AI Glossary

Hesap verebilirlik

Accountability assigns answerability and responsibilities for AI-system functioning and principles according to roles, context, and ability to act.

Example: A company names who owns a chatbot's decisions and who handles complaints.

Ajan

An AI agent is software that interacts with its environment, receives information, and undertakes self-directed actions in service of a larger, externally specified goal.

Example: An agent checks a calendar, proposes a meeting time and waits for approval to book it.

Ajan Geliştirme Çatısı

An AI/agent framework is a reusable software foundation that provides structures and components for building, running, or coordinating AI agents in an application.

Example: A developer uses Google ADK to connect a research agent to a search tool.

Ajan Belleği

In this entry, agent memory means application-managed information retained or retrieved from previous interactions so an agent can use it in later work.

Example: An assistant remembers a user's preferred date format for future reports.

Ajan orkestrasyonu

Agent orchestration is the coordination of which agents run in an application, in what order, and how their next actions are chosen.

Example: A travel agent searches flights, checks availability and then prepares an itinerary.

Ajan Yeteneği

In the Agent Skills open format, an agent skill is a reusable package of instructions and supporting resources, represented by a directory containing at minimum a SKILL.md file with metadata and instructions.

Example: A team gives an agent a SKILL.md with steps for checking a CSV before import.

Algoritma

An algorithm is a computable set of steps for achieving a desired result.

Example: A navigation app follows a route-finding algorithm to suggest a path.

Uygulama Programlama Arayüzü (API)

An API is a defined interface through which one software component requests a capability or data from another component.

Example: A weather app requests today's forecast through a weather service's API.

Genel yapay zekâ (AGI)

AGI denotes broad cognitive capability, commonly framed as matching human performance across many tasks rather than one specialty.

Example: A hypothetical AGI could learn to plan a trip and diagnose a machine fault without separate task-specific training.

Yapay zekâ

Artificial intelligence is a broad field of techniques for building computerized systems that can learn, solve problems, and autonomously pursue goals under varying conditions.

Example: A bank uses AI to flag unusual card purchases for a person to review.

Yapay süper zekâ (ASI)

ASI denotes artificial intelligence that would substantially surpass the best human cognitive abilities across virtually all relevant domains.

Example: A hypothetical ASI would surpass human experts across many areas; no such system has been demonstrated.

Otomasyon yanlılığı

Automation bias is over-reliance on automated cues or recommendations that can contribute to omission or commission errors.

Example: A reviewer accepts an AI risk score without checking the underlying application.

Kıyaslama testi

A benchmark is a defined comparative evaluation arrangement that combines tasks or data, a protocol, and one or more metrics.

Example: Two models answer the same fixed question set and their scores are compared.

Önyargı ve adalet

Bias and fairness concern how systematic patterns, choices, and impacts can produce unequal or unjust outcomes in an AI system and how those outcomes are evaluated in context.

Example: A hiring team checks whether its screening tool rejects qualified applicants from one group more often.

Düşünce zinciri (CoT)

A chain of thought is a sequence of intermediate reasoning steps expressed while working toward an answer.

Example: A worked math solution shows the steps used to reach its answer.

Bağlam

Context is information made available with a model request that can condition how the model produces an output.

Example: A support bot receives the customer's question alongside the relevant return policy.

Bağlam Mühendisliği

In this entry, context engineering is the deliberate selection and arrangement at run time of the information and tools supplied in the format needed for an LLM to perform a task.

Example: A policy assistant receives the latest policy and the user's question before answering.

Bağlam penceresi

A context window is the finite amount of input context a particular model interface can use for a request or conversation state.

Example: A long chat may exceed a model's context window, so older messages are summarized.

Telif hakkı

Copyright is a distinct intellectual-property regime that protects qualifying original expression under the applicable law.

Example: A publisher checks permission before reproducing a photographer's image in a generated brochure.

Veri yönetişimi

Data governance is the organizational practice of managing data used in a stated AI context, including collection, origin, preparation, suitability, possible bias examination, and lifecycle traceability.

Example: A company records who may use its customer data and when it must be deleted.

Veri kalitesi

Data quality is the degree to which data are fit for a defined purpose, considering attributes such as accuracy, completeness, relevance, consistency, reliability, and currency.

Example: A team fixes duplicate and wrongly labelled orders before training a forecast model.

Veri kümesi

A dataset is a collection of data.

Example: A folder of labelled product photos serves as a training dataset.

Karar desteği

Decision support supplies information or recommendations intended to inform a human decision.

Example: A system ranks maintenance requests; a manager decides which repair to prioritize.

Derin öğrenme

Deep learning is an approach to machine learning that uses models with layered representations of a target function.

Example: A camera app uses deep learning to identify objects in a photo.

Gömme

An embedding is a learned numerical vector representation that a system can compare with other vectors or use as model input.

Example: A search app turns similar questions into nearby numerical vectors.

Değerlendirme

AI evaluation is a structured process that produces evidence about how an AI system performs against stated criteria in a stated context.

Example: A team tests whether its support bot answers 100 representative questions correctly.

Değerlendirme veri kümesi

An evaluation dataset is a defined collection of inputs or reference information used to assess a system for a stated purpose.

Example: A team reserves 100 labelled tickets to test a classifier after training.

Kanıt kökeni

Evidence provenance is recorded information that makes an evidence item's origin, derivation, handling, and connection to a claim inspectable.

Example: A report records which dataset and test run produced its accuracy figure.

İnce ayarlama

Fine-tuning is further training that adapts a pre-trained model for a specific task or domain using task-specific data.

Example: A support team fine-tunes a model on reviewed ticket categories.

Temel model

In generative AI, a foundation model is trained on broad data using self-supervised learning and can be adapted for multiple application tasks.

Example: A developer adapts a broadly trained model to summarize company documents.

Üretken yapay zekâ

Generative AI is a class of AI models that generates derived synthetic content by emulating structures and characteristics of input data.

Example: A designer asks an image generator for three poster concepts.

Yönetişim

AI governance is the organizational arrangement of authority, roles, policies, processes, oversight, and evidence used to direct and account for AI across its lifecycle.

Example: A company assigns an owner and review process for its customer-service chatbot.

Referans gerçeklik

Ground truth is task-specific reference information or an accepted target used to compare a system's output.

Example: A verified delivery record is the reference used to check a bot's delivery answer.

Temellendirme

Grounding is the practice of connecting an AI system's output or behaviour to supplied or retrievable evidence and context.

Example: A policy bot shows the document passage behind its answer.

Koruma katmanı

A guardrail is a defined control that constrains, checks, or routes AI-system input, output, tool use, or actions to reduce specified risks.

Example: A payment assistant requires a person's approval before issuing a refund.

Halüsinasyon

An AI hallucination is a colloquial term for a generative-AI output that is erroneous, false, or inconsistent with its prompt or context.

Example: A chatbot cites a research paper that does not exist; the reader checks the citation.

İnsan değerlendirmesi

Human evaluation is structured human judgment of a system or its outputs under a stated task and protocol.

Example: Two reviewers rate whether an AI summary preserves a meeting's key decisions.

İnsan gözetimi

Human oversight enables assigned people to understand relevant limits, monitor, interpret, disregard or override output, intervene, or stop AI-system use as appropriate and proportionate.

Example: A clinician reviews an AI alert and can override it before taking action.

Döngüde insan

In the cited HITL NLP survey, human-in-the-loop is an arrangement that integrates human feedback into a model-development or deployment workflow.

Example: A reviewer must approve an AI-proposed invoice category before it enters the ledger.

İnsan–YZ iş birliği

Human–AI collaboration concerns human roles and responsibilities in Human-AI team configurations, together with feedback and opportunities to challenge AI output.

Example: An assistant drafts a support reply; an employee checks the facts and sends it.

Çıkarım

In this glossary’s operational sense, inference is the use of a deployed model to derive outputs from new inputs.

Example: A trained spam filter labels a newly arrived email as spam.

Fikrî mülkiyet

Intellectual property is an umbrella term for legally protected creations, inventions, identifiers, and confidential information under distinct legal regimes.

Example: A team checks a photo's licence before using it in an AI-generated campaign.

Güvenlik kısıtlarını aşma (Jailbreak)

In language-model security, a jailbreak is an attempt to bypass safeguards so a model produces otherwise restricted behavior.

Example: An authorized tester tries to make a test chatbot ignore its safety rules.

Bilgi damıtma

Knowledge distillation trains a student model using information supplied by a teacher model or ensemble.

Example: A compact model learns from a larger model's answers to classify support tickets.

Büyük dil modeli

A large language model is a deep-learning language model trained on a very large text dataset; a generative LLM commonly produces text through next-token or next-word prediction.

Example: A customer-service agent asks an LLM to draft a reply and checks it before sending.

Okuryazarlık

AI literacy is the skills, knowledge, and understanding that enable informed deployment of AI systems and awareness of their opportunities, risks, and possible harm in context.

Example: An employee learns what company data may be entered into an approved chatbot.

Makine öğrenmesi

Machine learning is a set of techniques that improves a system’s performance by using training data to identify patterns and usually produce models automatically.

Example: An email service learns from labelled messages to identify likely spam.

Metrik

A metric operationalizes a stated criterion for assessing a system or its output.

Example: A support team tracks the percentage of tickets routed to the right department.

Uzmanlar karışımı (MoE)

A mixture-of-experts architecture routes inputs to expert components and combines or selects their contributions.

Example: An MoE model activates selected expert components for each token instead of all of them.

Model

An AI model is a core component that uses inputs to make inferences and produce outputs.

Example: A retailer uses a demand model to estimate next week's sales.

Model Bağlam Protokolü (MCP)

Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems, including data sources, tools, and workflows.

Example: Claude Desktop can connect to an MCP server to access an authorized tool or resource.

Çok modlu model

A multimodal model processes and relates information from multiple sensory modalities, such as vision and touch.

Example: A user uploads a receipt photo and asks a model to read its total.

Yapay sinir ağı

A neural network is a model composed of layers of connected computational units and nonlinear operations.

Example: A phone's photo app uses a neural network to recognize a dog in a picture.

Gözlemlenebilirlik

AI observability is the use of operational signals and context to investigate how an AI-enabled system behaves in use and to investigate the causes of problems.

Example: A team traces a wrong chatbot answer back to the document it retrieved.

Cihaz üzerinde yapay zekâ

On-device AI executes AI inference on a local device, such as a phone or computer.

Example: A phone summarizes a note with a model running on the device.

YZ’nin kurumsal benimsenmesi

In the cited OECD public-sector evidence, organizational AI adoption integrates AI use with governance, workforce capability, institutional strategy, and changes in how work is organized.

Example: A company trains staff and sets data rules before rolling out a writing assistant.

Parametre

Learnable model parameters are fitted using training data during training, and model parameters can change while an AI model is being built.

Example: Training adjusts a model's parameters as it learns to classify photos.

Mahremiyet

Privacy concerns people’s interests and rights in relation to the processing and protection of information about them.

Example: A team removes customer phone numbers before using support chats to test a bot.

İstem

A prompt is input text or other input supplied at inference time to guide a generative model’s response for a task.

Example: “Summarize this meeting in three bullets” is a prompt.

İstem mühendisliği

Prompt engineering is the iterative design and refinement of instructions and context used to guide a model’s response for a task.

Example: A team tests two invoice-extraction prompts and keeps the clearer one.

İstem enjeksiyonu

Prompt injection is an attack in which untrusted content supplies instructions that conflict with or redirect an LLM-integrated system’s intended instructions or data handling.

Example: A web page hides “ignore previous instructions” in text read by an assistant.

Nicemleme

In model compression, quantization represents model values with fewer bits or a reduced set of numerical values.

Example: An engineer reduces a model's numerical precision so it uses less device memory.

Akıl yürütme modeli

A reasoning model is a language model trained or configured to work through intermediate steps before answering complex tasks.

Example: A reasoning model works through a complex scheduling problem before proposing a plan.

Kırmızı takım çalışması

AI red teaming deliberately probes an AI system for harmful outputs or failures so they can be documented and addressed.

Example: An authorized team probes a sandbox chatbot for accidental disclosure of private data.

Güvenilirlik

For an AI system, reliability is the goal of overall correct operation under expected-use conditions over a stated period.

Example: A helpdesk checks whether its routing assistant keeps working accurately each week.

Bilgi getirme

Retrieval is the process of selecting information items from a collection for use in response to a query or task.

Example: A company search tool finds the five policy pages most relevant to a question.

Bilgi getirim destekli üretim (RAG)

Retrieval-augmented generation is a generation approach that combines a model’s parametric knowledge with retrieved external information.

Example: A policy assistant finds the current leave policy before drafting an answer about vacation days.

Risk değerlendirmesi

AI risk assessment is the context-specific process of identifying, analyzing, prioritizing, and documenting potential AI-related harms and uncertainties for affected people and organizations.

Example: Before using AI to screen applications, a team records possible errors and affected people.

Sağlamlık

Robustness is the goal of appropriate system functionality across a broad stated set of conditions, including unanticipated uses.

Example: A team checks whether a document classifier still works on blurry scans.

Güvenlik

AI security is the protection of AI systems and dependent assets against threats to confidentiality, integrity, availability, and intended behavior.

Example: A company gives its document assistant read-only access to approved files.

Küçük dil modeli (SLM)

A small language model is a language model described as relatively compact in parameter count and resource needs for its intended use.

Example: A phone app uses a compact language model to sort short notes locally.

Yapılandırılmış çıktı

Structured output is model output requested in a specified machine-readable shape, such as fields governed by a schema.

Example: An invoice assistant returns a JSON object with amount, currency and due date.

Sentetik veri

Synthetic data is artificially generated data used in place of or alongside directly collected observations.

Example: A team generates fictional support tickets to test a classifier without using real customer messages.

Sistem

An AI system is a machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations, or decisions for explicit or implicit objectives.

Example: A support system reads a customer message and suggests which team should answer it.

Sistem talimatı

A system instruction is a high-level set of instructions and context supplied by an application to steer a chat model’s responses.

Example: An app instructs its assistant to answer in Turkish and cite supplied documents.

Çıkarım zamanı hesaplaması

Test-time compute is computation spent producing an answer at inference time; scaling it allocates extra work to a request.

Example: An app gives a model more time to compare candidate plans before answering.

Token

In OpenAI’s documented text processing, a token is a unit into which text is divided for a model to process or generate.

Example: A long document uses more tokens than a short question when sent to a chatbot.

Araç kullanımı

Tool use is an application pattern in which a model’s output is used to select or call a configured function or external capability and process its result; provider documentation may call this function calling or tool calling.

Example: A travel assistant calls a weather service before suggesting what to pack.

Eğitim

Training is the process of improving a system’s performance with machine-learning techniques, commonly by fitting learnable model parameters using training data.

Example: A team trains a spam filter with previously labelled email.

Transformer mimarisi

The Transformer is a neural-network architecture that relies on attention mechanisms to model relationships between elements of input and output sequences.

Example: A translation model uses a Transformer to relate words across a sentence.

Şeffaflık

Transparency is the provision of relevant information about an AI system to appropriate audiences at a usable depth for the stated context.

Example: A bank tells customers when a chatbot is answering and how to reach a person.

Kullanım senaryosu

A use case is a bounded description of an intended AI use in a stated process and context, including its system objective and people or groups likely to be affected.

Example: A team defines a use case: summarize meeting notes for employees to review.

Vektör veritabanı

A vector database stores vector representations and supports retrieving records by vector similarity.

Example: A search app stores document vectors to find passages similar to a question.

İş akışı yeniden tasarımı

Workflow redesign concerns task reorganisation reported where AI-adopting employers automate existing tasks and create new ones.

Example: A support team adds a fact-check step before sending AI-drafted replies.