Definitions aligned to FDA AI credibility guidance, EU AI Act, GMP Annex 11/22, and GxP data integrity — from agentic AI and ALCOA+ to validation and the jagged frontier.
- Agent / agentic AI
- AI that doesn't just answer questions but plans and executes multi-step tasks using tools: querying systems, drafting and filing documents, triggering workflows.
- AI literacy
- Defined in EU AI Act Article 3(56) as the "skills, knowledge and understanding" that allow providers, deployers, and affected persons "to make an informed deployment of AI systems, as well as to gain awareness about the opportunities and risks of AI and possible harm it can cause.
- AI system (EU AI Act definition)
- "A machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions" with some level of autonomy and possible post-deployment adaptiveness.
- ALCOA+
- The data-integrity mnemonic: records must be Attributable, Legible, Contemporaneous, Original, Accurate, plus Complete, Consistent, Enduring, Available.
- AlphaFold
- DeepMind's protein-structure prediction system, the subject of the 2024 Nobel Prize in Chemistry.
- Annex 11 (EU GMP, draft revision)
- The EU's rules for computerised systems in GMP, under revision in the same 2025 consultation as Annex 22.
- Annex 22 (EU GMP, draft)
- The first GMP-specific AI guideline, drafted by the EMA/PIC/S inspectors' working group (consultation July–October 2025; still draft).
- Article 4 (EU AI Act)
- The AI-literacy obligation: providers and deployers "shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf," proportionate to technical knowledge, experience, education, and context of use.
- Bias (model bias)
- Systematic error a model inherits from unrepresentative training data or flawed labels: an inspection model trained mostly on one vial format underperforming on another, a triage model that under-flags rare event types it seldom saw.
- Black box
- A model whose internal decision logic can't be directly inspected, true of most deep-learning systems, whose "reasoning" is distributed across millions of parameters.
- Centaur / Cyborg
- The two successful human-AI working styles identified in the Harvard/BCG "jagged frontier" experiment.
- Confidence score
- The model's own numeric estimate of how certain it is about a given output.
- Context of use (COU)
- FDA's pivotal term: "the specific role and scope of the AI model used to address a question of interest": what the model does, how its output will be used, and what other evidence sits alongside it.
- Context window
- The bounded amount of text an LLM can consider in one exchange.
- Credibility assessment (FDA)
- The seven-step process in FDA's January 2025 draft guidance for trusting an AI model in regulatory decision-making: (1) define the question of interest; (2) define the COU; (3) assess model risk (influence × consequence); (4) plan the credibility activities; (5) execute; (6) document results and deviations; (7) determine adequacy for the COU.
- CSA (Computer Software Assurance)
- FDA's modern philosophy of software validation: draft 2022, finalized September 24, 2025 (CDRH/CBER; updated February 2026), formally displacing the documentation-heavy CSV tradition for production and quality-system software.
- CSV (Computer System Validation)
- The traditional GxP discipline of documented evidence that a computerized system does what it's intended to do, historically implemented through IQ/OQ/PQ protocols and extensive scripted testing.
- Data drift / model drift
- The silent degradation mode of deployed AI: the real world shifts away from the training data (new product format, new camera, new supplier's raw material, new documentation style), and performance decays without any code change or error message.
- Data governance
- The organizational discipline covering how data is sourced, stored, accessed, retained, and used: consent, lineage, quality, security.
- Dataset (training / validation / test)
- The three-way split at the heart of honest ML evaluation: the *training set* teaches the model; the *validation set* (in the ML sense, tuning, not GxP validation) guides development choices; the *test set* is held back and touched only to measure final performance.
- De novo design (generative chemistry)
- Using generative models to propose entirely new molecules or proteins with desired properties, rather than screening libraries of existing compounds.
- Decision consequence (FDA)
- One of the two axes of FDA model risk: "the significance of an adverse outcome resulting from an incorrect decision.
- Deep learning
- Machine learning built on many-layered neural networks, responsible for the modern leaps in vision (Amgen's vial inspection) and language (every LLM).
- Deterministic
- Property of producing identical output for identical input, every time.
- Digital twin
- A continuously updated computational model of a physical asset or process, a bioreactor, a production line, even a patient population, used to simulate, predict, and optimize before touching the real thing.
- Explainability (XAI)
- Techniques that expose *why* a model produced an output: SHAP and LIME attribute a prediction to input features; heat maps show which image regions drove a vision model's call.
- "Falling asleep at the wheel"
- The over-reliance failure mode named in the jagged-frontier research: as AI output quality rises, human vigilance falls, and the errors that remain are precisely the ones no one is checking for.
- Federated learning
- Training a shared model across multiple organizations or sites without moving the underlying data: the model travels, the data stays put.
- Fine-tuning
- Additional training of a general-purpose model on domain-specific data (your deviation histories, your SOP corpus) to specialize its behavior.
- Foundation model / GPAI
- A very large model trained on broad data that many downstream applications build on.
- 4-eyes principle (Annex 22)
- The draft's control for test-data independence: a person *without access to the training data* independently reviews test outcomes, so no single individual can (accidentally or otherwise) let training data contaminate the test.
- GAMP 5 (2nd ed.) / GAMP AI Guide
- ISPE's practitioner frameworks for computerized-system quality.
- Generative AI
- AI that produces new content, text, images, code, structures, rather than only classifying or predicting.
- Grounding
- Constraining a model's outputs to verifiable sources, documents retrieved at question time, structured data, citations a human can follow, rather than the model's parametric memory.
- Guardrails
- The layered constraints on AI behavior: system-level (what the deployment will refuse), application-level (input/output filters, allowed data sources), and prompt-level (your own instructions, "flag missing information rather than guessing").
- Hallucination
- The generation of confident, fluent, well-formatted content that is factually wrong or fabricated, invented citations, plausible-but-false numbers, nonexistent guidance documents.
- High-risk AI (EU AI Act)
- The Act's most regulated tier, defined by two routes: safety components of products under listed EU legislation (medical devices via MDR/IVDR, note medicinal-products law is *not* on that list), or uses listed in Annex III (biometrics, critical infrastructure, employment, essential services…).
- Human-in-the-loop (HITL)
- An oversight design where a human reviews or approves AI output before it takes effect.
- In silico
- Performed by computer simulation, the third member of the classic triad with in vitro (glassware) and in vivo (living organisms).
- Inference
- Running a trained model on new data to produce an output, the "use" phase, as distinct from training.
- Input space
- The range and character of data a model was designed and validated for, vial formats, lighting conditions, document types, populations.
- Jagged frontier
- The defining metaphor for AI capability, from the 758-consultant Harvard/BCG field experiment: the boundary of what AI does well is uneven and invisible, +12% task completion, +25% speed, +40% quality inside it; −19 percentage points accuracy just outside it.
- Knowledge graph
- A data structure representing entities (drugs, targets, adverse events, batches, suppliers) and the relationships between them, so machines can traverse connections the way an expert's intuition does.
- Large language model (LLM)
- A deep-learning model trained on vast text corpora to predict likely next tokens, from which capabilities in drafting, summarization, reasoning-like behavior, and code emerge.
- Machine learning (ML)
- Building models whose behavior is learned from data rather than explicitly programmed.
- Model
- The trained artifact itself: the architecture plus the learned parameters that map inputs to outputs.
- Model influence (FDA)
- The second axis of FDA model risk: "the contribution of the evidence derived from the AI model relative to other contributing evidence.
- Model lifecycle
- The full arc: data acquisition → training → evaluation → deployment → monitoring → retraining → retirement, with governance at each stage.
- Multimodal AI
- Models that work across data types at once: text, images, structured data, molecular structures, genomic sequences.
- NIST AI Risk Management Framework (AI RMF)
- The voluntary US framework (v1.
- Overfitting
- When a model memorizes its training data instead of learning generalizable patterns, spectacular on the training set, brittle in production.
- Parameters
- The numerical weights inside a model, millions to trillions of them, set during training and encoding everything it "knows.
- Part 11 (21 CFR Part 11)
- FDA's 1997 regulation on electronic records and electronic signatures: trustworthy records (validation, audit trails, protection, access control) and signatures that legally bind a person to content.
- PCCP (Predetermined Change Control Plan)
- FDA's mechanism (final guidance December 2024, device-side) letting a manufacturer pre-specify how an AI-enabled product will be modified, retraining triggers, performance bounds, update protocols, so in-bounds changes proceed without new marketing submissions.
- Predictive maintenance
- ML on equipment sensor data to forecast failures before they occur: vibration, temperature, cycle patterns.
- Probabilistic
- Producing outputs as likelihoods over possibilities rather than a single guaranteed answer, ML's native mode.
- Prompt / prompt engineering
- The instruction given to a generative model, and the craft of writing it well.
- Question of interest (FDA)
- Step one of the credibility framework: the specific question the AI model's evidence will help answer; e.
- RAG (Retrieval-Augmented Generation)
- Architecture in which the system first retrieves relevant passages from a controlled corpus (your SOPs, your regulations, your records) and requires the LLM to answer from them, typically with citations.
- Real-world data / real-world evidence (RWD/RWE)
- Health data gathered outside conventional trials (claims, EHRs, registries, wearables) and the evidence derived from it.
- Self-driving lab
- Closed-loop experimentation in which AI proposes the next experiment, robotic automation runs it, and the results retrain the model, iterating around the clock.
- Shadow AI
- Employees using personal, unsanctioned AI tools for work: pasting company content into consumer chatbots.
- SHAP / LIME
- The two feature-attribution techniques named in draft Annex 22.
- Silicon ceiling
- BCG's term (AI at Work 2025, n=10,600+) for the adoption gap between leadership and the front line: >75% of leaders and managers use genAI regularly versus 51% of frontline employees, with training access, not attitude, as the main differentiator.
- Static model / locked model
- A model whose parameters are frozen after training and validation; it changes only through controlled retraining and re-release.
- Synthetic control arm
- A comparison group for a clinical trial constructed from historical trial data and real-world data instead of newly randomized placebo patients.
- 10-20-70
- BCG's empirical ratio for where AI-transformation effort must go: 10% algorithms, 20% technology and data infrastructure, 70% people, processes, and culture.
- Test set (independent)
- Data held back from all model development and used once, to measure honest performance.
- Training / training data
- Training is the optimization process that sets a model's parameters; training data is the example corpus it learns from.
- Transparency
- Operating openly about when AI is used, what it does, and its limits, toward employees, regulators, and affected persons.
- Validation (of AI)
- The umbrella activity: documented, risk-proportionate evidence that a model reliably does its job in its context of use.