These are the terms I use most when I'm validating systems and AI in pharma, written the way I'd explain them to a new member of a validation team. Where a term comes from a regulation or an industry guide, I've named it, so you can go to the source. Draft documents are marked as drafts, because their wording may still change. Each term has its own link, such as /glossary#computer-software-assurance, so you can share one definition.
Numbers
21 CFR Part 11
The US FDA regulation on electronic records and electronic signatures. It sets the conditions under which FDA accepts electronic records and signatures as equal to paper ones, covering system validation, audit trails, access control and signature controls. An AI system that creates or changes GxP records falls under it like any other computer system. More in computer system validation in pharma.
A
Acceptance criteria
The performance levels a system must meet to pass testing, written and approved before testing starts. For an AI model they're usually statistical, such as a minimum sensitivity or a maximum false-reject rate on the test set. The draft EU GMP Annex 22 asks for them to be defined by the process subject matter expert, and to be at least as high as the performance of the process the model replaces. More in AI validation in GxP.
AI agent
An AI system that plans steps and takes actions to reach a goal, such as reading files, calling other software or writing records. Agents add a second control question to "is the output right?", which is "what is it allowed to do?" In GxP work, an agent's permissions should be limited to the task, with a human approval step before anything reaches a GxP record. More in AI validation in GxP.
ALCOA+
The data integrity principles that GxP records should meet: attributable, legible, contemporaneous, original and accurate, plus complete, consistent, enduring and available. FDA's 2018 data integrity Q&A guidance uses ALCOA, and PIC/S PI 041-1 adds the 4 attributes after the plus. For AI they apply to the training and test data as well as to the records the system produces. More in computer system validation in pharma.
Annex 11
EU GMP Annex 11: Computerised Systems, part of EudraLex Volume 4. The version in force dates from 2011. A substantially revised draft went to public consultation from 7 July to 7 October 2025, and EMA's inspectors working group is targeting a final text for the European Commission in Q4 2026. More in computer system validation in pharma.
Annex 22
The draft EU GMP Annex 22, Artificial Intelligence, published for consultation with the Annex 11 and Chapter 4 drafts in July 2025. It covers AI models used in critical GMP applications and, in its current draft, limits those to static models with deterministic output. On 12 November 2026 it's still a draft. More in the Annex 22 tracker.
Audit trail
A secure, computer-generated, time-stamped record of who did what to a GxP record, and when, including the old and new values and the reason for a change where required. 21 CFR Part 11 and EU GMP Annex 11 both require it. For AI, I also want an audit trail on the test dataset, so I can prove the people who trained the model never had access to it. More in computer system validation in pharma.
C
Change control
The formal process for proposing, assessing, approving, implementing and verifying a change to a validated system. It's required under EU GMP and 21 CFR Parts 210 and 211, and described in GAMP 5. For AI, retraining a model, swapping a dataset or changing a prompt are all changes and go through it. More in computer system validation in pharma.
Computer software assurance (CSA)
FDA's risk-based approach to assuring that production and quality system software is fit for its intended use. It puts testing effort where the risk to product quality and patient safety is highest, and accepts unscripted testing and lighter records for lower-risk functions. FDA's CSA guidance was finalised September 2025, updated February 2026 to align with the QMSR. More in CSV vs CSA.
Computer system validation (CSV)
The documented process of proving that a computer system does what it's intended to do, consistently and reliably, across its life. It's required by EU GMP Annex 11, 21 CFR 211.68 and 21 CFR Part 11, and GAMP 5 is the main industry guide to doing it. CSA is a risk-based way of doing CSV, and doesn't replace the requirement. More in computer system validation in pharma. For help with one system, see my computer system validation consulting.
Confidence score
A number an AI model gives alongside its output that shows how sure it is of the result. The draft EU GMP Annex 22 asks for confidence to be recorded and for a threshold below which the output is flagged as undecided instead of being acted on. Setting that threshold, and deciding what happens to the cases under it, is part of the validation. More in the Annex 22 tracker.
Context of use
FDA's term for how and where an AI model is used to answer a specific question, including the model's role and scope. It's central to FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs and biologics. It plays the same part as intended use in GMP validation. More in AI validation in GxP.
Critical thinking
The habit, at the centre of GAMP 5 Second Edition and behind FDA's CSA guidance, of deciding what to test and document based on real risk to the product and the patient. It means asking what could go wrong, how likely it is, and what evidence would show it's controlled. In AI validation it's what stops a team from testing the easy things thoroughly and the risky things not at all. More in CSV vs CSA.
D
Data drift
A change over time in the data a model receives in use, compared with the data it was tested on. A new raw material supplier, a new camera or a new site can all cause it. The draft EU GMP Annex 22 asks firms to monitor the input sample space so they can detect this before performance suffers. More in AI validation in GxP.
Data integrity
The extent to which data is complete, consistent and accurate across its whole life, from creation to archive. Regulators set out what they expect in guidance such as PIC/S PI 041-1 and FDA's 2018 data integrity Q&A, and EU GMP Chapter 4 and Annex 11 build on it. For AI, data integrity covers the training and test data as well as the outputs. More in computer system validation in pharma.
Decision integrity
A term I use in my book for the assurance that a decision made with AI support can be reconstructed afterwards and shown to be sound: which model version, which input, what confidence, which person accepted it and why. It isn't a regulatory term. I use it because inspectors increasingly ask to walk through a single AI-supported decision end to end. More in AI validation in GxP.
Deterministic and probabilistic output
A deterministic system gives the same output every time for the same input. A probabilistic system can give different outputs for the same input, or gives a likelihood instead of a fixed answer. The draft EU GMP Annex 22 covers only models with deterministic output in critical applications. More in the Annex 22 tracker.
E
EU AI Act high-risk system
An AI system classed as high-risk under Regulation (EU) 2024/1689, either because it's a safety component of a product under the laws listed in Annex I (medical devices among them) or because its use is listed in Annex III. High-risk systems carry obligations on risk management, data governance, documentation and human oversight. The Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the dates to 2 December 2027 for Annex III and 2 August 2028 for Annex I. More in AI validation in GxP.
Explainability
The ability to show which inputs or features drove a model's output. The draft EU GMP Annex 22 asks for feature attribution to be captured, and for a review of those features to be part of approving the test results. A model that rejects a tablet because of the background lighting instead of the defect is the classic failure explainability catches. More in the Annex 22 tracker.
G
GAMP 5
ISPE's GAMP 5 guide, the main industry guide to computer system validation. The Second Edition, published in July 2022, added an appendix on artificial intelligence and machine learning. ISPE followed it with the separate GAMP Guide: Artificial Intelligence in July 2025. More in computer system validation in pharma.
Generative AI
AI that produces new content, such as text, images or code, in response to a prompt. Large language models are the most common type in pharma. The draft EU GMP Annex 22 says it doesn't apply to generative AI and that such models should not be used in critical GMP applications. In non-critical uses, qualified and trained people must be responsible for checking the output. More in the Annex 22 tracker.
GxP
The family of "good practice" quality regulations and guidelines in life sciences, such as good manufacturing practice (GMP), good laboratory practice (GLP), good clinical practice (GCP) and good distribution practice (GDP). A system is GxP-relevant when it can affect product quality, patient safety or the integrity of GxP data. That's what brings it into scope for validation. More in computer system validation in pharma.
H
Hallucination
When a generative AI model states something false with apparent confidence, such as an invented regulation, a wrong limit or a citation that doesn't exist. It's a property of how these models work, and it can't be fully tested out. The control is a qualified person checking every output that reaches a GxP record against the real source. More in AI validation in GxP.
Human in the loop
A design where a person reviews, and can change or reject, an AI output before it's acted on. The draft EU GMP Annex 22 asks for the operator's responsibility to be part of the intended use, and for the operator's training and performance to be monitored like any other manual process. The review has to be real: records of edits and overrides show whether it is. More in the Annex 22 tracker.
I
Intended use
A precise description of what a system is for, who uses it, what goes in, what comes out and what decision the output supports. It's the foundation of validation under GAMP 5, Annex 11 and FDA's CSA guidance. The draft EU GMP Annex 22 asks for it to describe the input sample space, divided into subgroups, and to be approved by the process SME. More in computer system validation in pharma.
IQ, OQ and PQ
Installation, operational and performance qualification: the classic 3 stages of proving a system or piece of equipment is installed correctly, works as specified and performs in real use. They come from equipment qualification practice and EU GMP Annex 15, and GAMP 5 uses broader verification terms for software. For AI, PQ is where performance on representative production data gets proven. More in computer system validation in pharma.
L
Large language model (LLM)
A generative AI model trained on very large amounts of text to predict and produce language. Examples include the models behind ChatGPT, Claude and Gemini. In the current draft EU GMP Annex 22 they're kept out of critical GMP applications, though EMA's 2026 workshop report shows the drafting group discussed widening the scope. More in the Annex 22 tracker.
M
Model card
A short document that records a model's purpose, training data, performance, limits and version in one place. It started as a practice in AI research and is now widely used in industry; GxP regulations don't require it by name. I find it the easiest way to give an inspector a one-page view of the model. More in AI validation in GxP.
Model drift
A fall in a model's performance over time in real use, often caused by data drift or by changes in the process it supports. It can happen with no change to the model itself. Production monitoring against the acceptance criteria is how you catch it, and the draft EU GMP Annex 22 asks for this monitoring. More in AI validation in GxP.
Model risk
In FDA's January 2025 draft guidance on AI for regulatory decision-making, the combination of model influence (how much the AI output contributes to the decision) and decision consequence (how bad a wrong decision would be). Higher model risk calls for more credibility evidence. It's a useful way to set validation depth in GMP too. More in AI validation in GxP.
P
Periodic review
A scheduled review confirming that a validated system is still in a validated state and still fit for its intended use. It's required by EU GMP Annex 11 and described in GAMP 5. For AI it also covers performance trends, drift alerts, overrides and whether people use the system the way the intended use says. More in computer system validation in pharma.
PIC/S
The Pharmaceutical Inspection Co-operation Scheme, a co-operation between more than 50 GMP inspectorates. It publishes a GMP guide that mirrors EU GMP closely, and it worked with the EMA inspectors working group on the 2025 drafts of Annex 11, Annex 22 and Chapter 4. When PIC/S adopts a text, inspectorates well beyond the EU apply it. More in the Annex 22 tracker.
Predetermined change control plan (PCCP)
An FDA mechanism for medical devices that lets a manufacturer describe, in its marketing submission, the changes it plans to make to an AI-enabled device and how it'll validate them. Once authorised, those changes can be made without a new submission. FDA finalised its guidance on PCCPs for AI-enabled device software functions in December 2024. The idea of pre-agreeing how a model may change is useful for GMP change control too. More in AI validation in GxP.
Prompt
The instruction or input given to a generative AI model. In a GxP system the fixed parts of the prompt (the system prompt, templates and retrieval sources) are configuration and belong under change control. Prompts and responses that feed a GxP record should be logged. More in AI validation in GxP.
Q
QMSR
FDA's Quality Management System Regulation, the amended 21 CFR Part 820 for medical devices, which incorporates ISO 13485:2016 by reference. It took effect on 2 February 2026, replacing the old Quality System Regulation. FDA's CSA guidance was finalised September 2025, updated February 2026 to align with the QMSR. More in CSV vs CSA.
Quality unit
The function responsible for approving or rejecting procedures, specifications and batches, defined in 21 CFR 211.22. It can't hand that responsibility to software. FDA's warning letter 320-26-58 of 2 April 2026 made the point directly: any output or recommendation from an AI agent must be reviewed and cleared by an authorised person in the firm's quality unit. More in AI validation in GxP.
R
Retrieval-augmented generation (RAG)
A design where a generative AI model looks up passages from a defined set of documents and answers from them, instead of relying only on what it learned in training. It cuts hallucination and lets the answer cite its source. In GxP use, the document set is controlled content, and changing it is a change. More in AI validation in GxP.
Risk-based approach
Scaling validation and control effort to the risk a system poses to patient safety, product quality and data integrity. It's set out in ICH Q9(R1) Quality Risk Management and runs through GAMP 5, Annex 11 and FDA's CSA guidance. For AI, risk depends on what the output touches and how much a person stands between it and the decision. More in how to move from CSV to CSA.
S
Static and dynamic models
A static model is locked after training and doesn't change in use. A dynamic model keeps learning from new data while in use, so its behaviour can change between one day and the next. The draft EU GMP Annex 22 covers only static models in critical GMP applications. More in the Annex 22 tracker.
Supplier assessment
The evaluation of a software or service supplier's quality system and capability before you rely on their product. EU GMP Annex 11 requires it, and GAMP 5 explains how to scale it to risk. For an AI supplier, I add questions on training data, model change notification and how my data is used. More in how to move from CSV to CSA.
T
Test data independence
The rule that the data used to test a model must never have been used to train or tune it, and must be kept away from the people who build it. The draft EU GMP Annex 22 sets this out in detail, with access control and audit trails as proof. Without it, test results flatter the model. More in the Annex 22 tracker.
Training, validation and test data
The 3 datasets used to build and prove a model. Training data teaches the model, validation data is used to tune it during development, and test data is held back to measure final performance once. In GxP use, each set needs documented origin, labelling checks and version control. More in AI validation in GxP.
U
Unscripted testing
Testing without step-by-step pre-written scripts, carried out by a knowledgeable tester and recorded by its goal, results and issues found. FDA's CSA guidance describes scenario (ad hoc) testing, error guessing and exploratory testing as types of it. It suits lower-risk functions and finding the unexpected failures of AI tools. More in CSV vs CSA.
URS
User requirements specification: the document that states what users need a system to do, in testable terms. GAMP 5 treats it as the starting point of validation. For AI, the URS should state the intended use, the input range and the performance needed, which then become the acceptance criteria. More in computer system validation in pharma.
Turn these terms into working practice
If your team needs these terms turned into working practice, I run in-house AI validation training built on your own systems. For the full method behind each one, my book covers it chapter by chapter.
Sources
Last checked: 12 November 2026. I checked the sources listed below on that date.
- European Commission, consultation on Chapter 4, Annex 11 and new Annex 22
- European Commission, draft Annex 22 consultation document
- EMA, GMDP Inspectors Working Group 3-year work plan 2026 to 2028
- EMA, Annex 22 workshop report
- EUR-Lex, Regulation (EU) 2026/1744
- FDA, draft guidance on AI to support regulatory decision-making (January 2025)
- FDA, Computer Software Assurance guidance
- FDA, CSA guidance text (PDF, February 2026)
- FDA, Quality Management System Regulation
- FDA, PCCP for AI-enabled device software functions (final guidance)
- FDA, Data Integrity and Compliance With Drug CGMP: Questions and Answers (December 2018)
- FDA, warning letter 320-26-58, 2 April 2026
- PIC/S, PI 041-1 (1 July 2021)
- PIC/S, participating authorities
- ISPE, GAMP 5 Second Edition
- ISPE, announcement of the GAMP Guide: Artificial Intelligence (29 July 2025)