Glossary · Category

Privacy & governance

35 plain-English definitions.

AI bill of materials (AIBOM): A structured inventory documenting all the models, datasets, and components that make up an AI system, for transparency and risk assessment.
AI risk management: The organizational process of identifying, assessing, and mitigating risks introduced by deploying AI systems.
AI usage policy: An internal organizational document defining approved AI tools, acceptable use cases, and data handling rules for employees.
air-gapped deployment: Running an AI system on infrastructure completely isolated from external networks, for maximum security in sensitive environments.
audit trail (AI): A recorded history of an AI system's requests, decisions, and outputs used for compliance review or incident investigation.
confidential computing: Hardware-based technology that encrypts data even while it's being processed in memory, protecting sensitive AI workloads from the infrastructure provider itself.
data processing agreement (DPA): A legal contract governing how a vendor processes and protects a customer's data, standard for enterprise AI procurement.
data residency: Requirements or guarantees about the geographic location where data is stored and processed, often for regulatory compliance.
data sovereignty: The principle that data is subject to the laws of the country in which it is collected or stored.
EU AI Act: European Union legislation that classifies and regulates AI systems by risk level, imposing compliance obligations on providers and deployers.
export controls (AI models): Government restrictions on which countries or entities advanced AI models and chips can be sold or transferred to.
GDPR compliance (AI): Meeting EU data protection regulations when processing personal data of EU residents through AI systems.
HIPAA compliance (AI): Meeting U.S. healthcare data privacy regulations when processing protected health information through an LLM pipeline.
ISO 42001: An international standard specifying requirements for an organization's AI management system, akin to ISO 27001 for security.
kill switch (AI system): A mechanism to immediately halt or disable an AI system's actions if it behaves unexpectedly or unsafely.
model access control: Permissions systems that govern which users, teams, or applications are allowed to call which models.
model governance: The policies and processes an organization uses to approve, monitor, and control which AI models are used and how.
model portability: The ability to switch between LLM providers or self-hosted models with minimal application code changes, often via a unified API.
model transparency: The degree to which a provider discloses information about a model's training data, architecture, and evaluation.
NIST AI Risk Management Framework: A U.S. government-published framework offering guidance for identifying and managing risks throughout the AI system lifecycle.
on-premises inference: Running LLM inference entirely within an organization's own physical data centers rather than any cloud provider.
PII redaction: Automatically detecting and removing or masking personally identifiable information from text before it's sent to or returned from a model.
private cloud deployment: Running LLM inference within an organization's own cloud environment (VPC) rather than a shared multi-tenant public API.
prompt logging: Recording the prompts and responses sent through an LLM application, raising both debugging value and privacy considerations.
responsible AI: An umbrella term for organizational practices aimed at ensuring AI systems are safe, fair, transparent, and accountable.
role-based access control (RBAC) for AI: Restricting who can access which models, data, or features in an AI platform based on assigned organizational roles.
self-hosting (LLM): Running open-weight models on infrastructure you control instead of calling a third-party hosted API.
shadow AI: Unauthorized or unmanaged use of AI tools within an organization outside of sanctioned, governed channels.
single sign-on (SSO) for AI tools: Letting users authenticate into an AI platform using their existing enterprise identity provider.
SOC 2 compliance: A security and privacy audit standard commonly required of vendors handling customer data, including AI infrastructure providers.
training data opt-out: A setting or policy ensuring a customer's API data is never used to train or improve a provider's models.
trusted execution environment (TEE): An isolated, encrypted region of a processor used to run sensitive computations (including AI inference) securely away from the rest of the system.
vendor lock-in (AI): The risk of becoming dependent on a single LLM provider's API, pricing, and capabilities, making it costly to switch later.
VPC deployment (AI): Deploying model inference inside a customer's virtual private cloud for network isolation and data control.
zero data retention: A contractual or technical guarantee that a provider does not store request/response data after processing it.