information about how or why a system produced an outcome. A model-generated explanation may sound convincing without faithfully describing its internal process.

Fairness#

Assurance & operations

A family of goals and measures concerning appropriate treatment of people or groups. Different definitions can conflict; select and justify the approach for the use case.

Fine-tuning#

Models & training

Additional training that adjusts model parameters for a narrower purpose or behavior. Supplying documents through retrieval is a different process and does not, by itself, fine-tune the model.

Foundation model#

Models & training

A model trained on broad data that can be adapted or used for a range of downstream tasks. Being broadly capable does not make it appropriate for every application.

Framework mapping#

Governance & compliance

A documented relationship between requirements or reference materials. The relationship can be partial; it does not automatically establish equivalence, effective controls, or legal compliance.

Explore the related guide ↗

Generative AI (GenAI)#

Foundations

AI that generates content such as text, code, images, audio, or video. A plausible or fluent result still needs checking against the purpose and evidence.

GPT (generative pre-trained transformer)#

Models & training

A name used for generative language models built around transformer architectures and pre-training. GPT is not a synonym for ChatGPT, which is an application built around models and additional features.

GRC (governance, risk, and compliance)#

Governance & compliance

Connected disciplines for directing organizational activity, managing uncertainty and potential harm, and meeting applicable obligations. Their AI application brings technical controls and business responsibility into the same discussion.

Grounding#

Retrieval & data

Connecting an answer to relevant information or evidence. Retrieval is one way to support grounding, but a retrieved source or citation does not guarantee that the answer represents it faithfully.

Guardrail#

Security

A control intended to constrain inputs, outputs, or actions. The word covers different mechanisms, so name what it checks, what it can block, and how its effectiveness is tested.

Hallucination#

Generative AI

An output that presents unsupported or incorrect information as if it were reliable. The term does not imply intent; check the output against sources and the task’s requirements.

Explore the related guide ↗

Human oversight#

Governance & compliance

An operating arrangement that enables people to review, change, stop, or escalate relevant system behavior. A nominal reviewer without time, evidence, or authority may provide little meaningful oversight.

Human-in-the-loop (HITL)#

Agents & integration

A workflow with human participation in a decision or action. Its effectiveness depends on what the person reviews, the information available, and their ability to change or stop the outcome.

Explore the related guide ↗

Impact assessment#

Governance & compliance

An examination of how a system may affect people, organizations, or other interests in context. Connect identified concerns to design decisions, ownership, and review.

Explore the related guide ↗

Incident response#

Assurance & operations

Preparation and coordinated action to detect, investigate, contain, recover from, and learn from an incident. AI workflows may require model, retrieval, memory, and tool context alongside conventional logs.

Explore the related guide ↗

Inference#

Generative AI

Using a trained model to produce an output for new input. Inference can generate an answer or prediction without changing the model’s learned parameters.

Jailbreak#

Security

An attempt to bypass a model or application’s intended behavioral restrictions. The term overlaps with some prompt-injection discussions, but not every attack on retrieved context has the same objective.

Large language model (LLM)#

Models & training

A large model trained to process and generate language. Many LLM applications use a transformer architecture; application behavior also depends on instructions, retrieved information, tools, and settings.

Least privilege#

Security

Giving an identity only the permissions needed for its task, for an appropriate period. Apply the idea to people, workloads, retrieval collections, and agent tools.

LLMOps#

Assurance & operations

Operating practices for applications built around large language models, including prompts, retrieval, tools, evaluation, cost, monitoring, and change management. The term does not name a single formal standard.

Machine learning (ML)#

Foundations

Methods that learn patterns from data to perform a task. Instead of specifying every rule by hand, developers train a model and evaluate how well it works on relevant examples.

Management system#

Governance & compliance

The responsibilities, policies, processes, and review arrangements an organization uses to achieve and improve stated objectives. An AI management system concerns organizational practice as well as individual technical controls.

MCP (Model Context Protocol)#

Agents & integration

A protocol for connecting AI applications with tools, resources, and contextual information through client-server interfaces. Using MCP does not by itself establish trust in a server or authorize an action.

Membership inference#

Security

An attempt to determine whether a particular record was part of a model’s training data. Its success depends on the model and the information available to the attacker.

MLOps#

Assurance & operations

Engineering practices for developing, deploying, monitoring, and maintaining machine-learning systems. They connect models and data with versioning, testing, access, and operational ownership.

Model drift#

Assurance & operations

A change in a model’s observed behavior or performance over time. Investigate changes in data, user behavior, system configuration, and the model rather than assuming one cause.

Model extraction#

Security

An attempt to obtain model information or reproduce its behavior through access such as repeated queries. The achievable result depends on the model, interface, exposure, and attacker resources.

Multimodal AI#

Generative AI

AI that works with more than one type of information, such as text, images, audio, or video. Each input and output path can introduce additional data-handling and security questions.

Natural language processing (NLP)#

Foundations

Methods for processing and generating human language. Examples include translation, text classification, speech-related workflows, and question answering.

Neural network#

Foundations

A model made of connected computational units with adjustable weights. Training changes those weights to improve performance on an objective; the name does not mean the model works like a human brain.

Parameter#

Models & training

An adjustable value within a model, such as a neural-network weight. Training learns parameter values. Parameter count alone does not establish answer quality, safety, or suitability.

Pre-training#

Models & training

The broad initial training stage that teaches a model useful representations or patterns before later adaptation. The training objective and data vary by model.

Predictive AI#

Foundations

AI used to estimate an outcome or assign a category, such as forecasting demand or flagging a transaction. Its output can support a decision without generating new prose or images.

Prompt#

Generative AI

Input supplied to a model to guide a task. A complete request may include several kinds of instructions, user content, examples, and retrieved information.

Prompt engineering#

Generative AI

Designing and testing instructions, examples, and context to improve task performance. It is an application technique, not a substitute for authorization, evaluation, or secure software design.

Prompt injection#

Security

Untrusted input that attempts to redirect an AI application’s intended behavior. It can arrive in a user message or indirectly through documents, webpages, emails, and tool results.

Explore the related guide ↗

RAG (retrieval-augmented generation)#

Retrieval & data

A workflow that retrieves external information and provides it to a generative model as context. It can improve access to relevant knowledge, but it does not eliminate incorrect answers or replace permission checks.

Explore the related guide ↗

Red teaming#

Assurance & operations

A scoped adversarial exercise intended to reveal weaknesses and improve a system. Authorization, safe targets, evidence capture, and retesting are part of the work.

Explore the related guide ↗

Reinforcement learning#

Models & training

Learning in which actions or outputs are adjusted using rewards or feedback about outcomes. The reward design influences what behavior the training encourages.

Reranking#

Retrieval & data

Reordering retrieved candidates using an additional relevance assessment. It can improve source selection, but a relevance score does not establish permission, factual truth, or source trustworthiness.

Residual risk#

Governance & compliance

The exposure remaining after considering controls and their effectiveness. Record the reasoning, uncertainty, and accountable decision about whether that exposure is acceptable.

Risk register#

Governance & compliance

A working record linking risk scenarios to owners, controls, decisions, evidence, and review dates. It should distinguish existing safeguards from measures that are only planned.

Explore the related guide ↗

Risk scenario#

Governance & compliance

A description of a possible event, the conditions that could lead to it, and the consequences that matter. A useful scenario is specific enough to assign an owner and choose a response.

RLHF (reinforcement learning from human feedback)#

Models & training

An approach that uses human feedback or preferences as part of a reinforcement-learning process to shape model behavior. It can improve behavior without guaranteeing factual accuracy or safety.

Self-supervised learning#

Models & training

Training in which the task’s targets are derived from the data itself, such as predicting a missing or next token. It can use large datasets without people labeling every example.

Semantic search#

Retrieval & data

Searching by representations of meaning or related concepts rather than relying only on exact words. A relevant result may still be outdated, unauthorized, or unreliable.

Shadow AI#

Governance & compliance

AI use outside an organization’s intended visibility, approval, or operating processes. It may include embedded product features as well as separate chat applications.

Supervised learning#

Models & training

Training with examples paired with target answers or labels. The resulting model is evaluated on relevant data beyond the examples it learned from.

System instruction#

Generative AI

A higher-priority instruction set by an application or platform to guide model behavior. Instructions can help shape a workflow, while permissions must still be enforced by the surrounding application.

Temperature#

Generative AI

A sampling setting that affects the probability distribution used when selecting generated tokens. Lower values often reduce variation; a setting of zero does not universally guarantee identical outputs.

Threat modeling#

Security

A structured examination of assets, actors, possible attack paths, and defenses in a specific system. Use the architecture and operating context to choose plausible scenarios.

Token#

Generative AI

A unit a model uses to represent input or output. It may be part of a word, a whole word, punctuation, or another encoded element; one token is not always one word.

Tool calling#

Agents & integration

A model proposing or requesting an operation through a defined tool interface. The application should validate parameters, authorize the operation, and report the actual execution result.

Training data#

Models & training

Examples used to learn or adjust a model’s parameters. Their quality, provenance, representativeness, and permitted use can affect both performance and governance obligations.

Transformer#

Models & training

A neural-network architecture that uses attention mechanisms to process relationships in data. It underlies many language and multimodal models, including both generative and non-generative systems.

Trust boundary#

Security

A point where information or actions cross between different levels of trust or authority. Document what is validated and enforced at that crossing.

Vector database#

Retrieval & data

A database designed to store and search vector representations, often by similarity. Its access controls, data lifecycle, and tenant isolation matter just as they do for other data stores.

Workload identity#

Agents & integration

An identity used by an application, service, or automated job. Keep its permissions tied to its actual purpose and distinguish it from the human user requesting work.