The idea that artificial intelligence should be developed, deployed, and used in a way that ensures that the legal responsibility for bad outcomes can be assigned to the correct individuals or organisations.

The process of evaluating and comparing AI models using standardized tasks and datasets to measure performance factors like accuracy, speed, efficiency, robustness, and fairness. It helps developers track progress, identify strengths and weaknesses, and ensure consistent evaluation across systems.

The ability of AI tools to contribute to human creativity by suggesting ideas, automating tasks, and generating new content. AI lacks true creative consciousness and relies on human-generated data, input and direction.

As AI systems learn from data which may be unbalanced and or reflect discrimination, they may produce outputs which have discriminatory effects on different groups of people (e.g. relating to gender or race).

A response generated by AI that contains false or misleading information presented as fact.

People's ability to understand, use, monitor, and critically reflect on AI applications as they become more common in their everyday lives.

A process for organizations to identify, analyze, and mitigate potential harms associated with AI systems, including technical, ethical, social, and legal implications.

Automated translation of text between languages using AI. Models learn from large bilingual corpora to convert text from one language to another. These systems perform translation tasks automatically or with human post-editing.

A step-by-step procedure for solving a mathematical problem or accomplishing a specific end: rules used by a computer to sort through a mathematical problem.

Describes systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others. Bias can emerge due to many factors, including the design of the algorithm or the unintended or unanticipated use or decisions relating to the way data is coded, collected, selected or used to train the algorithm.

AI automation is the use of artificial intelligence to perform routine or repetitive tasks with minimal human input. This is intended to increase efficiency by letting machines handle data processing, analysis, or decision-making that would otherwise require manual effort.

Backpropagation is a method used to train neural networks by helping them learn from mistakes. It works by comparing the network’s prediction to the correct answer, figuring out where it went wrong, and then adjusting the internal settings (called weights) to do better next time. This process moves backward through the network, step by step, improving accuracy with each round of learning.

A computer program that runs automated tasks over the internet, e.g., mimicking human behaviour online.

The change of the world's climate which is attributed directly or indirectly to human activity, e.g. increased carbon dioxide emissions. The negative impacts of climate change include rising temperatures, rising sea levels, drought, flooding. Somalia is one of the countries that is most vulnerable to the impacts of climate change.

Refers to a model of computing that delivers on-demand access to shared resources—like servers, storage, and applications—over the Internet.

Coding is the process of creating instructions, called programs, that a computer can follow to carry out tasks. It involves designing and writing algorithms—step-by-step methods for solving problems—in a formal language that computers can understand. Coding enables humans to build software, automate actions, and control how digital systems behave.

Refers to the processing capacity of computer hardware that is critical for developing and running AI systems. As AI models grow more complex, they demand greater computing resources, concentrating access among a few powerful firms and raising concerns about equity, accessibility, and environmental impact.

The process by which an online platform (e.g. a social media company) monitors user-generated content (e.g. social media posts) based on platform-specific rules and guidelines to determine if the content should be allowed to remain on the online platform, or not.

The process of removing or encrypting personally identifiable information from datasets so that individuals cannot be identified. Anonymization techniques (like removing names or masking IDs) is intended to protect privacy by preventing the linkage of data to individuals, while preserving the dataset’s utility for analysis.

Data augmentation is a technique to increase the diversity of training data without collecting new data. It creates new examples by transforming existing ones—such as rephrasing text, changing image lighting, or altering audio speed—to help models generalize better.

A temperature-controlled building that houses computing infrastructure, such as servers, data storage drives, and network equipment. These are used for the development and operation of generative AI.

The appropriation and exploitation of data, particularly from marginalized communities, by powerful actors, mirroring the patterns of historical colonialism and leading to new forms of control and inequality.

Branch of data security concerned with the proper handling of data – consent, notice, and regulatory obligations. More specifically, practical data privacy concerns often revolve around whether or how data is shared with third parties.

Tracking the origin, history, and lineage of data. This records where data came from, how it was processed, and by whom, to ensure integrity and reproducibility. For AI, this is supposed to help verify that training data is from legitimate sources and is correctly processed, aiding transparency and trust in the model’s development.

A usually automated process where a technology takes data from a particular program or websites for a variety of uses that usually results in aggregation for analysis.

Deep learning is a type of machine learning that uses artificial neural networks to learn from data. Artificial neural networks are inspired by the human brain, and they can be used to solve a wide variety of problems, including image recognition, natural language processing, and speech recognition.

Synthetic media in which a person in an existing image or video is replaced with someone else's likeness, created by deep learning and AI.

The unequal access to and use of digital technologies]], such as the Internet and digital devices, between different groups and regions, leading to disparities in opportunities and outcomes.

A state's control over national data, its local capacity for digital technology development, and its ability to shape and enforce laws in the digital environment.

False information that is given deliberately.

Edge AI refers to the deployment of AI models on local devices—such as smartphones, cameras, or IoT sensors—so data is processed near its source instead of in the cloud. This reduces latency, saves bandwidth, improves privacy, and allows for real-time decision-making, even without constant Internet access.

Embeddings in AI are numerical representations of data—such as words, sentences, or images—that capture their meaning and relationships in a format machines can understand. For instance, in natural language processing, words like "king" and "queen" are represented numerically as being close to each other in a multidimensional space, reflecting their semantic similarity. This enables AI systems to perform tasks like language translation, sentiment analysis, and image recognition by understanding and processing the underlying context and relationships within the data.

Explainable artificial intelligence (XAI) is a set of processes and methods that allows human users to understand the results and output created by machine learning algorithms. However explainability can be very difficult in relation to generative AI, because even the developers of the technology may not know or be able to find out why a particular output has been generated.

The process of transforming raw data into meaningful inputs—called features—that help machine learning models make better predictions. It involves identifying which aspects of data are most predictive (e.g., temperature, time of day) and converting raw data into those features. Good feature engineering often greatly boosts model accuracy.

Federated learning is a way to train AI models across many devices or servers without sharing raw data. Each device learns from its own private data and sends only updates to a shared model, helping protect privacy—especially when data is sensitive or spread out.

Fine-tuning is the process of taking a pre-trained machine learning model and adapting it to a specific task or dataset by continuing its training on new, relevant data. This approach builds on the model’s existing knowledge—learned from a large, general dataset—so it can do more specialized tasks, even with limited available data.

The use of artificial intelligence to monitor, filter, and suppress content, often with the goal of controlling information flow and limiting free expression. For example, a government may seek to prevent a AI chatbot from discussing particular topics for political reasons.

Laws, regulations and policies made by states or international organisations to control the development of AI technologies.

Generative AI is a type of artificial intelligence that can create new content like text, images, music, and code, based on the data it's trained on, mimicking human creativity. These AI systems learn patterns and structures from massive datasets in order to generate new outputs. Some generative AI (like ChatGPT) use large language models to create human-like text, stories, and code. Image-based AI models (like DALL-E and Midjourney) can create realistic or surreal images from text prompts. Generative AI models often require large-scale computing infrastructure and massive amounts of energy. Generative AI makes predictions based on existing data, it does not always represent the truth. Societies around the world are debating the opportunities and risks of generative AI as it becomes used more widely in society.

Image generation is a form of generative AI that creates new visual content—such as photos, illustrations, or graphics—by learning patterns from existing images.

The ability of an individual to understand and give their agreement for their data to be processed using AI, or for the training of AI.

AI infrastructure refers to the hardware and software systems required to build, train, and deploy AI applications.

An idea, a design, etc. that somebody has created and that the law prevents other people from copying. When AI training data includes copyrighted material (without permission) people have argued that this violates intellectual property laws.

The global system of interconnected computer networks.

A network of physical objects (“things”) embedded with sensors, processing ability, and connectivity, enabling them to exchange data with other devices or systems. IoT devices (such as smart home appliances or industrial sensors) communicate over the Internet to collect and share information, enabling applications like remote monitoring and automation.

Large language models (LLMs) are very large deep learning models that are pre-trained on vast amounts of data. They use neural networks to extract meanings from a sequence of text and understand the relationships between words and phrases in this text.

AI methods that are designed for languages with limited available training data. Such models often borrow knowledge from similar, data-rich areas to enable them to work with limited information.

The use and development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyse and draw inferences from patterns in data.

The design and structure of an AI system, outlining how data flows through the system, how information is processed, and how results are produced.

Model compression refers to techniques that reduce the size and complexity of neural networks by removing redundant parameters or making the model more efficient. This lowers storage and energy use, speeds up inference, and enables deployment on devices with limited resources, like smartphones.

Model watermarking is a way to secretly mark an AI model to prove who created or owns it. It works by slightly changing the model so it gives special, recognizable responses to certain inputs—helping detect if the model has been copied or stolen.

Multimodal AI refers to systems that can understand and combine different types of data—like text, images, or audio—at the same time. For example, it might use visual materials and language to answer questions about an image.

Natural language processing (NLP) is a type of computer science and AI that uses machine learning to enable computers to understand and communicate with human language.

A type of AI modeled on the human brain. Instead of having a set of rules coded in like a traditional computer program, an artificial neural network learns to detect and predict patterns over time.

Used to describe software for which the original source code is made freely available to anyone to adapt or share.

Overfitting is a phenomenon that occurs when a model learns the training data too well leading to poor performance on new, unseen data. It often happens when the model is too complex for the dataset, effectively “memorizing” examples instead of learning general patterns

A statement that says what you think will happen; the act of making such a statement.

Prompt engineering is the process of writing, refining and optimizing inputs to encourage GG Curiyeed (Generative AI) systems to create the specific outputs that the user wants.

Reinforcement learning is a machine learning method where an AI system learns to make decisions by interacting with its environment and receiving feedback—rewards for good actions and penalties for bad ones. Its goal is to maximize total rewards over time. This approach is important for building adaptive systems that can learn and improve in complex, changing environments.

The computational power required to train generative AI models can require a huge amount of electricity, which leads to increased carbon dioxide emissions. Large amounts of water are also needed to cool the hardware used for generative AI models.

An approach to developing and using AI systems that aligns with ethical principles and societal values, ensuring they are safe, trustworthy, and beneficial.

Ensuring that AI systems behave in ways aligned with human values and safety constraints. Alignment aims to make a model’s goals and actions match user intent or ethical norms, preventing harmful or unintended behavior.

Also called text-to-speech (TTS), this is the generation of speech by artificial means, i.e. converting written content into spoken audio using AI models of pronunciation and voice.

Supervised learning is a type of machine learning where models are trained on labeled data—meaning the correct answers are already known—to learn patterns and make accurate predictions. It's widely used in fields like healthcare, finance, and marketing to classify data or forecast outcomes based on past examples

Large language models (LLMs) can be used to compose responses to questions or prompts that are statistically likely, based on patterns in their training data and on subsequent fine tuning by techniques such as feedback from human testers.

Training data refers to a curated set of labeled data (or examples) —such as text, images, audio, or video—that is used to teach a machine learning (ML) model how to recognize patterns, make predictions, or perform specific tasks. Training data includes both the input data and the correct output (label), enabling the model to learn from experience. The quality, quantity, and diversity of training data directly impact the performance and generalization ability of the resulting AI system

A technique where knowledge from one task or dataset is reused to improve learning in another related but different task or dataset. This reduces training time and data needs, as the model transfers its learned representations to the new task.

Transformers are a type of AI model designed to understand and generate sequences—like sentences—by paying attention to all parts of the input at once. They are used in advanced language tools and in areas like image analysis

The visibility of how AI systems are built. Many people argue that this visibility should be part of every aspect of AI development and deployment, from understanding the machine learning model and the data it is trained on, to understanding how data is categorized and the frequency of errors and biases, to the communications among developers, users, stakeholders and regulators.

Unsupervised learning is a type of machine learning where models learn patterns and structures in data without labeled examples or human guidance. It helps uncover hidden insights—like grouping similar items or detecting anomalies—and is used in tasks such as recommendation systems, identifying different types of customers, and data exploration

Video generation is a type of GG Curiyeed (Generative AI) that creates new videos by learning patterns of motion and visuals from existing footage. It produces realistic or imaginative scenes based on inputs like text prompts or images.

Zero-shot learning (ZSL) is a machine learning scenario in which an AI model is able to recognize and categorize objects or concepts without having seen any examples of those specific categories or concepts during training, by relying on prior knowledge and learned relationships.