Who is this AI glossary for?

Managers, business teams and technical staff involved in AI projects, training or regulatory work. Definitions are kept short and need no technical background.

Do the definitions replace legal definitions?

No. Regulatory terms (EU AI Act, Turkish data protection law, etc.) are summarised for general information; binding definitions are in the legislation itself.

Is the glossary updated?

Yes. We add new terms and update existing definitions as new concepts and regulatory changes emerge.

AI Glossary: 102 Key Terms Explained

Core concepts

AI literacy

The knowledge and skills needed to use AI systems in an informed way and to understand their opportunities, limits and risks. Article 4 of the EU AI Act expects organisations using AI systems to support their staff's AI literacy.

AI maturity model

A framework that grades an organisation's AI capability in levels across dimensions such as strategy, data, technology, people and governance. It helps decide where to start and what comes next.

Artificial intelligence (AI)

The broad term for computer systems that perform tasks normally requiring human intelligence, such as learning, reasoning, understanding language, recognising images and making decisions. Most AI used in organisations today is based on machine learning.

Computer vision

The field of AI that extracts meaning from images and video. Quality control in manufacturing, document reading (OCR) and object recognition are common uses.

Data literacy

The ability to read, interpret and question data and to make data-driven decisions. It is a prerequisite for judging AI outputs correctly.

Deep learning

Machine learning with multi-layer artificial neural networks. It excels at unstructured data such as images, audio and text; large language models are deep learning models.

Generative AI (GenAI)

AI systems that create new content such as text, images, audio or code. Chat assistants such as ChatGPT, Claude and Gemini are the best-known examples.

Inference

Running a trained model on new data to produce outputs. Most production cost and latency arise at the inference stage.

Related: Model serving · GPU

Machine learning (ML)

An approach to AI in which a system learns patterns from example data instead of following hand-written rules, and uses them to make predictions or decisions. Credit scoring, demand forecasting and fraud detection are typical examples.

Model

A mathematical structure that captures patterns learned from data and produces outputs (predictions, classes, text) for new inputs. An AI system usually combines one or more models with data pipelines and application code.

Related: Training data · Inference

Natural language processing (NLP)

The field of AI concerned with enabling computers to understand, process and generate human language. Text classification, sentiment analysis, summarisation and question answering all belong to NLP.

Related: Large language model (LLM)

Neural network

A model made of interconnected, weighted computing units (neurons). During training the weights are adjusted to reduce the model's error.

Related: Deep learning

Proof of concept (PoC)

A short, limited-scope study that tests whether an AI idea works technically and commercially on real data before a larger investment.

Shadow AI

Employees using AI tools for work that the organisation has not approved or is not aware of. The main risk is personal or confidential data being entered into external services without control.

Training data

The dataset a model learns patterns from. Its quality, representativeness and freshness often affect model performance more than the choice of algorithm.

Use case

A definition of how AI will be applied to a specific business problem: who will use it, with which data, which decision or task it supports and how success will be measured.

Machine learning

Anomaly detection

Finding records that deviate markedly from normal behaviour. It is used for early detection of fraud, security incidents and equipment failures.

AutoML

Tools that automate steps such as model selection, feature preparation and parameter tuning. They speed up the start, but problem definition and data quality still need human expertise.

Related: Machine Learning Without Code

Classification

A modelling task that assigns an input to one of several predefined classes, such as fraud / not fraud, will churn / will stay, or the topic of an email.

Clustering and segmentation

Grouping similar records without labels. Building customer segments for marketing is the most common use.

Related: Unsupervised learning

Credit scoring

Models that estimate the likelihood that a customer will not repay a loan. Because they drive decisions that directly affect customers, they require regular validation, explainability and monitoring.

Cross-validation

Splitting data into several folds and testing the model on a different fold each time, giving a more reliable estimate of performance.

Related: Overfitting

Explainable AI (XAI)

Methods that show, in a way people can understand, why a model produced a particular result. Important for both trust and regulation in decisions such as credit, insurance and hiring.

Feature and feature engineering

A feature is a variable a model uses as input (income, number of transactions in the last three months, etc.). Feature engineering is deriving meaningful variables from raw data that make learning easier.

Related: Feature store

Gini coefficient (in credit scoring)

A measure widely used in banking to assess the discriminatory power of credit scoring models, calculated as Gini = 2 × AUC − 1.

Imbalanced data

Data in which one class is far rarer than others, for example when only a tiny share of transactions is fraudulent. It requires careful sampling and the right performance metrics.

Related: Precision and recall

Overfitting

When a model memorises the training data and learns noise instead of general patterns. It performs very well on training data and poorly on new data.

Precision and recall

Precision is the share of cases flagged positive that really are positive; recall is the share of real positives the model catches. In imbalanced problems such as fraud detection these are used instead of accuracy.

Regression

A modelling task that predicts a numerical value, such as sales volume, price or loss amount.

Related: Time-series forecasting

Reinforcement learning

A method in which an agent learns the best behaviour by interacting with an environment and receiving rewards and penalties. It is also used to refine language models with human feedback (RLHF).

ROC curve and AUC

A measure summarising how well a classification model separates positive and negative cases across all thresholds. An AUC of 0.5 means random guessing; 1 means perfect separation.

SHAP

A widely used explainability method, based on Shapley values from game theory, that calculates each variable's contribution to an individual prediction.

Related: Explainable AI (XAI)

Supervised learning

Training a model on past examples whose correct answer (label) is known, for example predicting default probability from loans that were and were not repaid.

Related: Classification · Regression

Time-series forecasting

Predicting future values such as demand, sales, cash flow or energy consumption from historical time-series data.

Related: Machine Learning Models

Train / validation / test split

Dividing data into separate parts to train a model, tune its settings and measure its final performance. The test set shows how the model behaves on unseen data; for time-dependent data an out-of-time test is important.

Unsupervised learning

Learning without labels, aimed at discovering structure and groups in data. Customer segmentation and anomaly detection are typical uses.

Generative AI and LLMs

AI agent

An AI system that plans its own steps towards a goal, takes actions using tools (search, databases, email, APIs) and continues based on the results. Because it can act, permission design and human approval are critical.

Chunking

Splitting documents into searchable pieces for RAG. Chunk size and preserving structure such as headings and tables directly affect answer quality.

Context window

The total number of tokens a language model can take into account at once, including instructions, documents, chat history and the answer. It affects whether methods such as RAG are needed for long documents.

Embedding

A list of numbers (a vector) that represents the meaning of a piece of text. Texts with similar meanings have nearby vectors; semantic search and RAG rely on this.

Fine-tuning

Further training a pre-trained model on examples specific to an organisation or task. It is effective for teaching style and behaviour; for adding up-to-date knowledge, RAG is usually more suitable.

Foundation model

A large model trained on broad, diverse data that can be adapted to many different tasks. The EU AI Act uses the concept of a general-purpose AI model for such models.

Groundedness

A quality measure of whether a RAG answer is actually based on the retrieved passages or whether the model has added information.

Guardrails

Rules and filters that check the inputs and outputs of an AI application, such as masking personal data, blocking off-topic or harmful content and requiring approval for certain actions.

Related: Prompt injection

Hallucination

When a language model produces false information in a convincing, confident way, such as invented sources, figures or legal references. RAG, verification steps and user training reduce it but do not eliminate it.

Hybrid search

Combining semantic (vector) search with keyword search. For queries containing product codes, clause numbers or proper names it usually works better than semantic search alone.

Large language model (LLM)

A deep learning model trained on very large text corpora that generates language by predicting the next piece of text. It is the foundation of assistants such as ChatGPT, Claude and Gemini.

LLM evaluation (evals)

Systematically testing the quality of a language model application with expert-prepared question–answer sets and automated metrics. It should be repeated after every model, prompt or data change.

Related: Groundedness · LLMOps

MCP (Model Context Protocol)

An open protocol that lets AI applications connect to tools and data sources in a standard way. Announced by Anthropic in 2024, it allows a connector built once for internal systems to be reused across different AI applications.

Multimodal model

A model that can take or produce more than one type of data, such as text, images, audio or video, for example an assistant that reads and summarises a photo of an invoice.

Open-weight model

A language model whose weights can be downloaded and run on an organisation's own infrastructure. It is the basis for on-premise deployment when data must not leave the organisation; licence terms vary by model.

Prompt

The instructions and context given to a generative AI model. Answer quality depends heavily on how clear, contextual and well-structured the prompt is.

Related: Prompt engineering

Prompt engineering

The skill of systematically designing instructions to get the desired result from AI: giving a role and context, defining the format, showing examples, setting limits and iterating.

Prompt injection

An attack in which instructions hidden in text the model processes (an email, web page or document) redirect the system. It is one of the most serious security risks, especially for agents with tool access.

Related: AI agent · Guardrails

Quantization

Storing model weights at lower numerical precision to reduce memory and hardware needs. It cuts the cost of on-premise LLM deployments, but the quality impact should be tested.

RAG (retrieval-augmented generation)

An architecture that retrieves relevant passages from an organisation's documents and passes them to the model together with the question. The model answers based on those passages and can cite them; it is the most common design for enterprise knowledge assistants.

System prompt

Persistent instructions given to the model in an AI application before user messages, defining its role, rules and limits.

Related: Guardrails

Temperature

A setting that controls randomness in a language model's answers. Lower values give more consistent, repeatable answers; higher values give more creative, varied ones.

Token

The basic unit in which language models process text: a word, part of a word or a punctuation mark. Model costs and context windows are measured in tokens.

Related: Context window

Tool use (function calling)

The ability of a language model to decide to call a defined function or API instead of only producing text. It is the basis for agents working with external systems.

Transformer

A neural network architecture introduced in 2017 that models relationships between words using an attention mechanism. Almost all of today's large language models are based on it.

Related: Large language model (LLM)

Vector database

A database that stores embeddings and quickly finds the records closest in meaning to a query. It is the retrieval layer of RAG systems, and access permissions must be enforced here too.

MLOps and infrastructure

AIOps

Using AI in IT operations for incident detection, root-cause analysis and automated response. It differs from MLOps: AIOps applies AI to IT, while MLOps operates AI models.

Amazon Bedrock

AWS's generative AI platform that provides access to foundation models from different providers through a single managed service.

Related: AWS Consulting

Amazon SageMaker AI

AWS's managed platform for building, training, deploying and monitoring machine learning models.

Related: AWS Consulting · MLOps & LLMOps

CI/CD (continuous integration and delivery)

A process in which code changes are automatically tested and merged (CI) and automatically released to production (CD). In ML it is extended with data and model tests.

Related: DevOps

Concept drift

When the relationship between inputs and the target changes, for example when fraudsters change tactics and old patterns no longer hold. It usually requires retraining.

Related: Data drift

Containers and Kubernetes

A container packages an application with all its dependencies so it runs the same way everywhere (e.g. Docker). Kubernetes is a platform that automates scaling and managing many containers.

Related: DevOps

Data drift

When the distribution of production data gradually differs from the data a model was trained on, for example customer profiles changing as economic conditions change.

Related: Concept drift · Model monitoring

Data pipeline

An automated flow that takes data from source systems, cleans and transforms it, and prepares it for models or reporting.

Related: Integration & Data Pipelines

DevOps

A culture and set of practices in which development and operations teams work together to deliver software frequently, safely and automatically. It underpins MLOps.

Experiment tracking

Recording every model experiment with the data, code, parameters and results used, so it can later be explained which model was chosen and why.

Related: Model registry

Feature store

A central system where model features are defined once and used identically in training and production, reducing inconsistencies between the two.

GPU

A graphics processor suited to parallel computation. It is the core hardware for training deep learning models and running large language models, and the largest cost item in on-premise LLM deployments.

LLMOps

MLOps principles adapted to language model applications: prompt and model versioning, evaluation sets, cost and latency tracking, safety controls and usage monitoring.

MLOps

A set of practices that automate and standardise the development, testing, deployment, monitoring and retraining of machine learning models, turning models into sustainable products instead of one-off projects.

MLOps maturity levels

A framework describing how automated and repeatable an organisation's model development and operations are, in levels from manual work to full automation.

Model monitoring

Continuously tracking a production model's input data, predictions and business outcomes to spot performance degradation early.

Related: Data drift · MLOps & LLMOps

Model registry

A central store of approved model versions with their metadata and approval status. It makes it easy to know which version is in production and to roll back.

Related: Experiment tracking · MLOps

Model serving

Deploying a trained model as a service, usually an API, that applications can call, either in real time or in batch.

On-premise deployment

Running software and models in the organisation's own data centre or infrastructure under its control. Preferred in sectors such as banking and the public sector where data cannot leave the organisation.

Governance, risk and regulation

AI acceptable use policy

An internal policy setting out which AI tools employees may use, with which data and under which rules. It is usually the first step in AI governance.

AI governance

The roles, policies, processes and controls that determine how AI is selected, developed, used and overseen in an organisation.

AI inventory

A register of all AI systems used and developed in an organisation, with their purpose, owner, data and risk level. It is the starting point for risk classification and regulatory compliance.

Anonymisation and pseudonymisation

Anonymisation makes data impossible to link to a person in any way. Pseudonymisation replaces identifiers with a code but can be reversed with additional information, so the data still counts as personal data.

Bias

When a model systematically advantages or disadvantages particular groups. It often comes from training data that reflects past decisions and creates discrimination risk in areas such as hiring and lending.

Related: Responsible AI

Cross-border data transfer

Transferring personal data to servers or parties outside Türkiye. Entering personal data into AI services operated abroad falls under this and is subject to the conditions of Article 9 of KVKK, amended in 2024.

Data residency

A requirement to store and process data within a particular country. In regulated sectors such as banking it directly affects the choice of infrastructure (cloud or on-premise).

EU AI Act

The European Union regulation (2024/1689) that governs AI systems according to their level of risk. Because it also covers systems whose output is used in the EU, it can affect Turkish companies serving the EU market.

General-purpose AI model (GPAI)

Under the EU AI Act, models that can perform a wide range of tasks and be integrated into different systems, such as large language models. Their providers have documentation and transparency obligations.

High-risk AI system

Under the EU AI Act, systems that can significantly affect people's health, safety or fundamental rights, such as those used for creditworthiness assessment and recruitment. They bring obligations such as risk management, data quality, record-keeping and human oversight.

Human oversight (human-in-the-loop)

A person reviewing, approving or, if needed, stopping AI outputs in critical decisions. It is a core control for high-risk systems and agents that take actions.

Related: High-risk AI system · AI agent

ISO/IEC 42001

The international standard for AI management systems. It gives organisations a framework to manage AI systematically through policy, risk assessment, controls and continual improvement, and can be certified by accredited bodies.

KVKK (Turkish Personal Data Protection Law No. 6698)

The Turkish law governing the processing of personal data. Entering personal data into AI tools, using it to train models and transferring it to services abroad are all assessed under KVKK.

Model risk

The risk of financial, operational or reputational loss from a model that is poorly designed, misused or has degraded over time. In banking it is managed through a model inventory, independent validation and monitoring.

Model validation

An independent review, by a team separate from the developers, of a model's conceptual soundness, data quality, performance and stability.

NIST AI RMF (AI Risk Management Framework)

A voluntary framework published by the US National Institute of Standards and Technology. It addresses AI risk through four functions: Govern, Map, Measure and Manage.

Related: AI governance

Provider and deployer

Under the EU AI Act, the provider develops a system or places it on the market under its own name; the deployer uses the system in its own activities. Obligations differ by role.

Related: EU AI Act: Am I in Scope?

Responsible AI

Principles for developing and using AI in a fair, transparent, safe, privacy-respecting and accountable way.

Risk categories (EU AI Act)

The four levels into which the Act divides AI systems: prohibited practices, high-risk systems, limited-risk systems with transparency obligations, and minimal-risk systems with no additional obligations.

Transparency obligation

The obligation to make sure people know they are interacting with AI or that content was generated by AI, for example informing users of a chatbot.

Related: Risk categories (EU AI Act)

Türkiye AI Action Plan (2026–2030)

Türkiye's plan setting out national AI goals and actions, put into effect by a Presidential circular. Its main headings are awareness and skilled human capital, data and computing infrastructure, and wider AI adoption in the public and private sectors.

AI Glossary

Core concepts

Machine learning

Generative AI and LLMs

MLOps and infrastructure

Governance, risk and regulation

Frequently Asked Questions

Let's adapt the template to your organisation